CBR for the management and reuse of image-processing
expertise:a conversational system
V.Ficet-Cauchard*,C.Porquet,M.Revenu
GREYC-ISMRA,6Boulevard du Mare?chal Juin,F14050Caen cedex,France
Received1December1998;accepted1July1999
Abstract
The development of an image-processing(IP)application is a complex activity,which can be greatly alleviated by user-friendly graphical programming environments.The major objective of the work described in this paper is to help IP experts reuse parts of their applications.A?rst step towards knowledge reuse has been to propose a suitable representation of the strategies of IP experts by means of IP plans(trees of tasks,methods and tools).This paper describes the CBR module of an interactive system for the development of IP plans.After a brief presentation of the overall architecture of the system and its other modules,the authors explain the distinction between an IP case and an IP plan,and give the selection criteria and functions that are used for similarity calculation.The core of the CBR module is a search/adaptation algorithm,whose main steps are detailed:retrieval of suitable cases,recursive adaptation of the selected one and memorization of new cases.The system's implementation is presently completed;its functioning is described in a session showing the kind of assistance provided by the CBR module during the development of a new IP application.#1999Elsevier Science Ltd.All rights reserved. Keywords:Case-based reasoning;Image-processing;Application reuse;Interactivity;Knowledge management;Recursive algorithms
1.Introduction
Research work is being undertaken by the authors into the design of an interactive system that can pro-vide assistance during the working out of image-pro-cessing(IP)applications;the system's architecture has been detailed by Ficet-Cauchard et al.(1998).The sys-tem is composed of several modules dealing with the tuning of IP applications through interactive acqui-sition and the representation of IP knowledge coming from IP experts,the execution of such IP applications, and the reuse of applications following a case-based reasoning approach(CBR).This paper is dedicated to a detailed description of the CBR module:in particu-lar,a description of IP cases,a similarity calculation between two cases and a recursive search/adaptation algorithm are presented and discussed.
In Section2,the framework of this research is brieˉy presented along two axes:the objectives with regards to IP and the modeling of an IP application. Sections3and4deal with the CBR module:?rst,a de?nition of an IP case and the functions used for similarity calculation are given(Section3).Then the search/adaptation algorithm is described,and expla-nations are given about the process of case selection, recursive adaptation of the selected solution and mem-orization of new cases(Section4).Finally,a complete session showing how to use the CBR module for devel-oping an IP application is described in Section5.
2.Research framework:the TMT model
The primary objective of this work is to represent and structure the knowledge of di erent IP experts so as to enable knowledge sharing and reuse.To achieve
Engineering Applications of Arti?cial Intelligence12(1999)733±747
0952-1976/99/$-see front matter#1999Elsevier Science Ltd.All rights reserved.
PII:S0952-1976(99)00041-
X https://www.doczj.com/doc/308883697.html,/locate/engappai
*Corresponding author.Tel.:+33-02-31-45-27-21;fax:+33-02-31-45-26-98.
E-mail address:valerie.?cet@greyc.ismra.fr(V.Ficet-Cauchard).
such a goal,an interactive system is advocated, enabling knowledge acquisition from IP experts,as it arises through the development of IP applications.In this section,an approach to the building of appli-cations is presented;the section describes a model for the representation of applications,and brieˉy gives an idea of the functioning of two essential modules of the system:the interactive creation module and the ex-ecution module.
2.1.Representation of applications by hierarchical plans The approach adopted here to the development of IP applications is based on the smart supervision of libraries of operators.An operator is a program that performs one basic operation on one or several images. It takes as input the image(s)to be processed,as well as parameters,and produces as output one or several images,as well as numerical and/or symbolic results. With such libraries,the building of an application then `simply'consists of linking operators and tuning their https://www.doczj.com/doc/308883697.html,ers can thus stand back from computer codes and perform programming at the`knowledge' level.
A?rst category of systems that facilitate the use of such libraries are the graphical programming environ-ments.Such systems(IRIS,1993;Rasure and Kubica, 1994)help users to select and organize operators into sequences,but they do not enable them to explain,nor model their reasoning.It is thus di cult to reuse sol-ution elements that were previously built for other ap-plications.
A second category of IP systems that use program libraries are the automatic planners such as OCAPI (Cle ment and Thonnat,1993)or BORG(Clouard et al.,1999),the planning system developed within the authors'own research team.Such systems are based on an explicit knowledge representation,enabling reuse.However,the knowledge is represented for com-putational purposes rather than for the user's under-standing,who seldom intervenes during the search for a solution.
In order to take full advantage of the interactive nature of graphical programming environments,this system should enable users to select and link IP oper-ators graphically,but it should also give them a means to represent and make explicit the reasoning that has led them to this series of operators,with the objective of reusing the implemented strategy,in the same way as automatic systems do.
However,a real-sized application can lead to sequences of up to tens of operators.In order to rep-resent the reasoning associated with such sequences, this paper suggests the use of a representation based on trees of tasks,called`IP plans'.Such trees corre-spond to hierarchical decompositions of problems into sub-problems,each problem or sub-problem being re-lated to an IP task.As shown in Fig.1,a plan rep-resents not only the linking of IP operators corresponding to the leaves of the tree,but also all the reasoning necessary for the creation of such a linking, which is represented by IP tasks represented as gray boxes.
2.2.The TMT model
In this system,both IP plans and control tasks (dealing with plan management and system control) are uniformly represented within the`taskDmethod Dtool'model.In this model,a task represents a goal or sub-goal;a method describes a piece of know-how, it speci?es how a task can be performed;a tool rep-resents a computer code(IP operator,Lisp or C func-tion)in conceptual terms with a link to the code enabling the user to run it.There exist two types of methods:`terminal'methods(Fig.2a),which achieve a task by calling a computer code through the medium of a tool,and`complex'methods(Fig.2b),which decompose a task into sub-tasks by means of a `THEN'tree.Finally,as there may exist several strat-egies to solve a particular IP problem,a task can be associated with several methods(Fig.2c)by means of an`OR'tree,the choice of the method to be applied being made at the time of execution.
2.3.System functionalities
The system is provided with a graphical interface,in which several functionalities have been de?ned for the interactive construction and the interactive execution of IP applications.In particular,they include the visu-alization of applications as trees of tasks,so that users can study the reasoning associated with any given IP plan.
In order to create a new IP plan,the user has to de-?ne his/her tasks and tools by?lling in?elds in appro-priate windows;he/she can then form links by de?ning methods and dataˉows between tasks and sub-tasks or tasks and tools.There are three ways for specifying the way to get the values of parameters for tasks and tools:computed from another task or tool,?xed once and for all,or to be required from
user.
Fig.1.Representation of an IP plan.
V.Ficet-Cauchard et al./Engineering Applications of Arti?cial Intelligence12(1999)733±747 734
When they want an application to be executed,users simply have to select the root task of the correspond-ing plan in a menu,and the plan is immediately visual-ized on screen as a schematic tree of tasks.The plan can then be executed interactively:users are required to choose between methods when several methods exist to perform some given task,and also to provide values for `user'parameters.Once the execution has been completed,they can have access to any information about tasks and tools that have actually been executed and,moreover,can visualize any intermediate image in order to assess critical points.
In addition to the creation and execution functional-ities that have just been described,the third and most original functionality integrated into the system con-sists of a second mode for creating applications through CBR.The CBR part is here to help users reuse knowledge already stored into the system,by providing a means of memorizing all cases of interest and retrieving the best one.The advantages of CBR in domains characterized by a weak or ill-structured the-ory,such as the IP domain,are manifold:
.representation of exceptions;
.allowing the use of missing or noisy data;
.solving a complex problem,through interactions between solutions of more simple problems;.dynamic learning.
The corresponding CBR module is detailed in the next two sections.
3.Case representation and similarity
A case is broadly composed of two parts:a descrip-tion of the solution and a description of the problem.In this system,a solution is represented as a TMT tree,which can be accessed through its root task.In case-based planning (Prasad,1995;Veloso et al.,1996)or case-based design (Smyth,1996),a solution is gener-ally built by combining parts of several plans,coming from several cases.In order to make this kind of de-sign possible,it was decided to associate several cases with one single plan:the ?rst case is associated with the root task of the plan,and the others with some sub-tasks of the same plan,which are considered as representative of speci?c IP techniques.In the example of Fig.3,cases are associated to tasks Ta1,Ta2and Ta4,which correspond to some speci?c strategy in IP;by contrast,no case is associated with tasks Ta3,Ta5,Ta6,Ta7and Ta8.
The problem description is composed by means of a set of discriminative criteria,which have been found from a thorough study of the IP domain.The results of this study are presented in Section 3.1;the similarity functions for comparing cases are described in Section 3.2.
3.1.Criteria for case selection
The identi?cation of a relevant set of similarity cri-teria enabling one to characterize an IP problem
is
Fig.2.Various possible links between tasks,methods and
tools.
Fig.3.Association of a set of cases with a TMT plan.
V.Ficet-Cauchard et al./Engineering Applications of Arti?cial Intelligence 12(1999)733±747735
based,on the one hand,on a study of IP systems detailed in (Ficet-Cauchard,1999)and,on the other hand,on the study of books and PhD dissertations dedicated to IP techniques (Elmoataz,1990;Russ,1995).
The major issue here is to choose an indexing voca-bulary that can be shared and accepted by any IP pro-grammer.Except for low-level actions (corresponding to operators from an IP library),there really exists no consensus on IP terms.In particular,this can be explained by the di culty of isolating oneself from the domain of application (most IP programmers work on one type of application at a time,and thus only use terms from their current domain of application).
The criteria put forward here come from a classi?-cation of the terms most often encountered in descrip-tions of IP actions and data.A distinction is made
between two broad categories of criteria:criteria re-lated to the task de?nition,and criteria related to the image description.
3.1.1.Criteria related to the task de?nition
This ?rst category includes data related to the oper-ation performed by a task and to its position in the plan in relation to other tasks.Such criteria include IP type or phase,problem de?nition and abstraction level.
IP type or phase broadly corresponds to the type of problem that is solved by a task.According to the task's abstraction level,one can take into account:.either the IP type (the root task of a complete plan de?nes a high-level processing,which belongs to an IP type Ddetection,segmentation,classi?cation ,F F F
),
Fig.4.Vertical division of a plan solving a segmentation
problem.
Fig.5.Horizontal division of a plan.
V.Ficet-Cauchard et al./Engineering Applications of Arti?cial Intelligence 12(1999)733±747
736
.or the IP phase(each sub-task of a plan de?nes one part of the complete processing,which corresponds to one speci?c stepDpre-processing,seed determi-nation,region determination,F F F).
The various IP phases correspond to a vertical division of the plan(Fig.4);for some types of problems,some phases may be optional.
The de?nition of the problem is composed of a set of keywords selected from three pre-de?ned lists:
.a list of verbs describing the operations performed by the task(detect,classify,binarize,smooth,F F F); .a list of nouns,corresponding either to objects on which the action is performed(contours,regions, image background,F F F),or to IP techniques(region growing,region division,F F F);
.a list of adjectives,qualifying either the objects on which the action is performed(small,local,F F F),or the action itself(partial,strong,F F F).
As can be noticed in previous examples,the vocabu-lary from these three lists of keywords is completely independent of the domain of application.
Finally,the abstraction levels that correspond to a horizontal division of the plan(Fig.5)are based on the abstraction levels of the automatic planner BORG (Clouard et al.,1999).
.Tasks belonging to the intentional level answer ques-tions such as`What is to be done?',and deal with IP objectives.
.Tasks belonging to the functional level answer ques-tions such as`How is it to be done?',and refer to some IP technique,leaving aside technical con-straints related to their implementation.
.Tasks belonging to the operational level answer questions such as`By what means is it to be done?', and represent IP technical know-how that can be implemented as algorithms.
3.1.2.Criteria related to the image description Among the criteria related to the images,some cor-respond to physical knowledge(related to image for-mation)and describe image quality(e.g.type of noise, amount of noise and quality of contrast).These criteria are of paramount importance for the choice of the pre-processing steps.
Other criteria correspond rather to perceptual knowledge(symbolic descriptions in terms of visual primitives).They include the presence or absence of an image background and the aspects of objects(homo-geneous gray level,light color,texture,thick boundaries, F F F).
The third group of criteria corresponds to semantic knowledge(scene analysis and components of the scene),and describes the appearance of what is to be detected,but in abstract terms,independent of the domain of application.These latter criteria include the forms of objects(convex,concave,elongated,compact, square,round,F F F),the relative sizes of objects,their positions(left,middle,right,top,bottom,center)and inter-object relations(proximity,connectivity,inclusion, F F F).
3.2.Similarity calculation between two cases
One can consider two principles for the determi-nation of similar cases,either maximized similarity (Caulier and Houriez,1995)or minimized adaptation e ort(Smyth,1996).Owing to the absence of any automatic method for evaluating IP results,the former was chosen here.First,the functions used for simi-larity calculation between a source case and a target case are described.Then comparison modes for each type of criterion are detailed.Finally,the management of missing values for a criterion is explained;in fact, as is the case in ISAC(Bonzano et al.,1997),not all the previously enumerated criteria need be taken into account in any application.
3.2.1.Similarity functions
The?rst group of criteria(i.e.criteria related to the task de?nition)is used to characterize the action per-formed by a task,and is thus closely dependent on the TMT model.Such criteria de?ne a set of tasks that can solve one`type of problem'.They are`compulsory' (each criterion of the target case must have a value), and are used to reduce the search space.A?rst simi-larity function F t using the criteria related to the task de?nition will thus be applied to reduce the set of can-didate target cases.This function is de?ned by formula (1)as the weighted average of the similarity results for each criterion:S is the source case,T is the target case,a Cr is the importance coe cient associated with criterion Cr,and j Cr(S,T)is the similarity between S and T related to criterion Cr.The result value of any j Cr function is between0(if the values of Cr in the two cases are very di erent from each other)and1 (when they are deemed identical).All a Cr coe cients also lie between0and1,in order to normalize the F t function(return values between0and1).
F t S,T
S a Cr?j Cr S,T
S a Cr
V Cr P f criteria to the task definition g 1
The second group of criteria(i.e.criteria related to the images)characterizes the objects to be detected,and depends on the current image.Such criteria are not meaningful for all applications:for instance,`contrast quality'has no sense when processing a region map.
V.Ficet-Cauchard et al./Engineering Applications of Arti?cial Intelligence12(1999)733±747737
This second group of criteria are `optional'ones (not all the criteria of the target case need to be ?lled in);they enable the user to select the nearest cases from among the candidates obtained after applying function F t .The second similarity function F i is thus used to reduce the set of selected cases,in order to get a list of reasonable size.This function is de?ned by formula (2)as the weighted average of similarity results on each criterion;notations and properties are the same as in formula (1).F i S ,T
S a Cr ?j Cr S ,T
S a Cr
V Cr P f criteria related to the image description g 2
The de?nitions of the functions j Cr that are respon-sible for the similarity calculation for each category of criterion are given in the next paragraph.The use of similarity functions F t and F i in the selection/adap-tation algorithm,as well as the adjustment of import-ance coe cients,is explained in Section 4.
3.2.2.Criterion comparison modes
It is clear that the list of criteria related to the description of an image cannot be exhaustive:the cri-teria put forward here come from a study of the IP lit-erature and the development of some particular applications.It should be completed in the course of further applications.Each criterion type is associated with a generic similarity function,in order to integrate new criteria easily.Here are the types of criteria that are presently available.
.Strictly numerical criteria:the value must be of inte-ger or real type,and a comparison between two values returns `1'when values are strictly equal and `0'otherwise.
.Strictly symbolic criteria:the value is a symbol,and a comparison between two values returns `1'when values are strictly equal and `0'otherwise (e.g.the presence of an image background).
.Gradual numerical criteria:the value belongs to integer or real intervals,and a comparison between
two values returns the di erence between the two values divided by the interval length (e.g.the relative sizes of objects).
.Gradual symbolic criteria:the value belongs to an ordered set of symbols,and a comparison between two values returns the di erence between the two values according to their order in the set,divided by the interval length (e.g.the amount of noise).
.Multi-valued criteria:the value is de?ned as a non-ordered list of symbols and/or numbers,and a com-parison between two values returns the ratio of the number of common elements in both lists to the length of the target case list (e.g.verbs used in the problem de?nition).A missing criterion value for a given case can result from several causes (no meaning,usefulness,F F F )and can be taken into account in several ways (do not take into account,consider as a speci?c value,F F F ).The line to be taken depends on whether one is considering the source case or the target one:
.The absence of a value in a target case means that the value is considered as irrelevant for this case (either it is meaningless,or it has been judged as useless by user),that absence will have no conse-quence on similarity calculation (j Cr (S ,T )=0and a Cr =0).
.The absence of a value in a source case (when this value is present in the target one)means that one similarity condition is not respected;that absence should lower the result of similarity calculation (j Cr (S ,T )=0and a Cr 60).Both conditions are respected by the set of generic functions that compute a similarity for each criterion type.
4.Recursive selection/adaptation algorithm
In the selection/adaptation process of most CBR systems,one can notice,on the one hand,the existence of a preliminary step in the selection process,aimed
at
Fig.6.Schema of the selection/adaptation process.
V.Ficet-Cauchard et al./Engineering Applications of Arti?cial Intelligence 12(1999)733±747
738
reducing the case base(Bonzano et al.,1997;Netten and Vingerhoeds,1996),and on the other hand,the fact that the selection/adaptation cycle must be applied iteratively,in particular in CBR planning(Prasad, 1995;Smyth,1996).The approach taken here(Fig.6) is also based on a selection/adaptation cycle,applied iteratively at various levels of the plan;in addition,at each cycle loop,a reduction step of the case base has been included.
Several levels of abstraction of cases are considered, but the notion of`level of abstraction'is di erent from that presented in(Bergmann and Wilke,1996):it does not reason with either concrete cases or abstract ones (the latter modeling the world in a less detailed way), but only with concrete cases that can be related,either to complete IP plans(tasks at a high level of abstrac-tion)or to parts of them(tasks at a lower level of abstraction).
Contrary to most CBR planning systems(Prasad, 1995;Smyth,1996;Veloso et al.,1996),which proceed by progressive re?nement of an abstract plan,here each tuning step produces a complete plan that can be tested and assessed.This choice is due,on the one hand,to the intention to develop a`conversational' system and not a completely automatic problem-solver, and on the other hand,to issues raised by result assessment of IP applications(no general functions exist for comparing the results produced with the desired ones).One can thus very rapidly obtain a sol-ution that will serve as a starting point for the IP expert,and the only evaluation method that can gener-ally be applied to any IP application is used:visual evaluation of output images by the expert.
The reduction of the case base can be achieved either by using criteria corresponding to strict con-straints,or by considering that two cases can only be compared when de?ned by the same set of criteria. The latter technique is not adapted to the domain being studied in this work.As a matter of fact,among the criteria related to the images,some of them bring no new knowledge about the target case,without dis-qualifying the source case.The reduction step can thus be achieved by means of function F t using the`com-pulsory'criteria related to the task de?nition,while the selection step makes use of function F i with the`op-tional'criteria related to the image description.
The objective of the CBR module is to provide some assistance to IP programmers when they are building applications,by helping them reuse solutions to pre-viously-solved problems that are somewhat analogous to their current problem.The selection/adaptation pro-cess must thus take place in co-operation with the user,according to the following algorithm.
1.Ask the user for values of criteria related to the task de?nition.
2.Determine the set S of cases matching the desired criteria by means of F t.
3.Ask the user for values of criteria related to the image description.
4.While S is not of reasonable size do:modify the weight of criteria;reduce the set S by mean of F i.
5.Ask the user to choose a case among the set S.
6.Present the plan associated with the chosen case to user and advise him/her to modify the unsuitable sub-tasks,either by re-running the algorithm,or by building it from scratch,via the interactive creation module.
Steps1and3correspond to the input of the descrip-tion of the target case.Step2is the reduction step of the case base.The selection of candidate source cases is done in Step4;Step5corresponds to the user's?nal choice.Finally,Step6consists of adapting the plan as-sociated with the selected source case to the current problem.
The principles for the selection and adaptation of cases used in this algorithm are detailed in the next two sections.
4.1.Selection of a source case
In the course of Step2of the algorithm,the re-duction of the search space consists of selecting source cases that solve the same type of problem as target case T.This corresponds to the selection of cases S such that F t(S,T)>a t where a t is a threshold?xed beforehand(as function F t returns a value between0 and1,a t is?xed to a default-value of0.5).The weights of each criterion in function F t are also?xed: the same importance is granted to all criteria.This step provides a?rst set of cases S.
So that the user can choose a case at Step5,the set of cases resulting from Step4must be of reasonable size.If the set is too small,the user's choice will lose importance,and if it is too large,the user's choice will be di cult.The iterative nature of Step4enables the user to get a set whose size can be shown as a list:he/ she can then examine each case in detail,before mak-ing the?nal choice,which deals well with the intuitive aspect that characterizes the way in which IP experts work.The modi?cation of set S at each iteration is done by means of a relaxation process,by modifying the weights of the criteria and/or the selection threshold.To implement this kind of relaxation,when the user enters the values of criteria for the target case, he/she must indicate whether the criterion is con-sidered as important or not.All importance criteria are initialized at0.5.At each iteration,the system keeps the cases S from set S such that F i(S,T)>a i where a i is the selection threshold.If the size of the resulting set is too small or too large(by default
V.Ficet-Cauchard et al./Engineering Applications of Arti?cial Intelligence12(1999)733±747739
between two and ?ve cases),the coe cients of the most important criteria are raised by 0.1,whereas those of the least important ones are lowered by 0.1for the next iteration.When it is no longer possible to modify the coe cients (coe cients of the least import-ant criteria have reached 0),if the set of source cases is still too small or too large,a second relaxation mode consisting of lowering threshold a i is applied.4.2.Interactive plan adaptation
Case adaptation by means of parts of other cases is particularly worthwhile in the domain of CBR plan-ning.In this system,a case can be adapted at several levels and in several ways:locally or globally,either by means of the CBR module,or by means of the interac-tive creation module.
The plan given as the solution to a case may only require minor local modi?cations.For instance,the parameters of an operator may need to be tuned,or an operator should be replaced by another one that better matches the current problem.This ?rst type of modi?cation can be taken into account by using the modi?cation menu of the interactive creation module.
But a plan may also require broader modi?cations,that is,necessitate the replacement of a whole sub-plan by another one.To achieve such modi?cations,Step 5of the selection/adaptation algorithm o ers a means to adapt the solution of the current case by replacing the root task of any sub-plan of the current plan by another task.The substitute task can be obtained,either by re-running the algorithm in order to retrieve a similar case,or by building it from scratch,via the interactive creation module.In the example of Fig.7,a plan is adapted along three successive steps:
.replacement of sub-plan A by sub-plan A ',which is obtained by re-running the selection algorithm;
.replacement of sub-plan B by sub-plan B ',which is built via the interactive creation module;
.transformation of tool C into tool C ',simply by changing the operator linked to tool C.This example shows the usefulness of having a recur-sive algorithm:a plan can be adapted,whatever its level within the tree of tasks (A is a high-level task,B a low-level task,C an operator)and for as long as necessary (A is replaced by A ',then A 'is adapted by replacing C by C ').Once a new plan is completed,
one
Fig.7.Adaptation of a solution plan in three steps.
V.Ficet-Cauchard et al./Engineering Applications of Arti?cial Intelligence 12(1999)733±747
740
has to decide whether the new cases associated to this plan should be added to the case library.This issue is discussed in the next section.4.3.The memorization step
Memorizing a new case should only be considered if it brings new knowledge to the base.It implies that a case must respect two conditions in order to be inte-grated:the corresponding knowledge must be correct,and it must be di erent enough from the knowledge of the cases that are already in the base.
Checking the ?rst condition consists of verifying the consistency and e ciency of the produced plan.A plan is consistent when its execution is normal,and it
is e cient if it produces satisfactory results.Consistency can be checked by the correct progress of the plan execution,while its e ciency must be assessed by the user,who is the only judge of its relevancy.The integration of new cases will thus be achieved,at the user's instigation,once the solution has been validated through a set of tests.
Several cases associated with one complete plan can be integrated into the base;in fact,if the complete plan represents the solution of a high-level problem,its various sub-plans represent solutions of problems at lower levels.When the integration of a case is required,a ?rst step consists of determining the list of plans and sub-plans that are candidates for inte-gration.This list corresponds to the plans that have been adapted,that is,the ancestors of replaced sub-plans that are large enough (at least three levels of tasks).If the substitute plan has been built via the interactive module,it will also be inserted into the list.Fig.8looks again at the plan adapted in Fig.7;the determination of the candidates for integration is achieved by examining the three replaced sub-plans..Sub-plan of root A ':D is inserted into the list;A 'is not inserted because it stems from a case of the base.
.Sub-plan of root B ':plans of roots E and F are inserted into the list;B 'has been manually built but it is not inserted because it has only two levels.
.Sub-plan of root C ':plans of roots A 'and G
are
Fig.8.Determination of candidate cases for
memorization.
Fig.9.Plan associated with the selected case with input and output images.
V.Ficet-Cauchard et al./Engineering Applications of Arti?cial Intelligence 12(1999)733±747741
inserted into the list,whereas H and C 'are not because they have less than three levels.
Then,for each plan in the list,the user has to provide values for the criteria of the corresponding cases that have been modi?ed.The system searches the case base for the most similar cases to the new cases,and inte-grates the latter if their similarity is lower than a given threshold (memorization threshold),that is,the new case is di erent enough from all the base cases,according to either the task criteria or the image cri-teria.The similarity considered here corresponds to the minimum of the similarity between task criteria and the similarity between image criteria:if that minimum is lower than the memorization threshold,this means that the case is considered as di erent enough from all the base cases,according to at least one of the two types of criteria.
5.The CBR module at work:an example
In this section,a session showing how the CBR module can be used during the creation of a new ap-plication is described.The new problem consists here of extracting objects in an image of industrial origin (image (2),Fig.9).The user begins by de?ning his/her target case through an input window:IP type is seg-mentation ,problem is de?ned as extract and object ,task's level is intentional ,amount of noise is low ,quality of contrast is medium ,there is an image background ,objects are characterized by their light grey level aspect ,convex form ,size relatively large and connec-tivity relation .Background,aspect,form and relation are considered as important by the user.
The selection algorithm is then run in two steps.The ?rst selection step keeps only cases whose similarity with regard to the target case according to the task cri-teria is higher than 0.5(i.e.cases related to a type of problem similar to the type of problem solved by the target case).Then,during the second selection step,the weights of the criteria and the selection threshold are tuned until the selection of cases according to the image criteria returns a set containing between two and ?ve cases.
In this example,a list of four cases is returned,of which the user chooses the case that seems to be the best match for his/her problem.The plan solution to the selected case can be visualized,so as to study its strategy,and it can also be executed.
The root task of the selected plan (Fig.9)is `isolate objects from background';this plan has been built for a cytology application (images (1)and (3)),for the extraction of some categories of cells.
The user can then start adapting the proposed plan to his/her new problem.The ?rst modi?cation deals with the `select background'task:in the initial plan,the problem was to isolate dark objects on a light background,whereas here the objects are light and the background is dark.The ?rst adaptation step simply consists of inverting the selection of objects (sub-plan F1,Fig.10),and is thus achieved via the interactive module.As the results after execution are still unsatis-factory (imprecise localization of contours,objects not properly separated,image (4)),the user considers a second adaptation step by re-running the selection al-gorithm in order to ?nd another sub-plan for the task `obtain objects from regions'.A new target case corre-sponding to this sub-problem is thus de?ned,the al-gorithm is re-run and the user ?nally chooses
a
Fig.10.Partial representation of plan after adaptation.
V.Ficet-Cauchard et al./Engineering Applications of Arti?cial Intelligence 12(1999)733±747
742
substitution (sub-plan F2,Fig.10).After replacement,the resulting plan (Fig.10)may further be improved by local modi?cations (e.g.the replacement of an oper-ator by another one).
Once all the adaptations have been completed,one has to de?ne the new cases to be integrated into the base.The system produces the candidates for inte-gration:they are the plans of roots `select back-ground',`obtain objects from regions',`isolate objects from background'and `eliminate background'.For these four tasks,the user is required to de?ne the cor-responding cases:two of these four cases are integrated into the base.
The assistance provided by the CBR module for the tuning of this plan shows the aptness of the selection criteria and the e ciency of the selection/adaptation algorithm:the interactive and recursive nature of this algorithm enables a satisfactory solution to be rapidly reached.However,the number of further local adap-tations that must be made reveals the sparsity of the present case base,which must now be enlarged by sys-tematically integrating all the plans and cases corre-sponding to the applications developed within the research team.
6.System validation
In this section,a few examples of applications devel-oped using the TMT system are presented.The corre-sponding IP plans served to test the system along three main axes:validation of the model and architecture,
experimentation with the interface by an inexperienced user,and searches for similarities between applications from di erent ?elds.
6.1.Validation of model and architecture
The ?rst two IP plans have enabled the authors to validate the TMT model and the system's architecture by taking actual applications into account.They were developed by A.Elmoataz (Elmoataz et al.,1996)and F.Angot (Angot,1999)to process biomedical images of cytology and histology provided by the cancer-research centre F.Baclesse of Caen.
The former (plan A)works on cytological images (Fig.11).The goal is to detect epithelial cells.A pre-cise description of the `objects'that can be found in such images was formulated in collaboration with a domain expert (e.g.image background is homogeneous and its grey level is lighter than the rest of the image).Plan A returns a region map where the various objects are labelled with di erent colours.From this region map,it is then possible to do further processing,such as eliminating objects that do not correspond to epi-thelial cells,according to size,shape or grey-level cri-teria.
Plan A was the ?rst plan that was integrated into the system.It has enabled validation of the TMT architecture by showing,on the one hand,that the de-composition of tasks into several abstraction levels gave a good representation of strategies and a good medium for dialogue between experts,and on the other hand,that the resulting plan was directly compu-tational (i.e.could be immediately executed).It has also allowed checks to be run on the functioning of the various execution modes of the tools (simple ex-ecution,multiple execution,execution until a con-straint is satis?ed).
In addition,it has enabled some functionalities of the graphical interface concerning plan creation and execution to be de?ned more precisely.Its integration has also brought to the fore the need for syntactic veri-?cation in the course of the modelling:as a matter of fact,problems due to the absence of syntactic consist-ency checking are manifested only at the time of
ex-
Fig.11.Input and output images of plan
A.
Fig.12.Input and output images of plan B.
V.Ficet-Cauchard et al./Engineering Applications of Arti?cial Intelligence 12(1999)733±747743
ecution,and it is then di cult to determine what causes them.
The second plan (plan B)was created for a histo-logical application.The goal was to detect signi?cant groups of cells,such groups suggesting the presence of tumoural lobules (Fig.12).The plan returns a region map,where each group of cells is labelled with a di er-ent colour.
As this second plan was relatively complex,it has enabled the functions for checking the syntactic con-sistency of plans to be complemented and validated.The input image of plan B is of the same type as the input image of plan A (same domain:biology,same acquisition device:microscope).The representation of both applications as TMT plans revealed the use of di erent IP techniques during the ?rst step,which cor-responds to background `elimination'(use of contrast on boundaries in plan A,and use of inter-class var-iance in plan B).As a matter of fact,as both tech-niques can be applied to plan A,a second method was added to the task `select background'of plan A (Fig.13);this additional method being a technique used in plan B.Such an operation shows,on the one hand,that it can be worthwhile to choose between several methods,in order to try various techniques for the same application,and on the other hand,that the rep-resentation of applications as TMT trees can be a means of knowledge sharing and discussion between experts.
6.2.Experimentation by a novice
A third plan (plan C)was integrated into the system by a novice,who was inexperienced both with the sys-tem and in IP.The goal of this experiment was to test whether a new user could rapidly take the system in hand,and also to validate the functionalities of the graphical interface.
Plan C was created to deal with images of human faces.The problem was set within the framework of `GDR-PRC ISIS',which is a French research group in Signal and Image Processing.It consists of localizing the inner mouth corners.The novice developed three di erent versions (C1,C2,C3)to achieve this goal.The results presented in this paper correspond to plan C1,which performs the ?rst step of the whole proces-sing,that is,extraction of the region corresponding to the teeth,and only works for open mouths (Fig.14).The area to be extracted is de?ned as a light area situ-ated in the centre of the image.
The integration of this plan by a novice has shown that the system TMT is actually easy to take in hand:even if it is not always easy to give relevant names to high-level tasks,the modelling principles appeared clear and handy.This work has also raised new issues about the validation of the integrated knowledge (need to check keyboard errors as much as possible)and about man/machine communication (inadequate voca-bulary or erroneous order of operations).This
plan
Fig.13.Addition of a second method to a sub-plan of plan
A.
Fig.14.Input and output images of plan C.
V.Ficet-Cauchard et al./Engineering Applications of Arti?cial Intelligence 12(1999)733±747
744
has also shown that the TMT system is not limited to operators in the library (although the library is quite exhaustive),as it includes a tool implemented as a C function and written on this occasion.
To become acquainted with the existing Pandore library of operators (Clouard et al.,1997),the novice ?rst implemented its application as a Shell script.Then it was modelled as a TMT plan so as highlight the underlying strategy,which demonstrated the edu-cational aspect of the approach.As a matter of fact,while the ?rst plan (C1)was built in a bottom-up man-ner,the two others (C2and C3)were developed in a top-down manner,by relying on the strategy discov-ered in the former.
6.3.Search for similarities between applications from di erent ?elds
The two last plans described below involve working on images of two di erent origins (synthesized images and industrial images).Their integration into the TMT system has enabled the determination of descriptive criteria that should be common to applications from di erent ?elds,and also extensive testing of the CBR module.
The fourth plan (plan D)was developed for testing both the TMT system and,more precisely,some oper-ators for selecting objects on the basis of shape cri-teria.It works on synthesized images,with the objective of sorting objects according to their geo-metrical shapes.The plan results in three distinct images,containing respectively rectangles,squares and ellipses (Fig.15).
This plan has enabled tests of the interweaving of tools of various types:tools calling to Pandore oper-ators and tools calling to Lisp functions.It shows that programming at the knowledge level facilitates the reuse of programming blocks written in di erent languages.Besides,this plan performs a new type of processing (detection of a speci?c shape)and works on a new image type.It has thus enabled the case base to be enriched with cases related to pattern-recognition tasks and has increased the set of indexing terms.
The ?fth plan (plan E)was entirely built using the CBR module.The goal is to isolate and separate the
objects in an industrial image.The result is a region map,where each object is labelled with a di erent col-our (Fig.16).
The ?rst case selected by the CBR module was the case associated with plan A;it was then adapted by using parts of plan B.The relevance of selection cri-teria and the e ciency of the selection/adaptation al-gorithm could thus be demonstrated:thanks to its interactive and recursive nature,this algorithm rapidly arrives at a ?rst solution.However,the number of further local adaptations that were needed,revealed the sparsity of the present case base,which must be enriched with plans performing more varied treat-ments.
6.4.Experimentation assessment
The experiments brieˉy described above have been conducted in the course of the system's design,in order to detect weaknesses as soon as possible,to determine their causes and to correct them.
Future experiments must include the integration of a wide variety of applications in di erent domains,with a view to enriching the vocabulary used for case description,and to increasing the CBR module role.In particular,the Pandore library is presently being aug-mented with image interpretation operators,which should enlarge the ?eld of investigation.The system is currently in practical use,but only its designers are playing the part of the IP experts.The system now needs to be tested in `real conditions',that is,to be validated by actual IP experts,and not only by mem-bers of the research
team.
Fig.15.Input and output images of plan
D.
Fig.16.Input and output images of plan E.
V.Ficet-Cauchard et al./Engineering Applications of Arti?cial Intelligence 12(1999)733±747745
7.Conclusions
In this paper,a CBR module providing assistance to knowledge reuse has been described.It enables an IP expert to retrieve an existing plan that solves a pro-blem similar to his/her current problem,and adapt it to the new situation.He/she can thus reuse his/her own knowledge,or knowledge previously modeled by other IP experts.The recursive selection/adaptation al-gorithm alternates between retrieval and adaptation steps,thus enabling users to build a plan by combining parts of other plans.Criteria for selecting cases are based on a de?nition of IP tasks and a description of images.
Similar ideas can be found in HICAP(Munoz-Avila et al.,1999),a general-purpose planning architecture that is applied to the planning of military evacuation operations.It is also a CBR system that can assist users during the construction of hierarchical plans of tasks.The system integrates a user-friendly task editor, conducting an interactive conversation with the user. For tasks that can be decomposed in multiple ways (i.e.problem-speci?c tasks),a case is associated with each available decomposition method(whereas in the system described here,cases are associated with tasks and not with methods).Thus in HICAP,the user has to de?ne a case in order to select each method used in the plan,which seems to be more constraining and time-consuming for the user.
In order to restrain the scope of the problem,tests have presently been limited to segmentation appli-cations.Further work will consist in diversifying the contents of the libraries(plans and cases)by integrat-ing applications dealing with more varied treatments (from image restoration to image interpretation)and applied to images from various domains.This should also enable the vocabulary used for the description of cases to be enriched,and thus complete the set of cri-teria,so as to obtain a more exhaustive list of terms. One could accordingly build a more exhaustive IP ontology including a classi?cation of terms and the de?nition of relations between them:such an ontology would support the dynamic indexing of the case base. Guidelines for achieving such an objective can be found in the work of Fensel and Benjamins(1998), who state that ontological engineering is one of the key issues for problem-solving methods reuse,and pro-poses an`ontologist'module that acts as a kind of negotiator between user and system during method selection.
In addition,one should consider means to alleviate the user's task in the course of the adaptation step.By using`simple'rules based on the comparison of some criterion values,the system could provide more assist-ance to the user by indicating which parts of the plan need adaptation.As the lack of theory in image pro-cessing does not allow such rules to be established in advance,this objective could be achieved by perform-ing an automatic search of adaptation rules.By keep-ing a history of all the adaptations performed,one could automatically extract`simple'adaptation rules that are representative of the experts'work habits.The rule premises would correspond to comparisons between values of some criteria,and their conclusions would propose some modi?cations of the solution plan.
Acknowledgements
This research on IP is carried out within the frame-work of the`Po le Traitement et Analyse d'Image,TAI, de Basse-Normandie'.This work is partially supported by a grant from the Ministry of University Education and Research.Special thanks are due to IP experts F. Angot and A.Elmoataz for their co-operation.
References
Angot,F.,1999.Segmentation d'images3D;application a la quanti-?cation d'images de tissus biologiques obtenues par microscopie confocale,PhD Thesis,Caen.
Bergmann,R.,Wilke,W.,1996.On the role of abstraction in case-based reasoning.EWCBR1996,Springer.
Bonzano,A.,Cunningham,P.,Smyth,B.,https://www.doczj.com/doc/308883697.html,ing introspective learning to improve retrieval in CBR:A case study in air tra c control.ICCBR1997,Rhode Island,USA.
Cle ment,V.,Thonnat,M.,1993.A knowledge-based approach to in-tegration of image processing procedures.In:CVGIP:Image Understanding,vol.57.Academic Press,pp.164±184. Clouard,R.,Elmoataz,A.,Angot,F.,1997.PANDORE:une bib-liothe que et un environnement de programmation d'ope rateurs de traitement d'images,Rapport interne du GREYC,Caen,France.
http://www.greyc.ismra.fr/~regis/Pandora/.
Clouard,R.,Elmoataz,A.,Porquet,C.,Revenu,M.,1999.BORG:
a knowledge-based system for automatic generation of image pro-
cessing programs.IEEE Trans.on Pattern Analysis and Machine Intelligence21(2),128±144.
Elmoataz, A.,1990.Me canismes ope ratoires d'un segmenteur d'images non de die:de?nition d'une base d'ope rateurs et im-ple mentation,The se de Doctorat,Caen.
Elmoataz,A.,Belhomme,P.,Herlin,P.,Schupp,S.,Revenu,M., Bloyet,D.,1996.Automated segmentation of cytological and his-tological images for the nuclear quanti?cation:an adaptive approach based on mathematical morphology.Microscopy, Microanalysis,Microstructure7,331±337.
Fensel,D.,Benjamins,R.,1998.Key issues for automated problem-solving methods reuse.ECAI1998,Brighton,UK,pp.63±67. Ficet-Cauchard,V.,Porquet,C.,Revenu,M.,1998.An interactive case-based reasoning system for the development of image proces-sing applications.EWCBR1998,Dublin,Ireland,pp.437±447. Ficet-Cauchard,V.,1999.Re alisation d'un syste me d'aide a la con-ception d'applications de traitement d'images:une approche base e sur le raisonnement a partir de cas,The se de Doctorat, Caen.
IRIS Explorer User's Guide,1993.Silicon Graphics Inc,Mountain View,California,No.007-1371-020.
V.Ficet-Cauchard et al./Engineering Applications of Arti?cial Intelligence12(1999)733±747 746
Munoz-Avila,H.,Aha,D.,Breslow,L.,Nau,D.,1999.HICAP:an interactive case-based planning architecture and its application to noncombatant evacuation operations.IAAI-99.
Netten,B.D.,Vingerhoeds,R.A.,1996.Structural adaptation by case combination in EADOCS,GWCBR1996,Bad Honnef, Germany.
Prasad,B.,1995.Planning with case-based structures.In:AAAI Fall Symposium,MIT Campus,Cambridge,Massachusetts. Rasure,J.,Kubica,S.,1994.In:Christensen,H.I.,Crowley,J.L.
(Eds.),The Khoros Application Development Environment, Experimental Environments for Computer Vision and Image Processing.Word Scienti?c,Singapore No.1±32.
Russ,J.C.,1995.The Image Processing Handbook,2nd ed.CRC Press.
Smyth,B.,1996.Case-based design,Doctoral Thesis of the Trinity College,Dublin,Ireland.
Veloso,M.,Munoz-Avila,H.,Bergmann,R.,1996.Cased-based planning:selected methods and systems.AI Communications9
(3).
Vale rie Cauchard received a PhD degree in Computer Science from the University of Caen in January1999.Since1995,she has been teaching Computer Science at the University of Caen and at the Engineering School of Caen(ISMRA)as a temporary lecturer.She is carrying out research at the GREYC in the Image Team.Her research interests are knowledge acquisition,representation and reuse in image processing,co-operative problem-solving architec-tures,man/machine interaction,and case-based reasoning.Christine Porquet received an Engineer Diploma from the Engineering School of Caen(ISMRA)in1983and a PhD degree from the University of Caen in1986.Since1984,she has been teach-ing Computer Science,arti?cial intelligence,and computer vision at the University of Caen and at ISMRA and became an associate pro-fessor at the ISMRA in1989.She is carrying out research at the GREYC in the Image Team.Her current research interests include knowledge-based systems in image processing and image understand-ing,co-operative problem-solving architectures,and knowledge ac-quisition and reuse.
Marinette Revenu received a Diploma in Electronic Engineering in 1969and a Doctoral Thesis degree(PhD)in Computer Science from the University of Paris6in1985.Since1978,she has been teaching arti?cial intelligence,knowledge acquisition,image processing,and general Computer Science at the Engineering School of Caen as an associate professor and now as a professor.She is also in charge of the Computer Science track in this school.Since1994,she has been at the head of the Image Team of the GREYC.This team has eight permanent researchers and ten PhD students.Her main research interests are:knowledge acquisition and representation applied to computer vision.More precisely,she is presently focused on image segmentation and interpretation.Applications concern the biomedi-cal domain,notably cell microscopy and MRI.
V.Ficet-Cauchard et al./Engineering Applications of Arti?cial Intelligence12(1999)733±747747
长春理工大学——professor——景文博——旗下出品 1基于形态学运算的星空图像分割 主要内容： 在获取星图像的过程中，由于某些因素的影响，获得的星图像存在噪声，而且星图像的背景经常是不均匀的，为星图像的分割造成了极大的困难。膨胀和腐蚀是形态学的两个基本运算。用形态学运算对星图像进行处理，补偿不均匀的星图像背景，然后进行星图像的阈值分割。 要求： 1> 图像预处理：对原始星空图像进行滤波去噪处理； 2> 对去噪后的图像进行形态学运算处理； 3> 选取自适应阈值对形态学运算处理后的图像进行二值化； 4> 显示每步处理后的图像； 5> 对经过形态学处理后再阈值的图像和未作形态学处理后再阈值的图像进行对比分析。 待分割图像直接分割图像处理后的分割图像 2基于数字图像处理的印刷电路板智能检测方法 主要内容： 通过对由相机实时获取的印刷电路板图像进行焊盘识别，从而提高电子元件的贴片质量，有效提高电路板的印刷效率。 要求： 1> 图像预处理：将原始彩色印刷电路板图像转成灰度图像，对灰度图像进行背景平滑和滤波去噪； 2> 对去噪后的图像进行图像增强处理，增强边缘提取的效果。 3> 对增强后的图像进行边缘提取（至少两种以上的边缘提取算法）； 4> 显示每步处理后的图像（原始电路板图像可自行查找）； 5> 图像处理后要求能对每个焊盘进行边缘提取，边缘清晰。 3静止背景下的移动目标视觉监控 主要内容：
基于视觉的人的运动分析最有前景的潜在应用之一是视觉监控。视觉监控系统的需求主要来自那些对安全要求敏感的场合，如银行、商店、停车场、军事基地等。通过对静止背景下的目标识别，来提醒监测人员有目标出现。 要求： 1>对原始参考图和实时图像进行去噪处理； 2>对去噪后的两幅图像进行代数运算，找出目标所在位置，提取目标，并将背景置黑； 3> 判断目标大小，若目标超过整幅图像的一定比例时，说明目标进入摄像保护区域，系统对监测人员进行提示（提示方式自选）。 4>显示每步处理后的图像； 5>分析此种图像监控方式的优缺点。 背景目标出现目标提取 4车牌识别图像预处理技术 主要内容： 车辆自动识别涉及到多种现代学科技术，如图像处理、模式识别与人工智能、计算机视觉、光学、机械设计、自动控制等。汽车作为人类生产、生活中的重要工具被广泛的使用，实现自动采集车辆信息和智能管理的车牌自动识别系统具有十分重要的意义: 要求： 1>对原始车牌图像做增强处理； 2>对增强后的彩色图像进行灰度变换； 3>对灰度图像进行直方图均衡处理； 4>选取自适应的阈值，对图像做二值化处理； 5>显示每步处理后的图像； 6>分析此种图像预处理的优缺点及改进措施，简要叙述车牌字符识别方法 原始车牌图像处理后的车牌图像 5医学细胞图像细胞分割图像增强算法研究 主要内容： 医学图象处理利用多种方法对各种图像数据进行处理，以期得到更好的显示效果以便医生根据细胞的外貌进行病变分析。 要求： 1>通过对图像的灰度变换调整改变细胞图像的灰度，突出感兴趣的细胞和细胞核区域。 2>通过直方图修改技术得到均衡化或规定化等不同的处理效果。 3>采用有效的图像平滑方法对细胞图像进行降噪处理，消除图像数字化和传输时所混入的噪声，提高图像的视觉效果。 4>利用图像锐化处理突出细胞的边缘信息，加强细胞的轮廓特征。 5>显示每步处理图像，分析此种细胞分割图像预处理方法的优缺点。 原始细胞图像 图像处理后的细胞图像 6瓶子灌装流水线检测是否液体灌装满瓶体 当饮料瓶子在罐装设备后要进行液体的检测，即：进行判断瓶子灌装流水线是否灌装满瓶体的检测，如液面超过瓶颈的位置，则装满，否则不满，如果不满则灌装液体不合格，需重新进行灌装。 具体要求： 1）将原进行二值化 2）二值化后的图像若不好，将其滤波再进行膨胀处理,并重新进行二值化
数字图像处理实验指导书 自动化工程学院 2014年11月试行
实验一数字图像的获取 一．实验目的 1. 熟悉MATLAB软件的操作环境，掌握基本的程序调试方法和简单的编程； 2．学会读出MATLAB目录下图像文件并进行旋转、加亮、取反、多幅图像显示等基本操作。 二．实验内容 1.读取和显示 I=imread(‘rice.png’);%读取图像 imshow(I) ; 2.直方图 在matlab环境中，程序首先读取图像，然后调用直方图函数，设置相关参数，再输出处理后的图像。 I=imread('cameraman.tif');%读取图像 subplot(1,2,1),imshow(I) %输出图像 title('原始图像') %在原始图像中加标题 subplot(1,2,2),imhist(I) %输出原图直方图 title('原始图像直方图') %在原图直方图上加标题 3.加亮 RGB=imread(‘peppers.png’);%读取图像 RGB2=imadd(RGB,50); subplot(2,2,1);imshow(RGB); subplot(2,2,2);imshow(RGB2); 4.取反 I=imread(‘rice.png’); J=imcomplement(I); Imshow(J); 5.翻转 I = imread('cameraman.tif'); １
figure,imshow(I); theta = 30; K = imrotate(I,theta); % Try varying the angle, theta. figure, imshow(K) 6.图片大小 I=imread(‘rice.png’); [m,n]=size(I) 7.缩放 在matlab环境中，程序首先读取图像，然后调用图像缩放函数，设置相关参数，再输出处理后的图像。 I = imread('cameraman.tif'); figure,imshow(I); scale = 0.5; J = imresize(I,scale); figure,imshow(J); 8.图像二值化操作 I=imread('rice.png'); J=im2bw(I,0.4); imshow(J) 9.图像合成 (1)图像加运算 i=imread('cameraman.tif') j=imread('rice.png') k=imadd(i,j); imshow(k) 注意：两幅图像的大小和类型必须是相同的。 （2）图像减运算 i=imread('AT3_1m4_01.tif'); j=imread('AT3_1m4_02.tif'); K=imsubtract(i,j);%函数将负值截断为0，显示效果为黑，可用于探查同一场景的多幅图像的变化。 imshow(K) ２
2013-2014年第一学期《数字图像处理》科目考查卷 专业：通信工程班级：任课教师：王新新 姓名：学号：成绩： 一 Deblurring Images Using the Wiener Filter ——使用维纳滤波器进行图像去模糊简介 在人们的日常生活中，常常会接触很多的图像画面，而在景物成像的过程中有可能出现模糊，失真，混入噪声等现象，最终导致图像的质量下降，我们现在把它还原成本来的面目，这就叫做图像还原。引起图像的模糊的原因有很多，举例来说有运动引起的，高斯噪声引起的，斑点噪声引起的，椒盐噪声引起的等等，而图像的复原也有很多，常见的例如逆滤波复原法，维纳滤波复原法，约束最小二乘滤波复原法等等。它们算法的基本原理是，在一定的准则下，采用数学最优化的方法从退化的图像去推测图像的估计问题。因此在不同的准则下及不同的数学最优方法下便形成了各种各样的算法。而我接下来要介绍的算法是一种很典型的算法，维纳滤波复原法。它假定输入信号为有用信号与噪声信号的合成，并且它们都是广义平稳过程和它们的二阶统计特性都已知。维纳根据最小均方准则，求得了最佳线性滤波器的的参数，这种滤波器被称为维纳滤波器。 维纳滤波器是最小均方差准则下的最佳线性滤波器，它在图像处理中有着重要的应用。本文主要通过介绍维纳滤波的结构原理，以及应用此方法通过MATLAB 函数来完成图像的复原。关键词：维纳函数、图像复原。
二维纳滤波器结构 维纳滤波自身为一个FIR或IIR滤波器，对于一个线性系统，如果其冲击响应为h（n），则当输入某个随机信号x (n)时， 式（1） 这里的输入 式（2） 式中s(n)代表信号，v(n)代表噪声。我们希望这种线性系统的输出是尽可能地逼近s(n)的某种估计，并用s^(n)表示，即 式（3） 因而该系统实际上也就是s(n)的一种估计器。这种估计器的主要功能是利用当前的观测值 x(n)以及一系列过去的观测值x(n-1),x(n-2),……来完成对当前信号值的某种估计。维纳滤波属于一种最佳线性滤波或线性最优估计，是一最小均方误差作为计算准则的一种滤波。设信号的真值与其估计值分别为s(n)和) s^(n)，而它们之间的误差 式（4） 则称为估计误差。估计误差e(n)为可正可负的随机变量，用它的均方值描述误差的大小显然更为合理。而均方误差最小，也就是 式（5） 最小。利用最小均方误差作为最佳过滤准则比较方便，它不涉及概率的描述，而且以它导出的最佳线性系统对其它很广泛的一类准则而言是属最佳。 图1 维纳滤波器一般结构
《数字图像处理》实验指导书 编写: 罗建军 海南大学三亚学院 10月
目录 一、概述 ....................................................................... 错误!未定义书签。 二、建立程序框架 ....................................................... 错误!未定义书签。 三、建立图像类 ........................................................... 错误!未定义书签。 四、定义图像文档实现图像读/写.............................. 错误!未定义书签。 五、实现图像显示 ....................................................... 错误!未定义书签。 六、建立图像处理类................................................... 错误!未定义书签。 七、实现颜色处理功能............................................... 错误!未定义书签。 (一) 亮度处理................................................................. 错误!未定义书签。 (二) 对比度处理............................................................. 错误!未定义书签。 (三) 色阶处理................................................................. 错误!未定义书签。 (四) 伽马变换................................................................. 错误!未定义书签。 (五) 饱和度处理............................................................. 错误!未定义书签。 (六) 色调处理................................................................. 错误!未定义书签。 八、实现几何变换功能............................................... 错误!未定义书签。 (一) 图像缩放................................................................. 错误!未定义书签。 (二) 旋转......................................................................... 错误!未定义书签。 (三) 水平镜像................................................................. 错误!未定义书签。 (四) 垂直镜像................................................................. 错误!未定义书签。 (五) 右转90度................................................................. 错误!未定义书签。 (六) 左转90度................................................................. 错误!未定义书签。 (七) 旋转180度............................................................... 错误!未定义书签。 九、实现平滑锐化功能............................................... 错误!未定义书签。 十、图像处理扩展编程............................................... 错误!未定义书签。
数字图像处理技术在识别领域的应用 1、定义 数字图像处理是利用计算机对图像进行处理，常用的方法技术有去除噪声、复原、增强、分割、提取特征等。数字图像发展初期，主要应用于提高图片质量，第一次应用该技术是对伦敦和纽约之间海底电缆发送的图片进行改善。图像处理的应用领域涉及到人类生活的方方面面。 2、数字图像处理的优点 数字图像处理应用于人类依靠图像获取外界的信息经过处理，具有如下优点：①重现性能好，数字图像处理在进行传输、存储、复制等处理从而用来服务于生活。 ②数字化处理精度高。 ③数字信号处理技术适用面宽。 ④数字图像处理的灵活性高。 3、主要研究内容 数字图像处理技术是利用计算机图像处理系统对图像进行输入、加工和输出，主要研究的内容包括以下几项：图像变换；图像增强和复原；图像编码压缩；图像分割。 因为数字图像处理技术应用太过广泛，我在这里仅探究它在识别领域的应用，从这里认识数字图像处理技术的方方面面。 4、数字图像处理在识别领域的应用： （一）数字图像处理在指纹识别中的应用 传统的利用密码、证件作为身份识别的方式具有易遗忘、易破解、易丢失、易伪造等特点，已不再符合现代数字社会的需求。指纹，作为人体独一无二的生理特征，虽然只是人体皮肤的一小部分，但是它的纹理复杂度可以提供用于识别的足够特征，具有极高的安全性，并且指纹还具有易获取、无侵犯性、唯一性和不变性等优点，使其成为生物识别技术中的焦点。 为了弥补指纹图像的质量缺陷，保证指纹后处理算法对指纹图像具有足够的鲁棒性，图像增强是十分必要的，采用数字图像处理则可以实现图像的增强。指纹图像增强目的是为了消除噪声，增强脊线和谷线的对比度，将断裂的脊线和谷线连接起来，消除由于噪声、变形等带来的粘连及由于油污等产生的毛刺等，改善图像质量，保证特征信息提取的准确性和可靠性。指纹图像的增强由图像规格化、图像再处理、滤波几个部分组成。
毕业论文声明 本人郑重声明： 1．此毕业论文是本人在指导教师指导下独立进行研究取得的成果。除了特别加以标注地方外，本文不包含他人或其它机构已经发表或撰写过的研究成果。对本文研究做出重要贡献的个人与集体均已在文中作了明确标明。本人完全意识到本声明的法律结果由本人承担。 2．本人完全了解学校、学院有关保留、使用学位论文的规定，同意学校与学院保留并向国家有关部门或机构送交此论文的复印件和电子版，允许此文被查阅和借阅。本人授权大学学院可以将此文的全部或部分内容编入有关数据库进行检索，可以采用影印、缩印或扫描等复制手段保存和汇编本文。 3．若在大学学院毕业论文审查小组复审中，发现本文有抄袭，一切后果均由本人承担，与毕业论文指导老师无关。 4.本人所呈交的毕业论文，是在指导老师的指导下独立进行研究所取得的成果。论文中凡引用他人已经发布或未发表的成果、数据、观点等，均已明确注明出处。论文中已经注明引用的内容外，不包含任何其他个人或集体已经发表或撰写过的研究成果。对本文的研究成果做出重要贡献的个人和集体，均已在论文中已明确的方式标明。 学位论文作者（签名）： 年月
关于毕业论文使用授权的声明 本人在指导老师的指导下所完成的论文及相关的资料（包括图纸、实验记录、原始数据、实物照片、图片、录音带、设计手稿等），知识产权归属华北电力大学。本人完全了解大学有关保存，使用毕业论文的规定。同意学校保存或向国家有关部门或机构送交论文的纸质版或电子版，允许论文被查阅或借阅。本人授权大学可以将本毕业论文的全部或部分内容编入有关数据库进行检索，可以采用任何复制手段保存或编汇本毕业论文。如果发表相关成果，一定征得指导教师同意，且第一署名单位为大学。本人毕业后使用毕业论文或与该论文直接相关的学术论文或成果时，第一署名单位仍然为大学。本人完全了解大学关于收集、保存、使用学位论文的规定，同意如下各项内容：按照学校要求提交学位论文的印刷本和电子版本；学校有权保存学位论文的印刷本和电子版，并采用影印、缩印、扫描、数字化或其它手段保存或汇编本学位论文；学校有权提供目录检索以及提供本学位论文全文或者部分的阅览服务；学校有权按有关规定向国家有关部门或者机构送交论文的复印件和电子版，允许论文被查阅和借阅。本人授权大学可以将本学位论文的全部或部分内容编入学校有关数据库和收录到《中国学位论文全文数据库》进行信息服务。在不以赢利为目的的前提下，学校可以适当复制论文的部分或全部内容用于学术活动。 论文作者签名：日期： 指导教师签名：日期：
数字图像处理课程论文图像复原算法研究 学院：信息科学与工程学院 专业：通信工程 姓名： 学号： 任课教师：
2017年5月 摘要 数字图像恢复是数字图像处理的一个基本的和重要的课题，它是后期图像处理的前提。图像在获取、上传、保存的过程中不可避免地引起图像退化和图像质量的下降，图像恢复就是试图利用退化过程的先验知识使已退化的图像恢复本来面貌。本论文主要研究引起退化的环境因素，建立相应的数学模型，并沿着使图像降质的逆过程恢复图像。本文首先对测试图像进行模糊及加噪处理，在已知系统退化模型的情况下，对观测图像分别使用逆滤波、维纳滤波、有约束的最小二乘方滤波算法进行复原，并比较它们的处理效果。在这几种算法的参数选取上得到了丰富的经验数据，并对实验结果进行了分析总结。发现维纳滤波较约束最小二乘法滤波效果要好，这是因为前者利用了原图像的统计信息，采用了真实的PSF函数来恢复。无论何种算法，它们都要依据获取的相关信息才能有效地实施，算法利用的信息越多，信息的准确性越高，复原图像的质量也就越高。 关键词:图像复原;逆滤波;维纳滤波;有约束的最小二乘方滤波
一、引言 MATLAB 语言是由美国MathWorks公司推出的计算机软件，经过多年的逐步发展与
不断完善，现已成为国际公认的最优秀的科学计算与数学应用软件之一，是近几年来在国外广泛流行的一种可视化科学计算软件。它集数值分析、矩阵运算、信号处理和图形显示于一体，构成了一个方便的、界面友好的用户环境，而且还具有可扩展性特征。MathWorks 公司针对不同领域的应用，推出了信号处理、控制系统、神经网络、图像处理、小波分析、鲁棒控制、非线性系统控制设计、系统辨识、优化设计、统计分析、财政金融、样条、通信等30 多个具有专门功能的工具箱，这些工具箱是由该领域的学术水平较高的专家编写的，无需用户自己编写所用的专业基础程序，可直接对工具箱进行运用。同时，工具箱的函数源程序也是开放性的，多为M 文件，用户可以查看这些文件的代码并进行更改，MALAB 支持用户对其函数进行二次开发，用户的应用程序也可以作为新的函数添加到相应的工具箱中。MATLAB 中的数字图像是以矩阵形式表示的，这意味着MATLAB 强大的矩阵运算能力用于图像处理非常有利，矩阵运算的语法对MATLAB 中的数字图像同样适用。本文对MATLAB 图像处理工具箱进行探索及应用，实验证明该软件功能强大，语言简洁易学，人机界面友好，工具箱具有丰富的技术支持并集成了该领域专家的智慧，应用简单而效果良好。 二、实验原理 2.1图像退化模型构建 图像复原处理是建立在图像退化的数学模型基础上的，这个退化数学模型能够反映图像退化的原因。图像的退化过程可以理解为施加于原图像上的运算和噪声两者联合作用的结果，图像退化模型如图1所示，可以表示为： g ( x, y)H [ f ( x, y)]n( x, y)f ( x, y)h( x, y) n( x, y) （1）
实验一 Matlab图像处理工具箱的初步练习 一、实验目的和任务 1、初步了解与掌握MA TLAB语言的基本用法； 2、掌握MA TLAB语言中图象数据与信息的读取方法； 3、掌握在MA TLAB语言中图像类型的转换。 二、实验仪器、设备及材料 1、计算机 2、MatLab软件/语言包括图像处理工具箱(Image Processing Toolbox) 3、实验所需要的图片 三、实验原理 将数字图像的RGB表示转换为YUV表示； Y=0.30R+0.59G+0.11B U=0.70R-0.59G-0.11B V=-0.30R-0.59G+0.89B 四、实验步骤 1、阅读资料并熟悉MatLab的基本操作 2、读取MATLAB中的图象数据 3、显示MATLAB中的图象文件。用MATLAB在自建的文件夹中建立m文件，在这个文件的程序中，将MA TLAB目录下work文件夹中的tree.tif 图象文件读出，用到imread，imfinfo等命令，观察一下图象数据，了解一下数字图象在MA TLAB中的处理就是处理一个矩阵的本质。 4、将3中的图象显示出来（用imshow）。 5、对MA TLAB目录下work文件夹中的flowers.tif进行真彩色图像、索引色图像、灰度图像、二值图像之间的相互变换，并显示。 6、进行真彩色图像RGB（lenacolor.jpg）、YIQ图像、HSV图像、YcbCr图像的相互转换，并显示。 五、实验报告要求 1、描述实验的基本步骤； 2、用图片给出步骤4、5、6中取得的实验结果； 六、实验所需图片
lenacolor.jpg 七、实验注意事项 1、学生应提前预习 2、请大家在E盘建一个目录（matlab），在每次启动时都要将这个目录加入到MATLAB的搜索路径中，添加的方法为File----Set Path----Tool---Add Path 八、思考题 1、图像之间转换的基础是什么，为什么可以实现相互的转换 九、附录 MATLAB简介 (1) MATLAB全称是Matrix Laboratory（矩阵实验室），一开始它是一种专门用于矩阵数值计算的软件，从这一点上也可以看出，它在矩阵运算上有自己独特的特点。实际上MATLAB中的绝大多数的运算都是通过矩阵这一形式进行的。这一特点也就决定了MA TLAB在处理数字图像上的独特优势。理论上讲，图像是一种二维的连续函数，然而在计算机上对图像进行数字处理的时候，首先必须对其在空间和亮度上进行数字化，这就是图像的采样和量化的过程。二维图像进行均匀采样，就可以得到一幅离散化成M×N样本的数字图像，该数字图像是一个整数阵列，因而用矩阵来描述该数字图像是最直观最简便的了。而MATLAB的长处就是处理矩阵运算，因此用MA TLAB处理数字图像非常的方便。MATLAB支持五种图像类型，即索引图像、灰度图像、二值图像、RGB图像和多帧图像阵列；支持BMP、GIF、HDF、JPEG、PCX、PNG、TIFF、XWD、CUR、ICO等图像文件格式的读，写和显示。MATLAB对图像的处理功能主要集中在它的图像处理工具箱（Image Processing Toolbox）中。图像处理工具箱是由一系列支持图像处理操作的函数组成，可以进行诸如几何操作、线性滤波和滤波器设计、图像变换、图像分析与图像增强、二值图像操作以及形态学处理等图像处理操作。 1、MATLAB中图象数据的读取 A、imread imread函数用于读入各种图象文件，其一般的用法为 [X，MAP]=imread(‘filename’,‘fmt’) 其中，X，MAP分别为读出的图象数据和颜色表数据，fmt为图象的格式，filename为读取的图象文件（可以加上文件的路径）。 例：[X，MAP]=imread（’flowers.tif’,’tif’）;
数字图像处理应用论文数字图像处理技术论文 关于数字图像处理及其应用的研究 摘要：首先对数字图像处理的关键技术以及相应的处理设备进行详细的探讨，然后对数字图像处理的应用领域以及发展趋势进行详尽论述。 关键词：数字图像处理：关键技术；应用领域 0 引言 人类通过眼、耳、鼻、舌、身接受信息，感知世界。约有75％的信息是通过视觉系统获取的。数字图象处理是用数字计算机处理所获取视觉信息的技术，上世纪20年代Bartlane电缆图片传输系统(纽约和伦敦之间海底电缆)传输一幅图片所需的时间由一周多减少到小于3个小时；上世纪50年代，计算机的发展，数字图像处理才真正地引起人们的巨大兴趣；1964年，数字图像处理有效地应用于美国喷气推进实验室(J.P.L)对“徘徊者七号”太空船发回的大批月球照片的处理；但是直到上世纪六十年代末至七十年代扔，由于离散数学理论的创立和完善，使之形成了比较完整的理论体系，成为一门新兴的学科。数字图像处理的两个主要任务：如何利用计算机来改进图像的品质以便于人类视觉分析；对图像数据进行存储、传输和表示，便于计算机自动化处理。图像处理的范畴是一个受争论的话题，因此也产生了其他的领域比如图像分析和计算机视觉等等。
1 数字图像处理主要技术概述 不论图像处理是基于什么样的目的，一般都需要通过利用计算机图像处理对输入的图像数据进行相关的处理，如加工以及输出，所以关于数字图像处理的研究，其主要内容可以分为以下几个过程。图像获取：这个过程基本上就是把模拟图像通过转换转变为计算机真正可以接受的数字图像，同时，将数字图像显示并且体现出来(例如彩色打印)。数据压缩和转换技术：通过数据压缩和数据转换技术的研究，减少数据载体空间，节省运算时间，实现不同星系遥感数据应用的一体化。图像分割：虽然国内外学者已提出很多种图像分割算法，但由于背景的多变性和复杂性，至今为止还没有一种能适用于各种背景的图像分割算法。当前提出的小波分析、模糊集、分形等新的智能信息处理方法有可能找到新的图像分割方法。图像校正：在理想情况下，卫星图像上的像素值只依赖于进入传感器的辐射强度；而辐射强度又只与太阳照射到地面的辐射强度和地物的辐射特性(反射率和发射率)有关，使图像上灰度值的差异直接反映了地物目标光谱辐射特性的差异，从而区分地物目标。图像复原，以图像退化的数学模型为基础，来改善图像质量表达与描述，图像分割后，输出分割标记或目标特征参数；特征提取：计算描述目标的特征，如目标的几何形状特征、统计特征、矩特征、纹理特征等。图像增强：显示图像中被模糊的细节。或是突出图像中感兴趣的特征。图像识别：统计模式识别、模糊模式识别、人工神经网络等。
华东交通大学理工学院课程设计报告书 所属课程名称数字图像处理期末论文分院电信分院专业班级14 计科 学号20140210440214 学生姓名习俊 指导教师熊渊 2016 年12 月13 日
摘要 数字图像处理是用计算机对图像信息进行处理的一门技术，主要是为了修改图形，改善图像质量，或是从图像中提起有效信息，还有利用数字图像处理可以对图像进行体积压缩，便于传输和保存。本文论述了用Matlab编程对数字图像进行图像运算的基本方法。图像运算涵盖了MA TLAB程序设计、图像点运算、代数运算、几何运算等基本知识及其应用（点运算是图象处理的一个重要运算）。以及对图像加入噪声、图像缩放和图像旋转。 关键词图像点运算；代数运算；几何运算；图像缩放；图像旋转
目录 绪论 第一章图像运算 2.1点运算 2.2代数运算 2.3几何运算 第二章程序设计与调试 结束语 参考文献
绪论 早期的计算机无论在计算速度或存储容量方面，难于满足对庞大图像数据进行实时处理的要求。随着计算机硬件技术及数字化技术的发展，计算机、内存及外围设备的价格急剧下降，而其性能却有了大幅度的提高。 图像信息是人类获得外界信息的主要来源，数字图像处理技术越来越多的应用于人们日常工作、学习和生活中。和传统图像处理相比，它具有精度高、再观性好、通用性和灵活性强等特点。在近代科学研究、军事技术、工农业生产、医学、气象及天文学等领域中也得到了广泛应用。 近几年来，随着计算机和各个相关领域研究的迅速发展，科学计算可视化、多媒体技术等研究和应用的兴起，数字图像处理从1个专门领域的学科，变成了1种新型的科学研究和人机界面的工具。数字图像作为一门新兴技术，它是二十一世纪五十年代数字计算机发展到相当水平后开拓出来的计算机应用新领域，它把图像转换成数据矩阵存放于计算机中，并进行滤波、增强、删除等处理，包括图像输入输出技术、图像分析、变换于处理技术以及图像识别和特征提取等方面。六十到七十年代数字处理技术的理论和方法更加完善，其准确性、灵活性和通用性逐步提高。 在日常生活中，电脑人像艺术，电视中的特殊效果，自动售货机钞票的识别，邮政编码的自动识别和利用指纹、虹膜、面部等特征的身份识别等均是图像处理的广泛应用。 进行数字图像处理时主要涉及数字图像点运算处理，针对图像的像素进行加、减、乘、除等运算，有效地改变了图像的直方图分布。
研究生课程考核试卷 （适用于课程论文、提交报告） 科目：数字图像处理教师：黄鸿 姓名：潘世强学号：20110802096 专业：仪器科学与技术类别：（学术）上课时间：2011年10月至2012年01月 考生成绩： 阅卷评语： 阅卷教师(签名) 重庆大学研究生院制
CHONGQING UNIVERSITY 数字图像处理 ——基于内容的图像检索系统 学院：光电工程学院 姓名：潘世强 学号：20110802096 指导教师：黄鸿 时间： 2012年01月08日
基于内容的图像检索系统 摘要：随着多媒体技术的迅速发展，图像数据库也急剧膨胀起来，如何高效、快速地从像资源中获取有用的图像成了信息检索技术研究的热点。 本文主要针对基于内容的图像检索技术（CBIR）做了相关的介绍，对基于图像检索技术中的特征提取技术进行了较为详细的阐述，研究了图像颜色的提取方法，以及图像间相似性度量方法。本文运用的特征值提取方法为颜色直方图的方法，对图像提取颜色特征，并根据这些特征对目标图片与图片库中的图片进行了相似度排序，最后运用Matlab软件对上述方法进行验证，得到图像检索结果，从而实现基于内容的图像检索。 关键词：直方图HIS彩色空间基于内容图像检索 1．引言 图像是对客观对象的一种相似性的、生动性的描述或写真。或者说图像是客观对象的一种表示，它包含了被描述对象的有关信息。它是人们最主要的信息源。据统计，一个人获取的信息大约有75%来自视觉。俗话说“百闻不如一见”，“一目了然”，都反映了图像在信息传递中的独特效果。所谓基于图像内容检索，即从图像库中查找含有特定目标的图像，也包括从连续的视频图像中检索含有特定目标的视频片段。它区别于传统的图像检索手段，融合了图像理解技术，从而可以提供更有效的检索手段。 本文主要针对基于内容的图像检索技术中的特征提取方法展开论述，简要地介绍了近年来基于内容的图像检索中颜色、纹理、形状及语义特征的描述方法，并对颜色特征的描述方法以及特征相似性做了详细的论述。
1 选题 课程论文选题如下，每人任选一题，题目自拟，本学期6月3日前交至计算机学院411办公室。 1.图像XX增强方法综述与MATLAB实现（至少3种） 2.图像增强方法的深入研究（学习一种或两种课本上没有的图像平 滑/锐化方法与课本上介绍的进行对比研究）（需实验） 3.图像XX特征分析方法综述与MATLAB实现（至少3种） 4.结合人脸图像讨论各种图像特征分析方法的适用性（需实验） 5..灰度共生矩阵与灰度差分直方图在图像处理中实际应用（需实验） 6.不同图像分割方法的分析与比较（需实验） 7.基于数字图像处理的森林火灾识别方法研究 基于摄像机摄取的视频图像对现场进行火灾的自动探测、监视，同时将摄得的图像，利用各种图像处理技术不断进行图像处理和分析，通过早期火灾的图像变化特征来探测火灾是否发生。 测试要求：首先从彩色摄像机获取视频流图像，并转换成BMP格式图像，先判断图像中有红色区域存在。 l)火灾图像预处理，包括图像抽样、图像分割、图像灰度化、二值化、图像平滑处理; 2)研究火焰目标的特征提取方法 (l)轮廓特征提取:该模块主要功能为提取火焰轮廓上的尖点特征和圆形度。在火焰轮廓特征图中，从下至上从左至右逐点扫描，将火焰的边缘编成链码。当链码在一定步数内，出现一次有效上升和一次有效下降时，我们就得到一个尖角。 (2)颜色特征提取:火焰一般从焰心到外焰其颜色应从白色到黄色再向红 色移动，在图像中表现为像素值的变化不明显，可以用图像像素方差值来反映这种变化。 8.基于数字图像处理的答题卡识别方法 9.车牌识别方法研究（要求本地苏L车牌照）
2 格式要求 （1）页面设置： A4纸，页边距正常（上、下各2cm，左3cm、右2.0cm）， 页码（页面底端居中，小五号，Times New Roman字体）， 装订线：0.5厘米，装订位置：左侧3、7两颗钉（2）题目： 不多于30字，黑体、小三号、不加粗、居中排列，1.25倍 行距，段前断后各空0.5行。 （3）内容： 不少于5000字，宋体，小四，不加粗，1.25倍行距，段前 空2字符。 （4）标题要求： 一级标题：小三号、宋体、加粗，段前断后各空0.5行 二级标题：四号、宋体、加粗，段前断后各空0.5行 三级标题：小四号、宋体、加粗，段前断后各空0.5行 四级标题：小四号、宋体、不加粗，段前断后各空0.5行 图片要求：图片嵌入到文字中，文字不环绕，图片居中，图 标题为宋体五号字，不加粗 表格要求：三线表，表标题及表中文字为宋体五号字，不加 粗 （5）参考文献： 不少于3篇，宋体五号字，不加粗，1.0倍行距，段前不空
江苏科技大学 数字图像处理本科生课程论文 论文题目：图像增强方法综述与matlab实现完成时间：___2016年6月2日________ 所在专业：____软件工程____________ ____ 所在年级：____13419042___________ __
图像增强方法综述与matlab实现 软件工程专业 1341904222 陆建伟 摘要：本文介绍图像增强的内容，并就内部几种方法进行更深一步的探索，利用matlab 使得算法实现并对比。 关键词：图像增强；数字图像处理；灰度变换；直方图；matlab； 一、研究背景 1.1研究目的 经过图像的传送和转换，如成像、复制、扫描、传输和显示等，经常会造成图像质量的下降。光学系统的失真、相对运动、大气流动等都会使图像模糊，传输过程中会引入各种类型的噪声。总之输入的图像在视觉效果和识别方便性等方面可能存在诸多问题。通过本课题的研究能够使图像有更好的视觉感受效果，更能够满足社会生活和生产的需要是本文的最终目的。 1.2研究现状 计算机图像处理的发展历史并不长,但是引起了人们的足够重视。总体来说，图像处理技术的发展大致经历了初创期、发展期、普及期和实用化期4 个阶段。随着对图像技术研究的不断深入和发展,新的图像增强方法不断出现。图像作为自然界景物的客观反映是人类感知世界的视觉基础也是人类获取信息、表达信息和传递信息的重要手段。 二、主要理论概况 图像增强是指根据特定的需要突出图像中的重要信息同时减弱或去除不需要的信息。从不同的途径获取的图像通过进行适当的增强处理可以将原本模糊不清甚至根本无法分辨的原始图像处理成清晰的富含大量有用信息的可使用图像有效地去除图像中的噪声、增强图像中的边缘或其他感兴趣的区域从而更加容易对图像中感兴趣的目标进行检测和测量。处理后的图像是否保持原状已经是无关紧要的了不会因为考虑到图像的一些理想形式而去有意识的努力重现图像的真实度。图像增强的目的是增强图像的视觉效果将原图像转换成一种更适合于人眼观察和计算机分析处理的形式。它一般要借助人眼的视觉特性以取得看起来较好地视觉效果很少涉及客观和统一的评价标准。图像增强的方法可以大致分为两类，一类是空域处理方法，一类是频域的处理法1
数字图像处理结课作业 --数字图像频域增强方法 及在matlab中的实现 学生姓名： 学号： 学院：理学院 班级：电科班 指导教师：
摘要:图像增强的目的是使处理后的图像更适合于具体的应用，即指按一定的需要突出一幅图像中的某些信息，同时削弱或去除某些不需要的信息，使之改善图像质量，加强图像判读和识别效果的处理技术。从总体上可以分为两大类:空域增强和频域增强。频域处理时将原定义空间中的图像以某种形式转换到其他空间中，利用该空间的特有性质方便的进行图像处理。而空域增强是在图像空间中借助模板对图像进行领域操作，处理图像每一个像素的取值都是根据模板对输入像素相应领域内的像素值进行计算得到的。空域滤波基本上是让图像在频域空间内某个范围的分量受到抑制，同时保证其他分量不变，从而改变输出图像的频率分布，达到增强图像的目的。本文主要从空域展开图像增强技术，重点阐明数字图像增强处理的基本方法，介绍几种空域图像增强方法。 关键词：图像增强 MATLAB 空域增强锐化空间滤波平滑空间滤波
目录： 1、何为数字图像处理及MATLAB的历史 2、空间域图像增强技术研究的目的和意义 3、空间域的增强 3.1 背景知识 3.2 空间域滤波和频域滤波之间的对应关系 3.3 锐化滤波 3.4 平滑滤波 4、结论 1、何为数字图像处理及MATLAB的历史 数字图像处理（digital image processing）,就是利用数字计算机或者其他数字硬件,对从图像信息转换而得到的电信号进行某些数学运算,以提高图像的实用性。例如从卫星图片中提取目标物的特征参数,三维立体断层图像的重建等。总的来说,数字图像处理包括运算、几何处理、图像增强、图像复原、图像形态学处理、图像编码、图像重建、模式识别等。目前数字图像处理的应用越来越广泛,已经渗透到工业、医疗保健、航空航天、军事等各个领域,在国民经济中发挥越来越大的作用。 MATLAB是由美国Math Works公司推出的软件产品。MATLAB是“Matric Laboratory”的缩写,意及“矩阵实验室”。MATLAB是一完整的并可扩展的计算机环境,是一种进行科学和工程计算的交互式程序语言。它的基本数据单元是不需要指定维数的矩阵,它可直接用于表达数学的算式和技术概念,而普通的高级语言只能对一个个具体的数据单元进行操作。它还是一种有利的教学工具,它在大学的线性代数课程以及其它领域的高一级课程的教学中,已成为标准的教学工具。
计算机科学与信息学院 《数字图像处理》 实验指导书 适用专业：信息安全、网络工程、计算机 贵州大学 二O一三年五月
前言 本指导书是根据数字图像处理教学大纲和实验大纲编写的，在教学过程中指导学生实验时使用。运用MATLAB软件平台，结合图像处理工具箱，对图像处理相关算法进行编程和实现。通过学生上机操作实践与教师指导，使学生深入理解和掌握数字图像处理的技术和方法，增强处理实际问题的能力。 考虑到《数字图像处理》课程的自身特点，以及软件的升级更新性，本实验指导书具有适应性。 本实验指导书主要适用于计算机科学与信息学院的各个相关专业。
目录 实验一图像基本操作 (4) 实验二图像增强 (7) 实验三图像分割 (11) 实验四汽车牌照自动识别 (16) 实验报告的基本内容及要求 (18) 贵州大学实验报告 (19)
实验一图像基本操作 实验学时：2 实验类型：验证 实验要求：必做 一、实验目的 利用MATLAB软件，熟悉图像的数据矩阵操作、图像的类型转换及图像的存储等基本操作。 1.熟悉图像矩阵的基本操作 2.掌握图像数据类型转换及图像类型转换 3.掌握图像文件的读写 4.掌握图像及灰度图像直方图的显示 5.掌握图像缩放和旋转 二、实验原理和方法 1.关于图像矩阵 MATLAB中图像数据以矩阵方式的存储。所以有必要学会关于矩阵的操作，由于篇幅有限，这里只作简要的介绍。 生成矩阵的函数有： eye 生成单位矩阵 ones全1阵 zeros 全零阵 rand 均匀随机阵 randn 正态随机阵 2.图像数据类型及图像类型 2.1 图像数据类型转换 MATLAB中图像数据矩阵的存储方式为双精度(double)类型即64位浮点数。而存储图像时MATLAB有时采用无符号整型（uint8)即图像矩阵中的每个数据占用一个字节。由于大多数运算和函数(比如最基本的矩阵加减运算)都不支持uint8类型，所以运算时通常要将图像转换成 double型。 函数double将数据转换为双精度浮点类型，调用格式为： X64=double(x8) /256 2.2 图像类型及转换 在MATLAB中，一幅图像可能包含一个数据矩阵，也可能有一个颜色映像表矩阵。MATLAB图像处理工具箱支持四种图像类型，其区别在于数据矩阵元素的不同含意。它们是：● 真彩色图像 ● 索引图像 ● 灰度图像 ● 二值图像 （１）真彩色图像 真彩色图像又称RGB图像，对于一个尺寸为M×N的彩色图像来说，在MATLAB中则存储为一个M×N×3的多维数组，像素的颜色由保存在像素位置上的R、G、B的强度值的组合来确定。如果需要知道图像A中（x，y）处的像素值，则可以使用这样的代码A（x,y,1:3）。 （２）索引图像
树叶分类数字图像处理 在树叶识别中的应用 TTA standardization office【TTA 5AB- TTAK 08- TTA 2C】
数字图像处理研究报告 数字图像处理在树叶识别中的应用 侯杰：土木系 侯晓鹏：林科院 苏东川：航院 张伟：精仪 指导教师：马慧敏教授 日期：数字图像处理在树叶识别中的应用 一、课题意义及背景 1 课题背景 植物的识别与分类对于区分植物种类，探索植物间的亲缘关系，阐明植物 系统的进化规律具有重要意义。因此植物分类学是植物科学乃至整个生命科学 的基础学科。然而，由于学科发展和社会等原因，全世界范围内目前从事经典 分类（即传统的形态分类）的人数急剧下降，且呈现出明显的老龄化趋势，后 继乏人，分类学已经成为一个“濒危学科”(Buyck,1999)。这不仅对于植物分类学 本身，而且对整个植物科学和国民经济的发展带来重大的不利影响。目前植物 识别和分类主要由人工完成。然而地球上仅为人所知的有花植物就有大约25万 种，面对如此庞大的植物世界，任何一个植物学家都不可能知道所有的物种和 名称，这就给进一步的研究带来了困难。在信息化的今天，我们提出的一种解
决方案是：建立计算机化的植物识别系统，即利用计算机及相关技术对植物进行识别和管理[1]。 2 课题意义[2-3] （1）人工进行植物叶形的分类难度很大。这种传统的判别方法要求操作者具有丰富的分类学知识和长期的实践经验,才能开展工作。要做到准确和快速地识别手中的植物是非常困难。并且相应人才极为短缺。 （2）仅为人所知的有花植物就有大约25万种，面对如此庞大的植物世界，任何一个植物学家都不可能知道所有的物种和名称。建立植物识别系统和数据库十分必要。 （3）植物学研究人员在野外考察时, 时常需要获取植物叶片面积等参数。（4）叶子面积大小对植物的生长发育、作物产量以及栽培管理都具有十分重要的意义。 因此，基于计算机图像处理识别技术的树叶图像识别技术对于植物学，农业科学等都具有重大意义。 二、相关理论综述 1 图像预处理 （1）边缘检测[4] 图像的边缘是指图像局部亮度变化最显着的部分，即在灰度级上发生急剧变化的区域。从空域角度看，二维图像上的边缘相邻像素灰度从某一个值跳变