Peripheral-Foveal Vision for Real-time Object Recognition and Tracking in Video Stephen Gould,Joakim Arfvidsson,Adrian Kaehler,Benjamin Sapp,Marius Messner, Gary Bradski,Paul Baumstarck,Sukwon Chung and Andrew Y.Ng
Stanford University
Stanford,CA94305USA
Abstract
Human object recognition in a physical3-d envi-
ronment is still far superior to that of any robotic
vision system.We believe that one reason(out
of many)for this—one that has not heretofore
been signi?cantly exploited in the arti?cial vision
literature—is that humans use a fovea to?xate on,
or near an object,thus obtaining a very high resolu-
tion image of the object and rendering it easy to rec-
ognize.In this paper,we present a novel method for
identifying and tracking objects in multi-resolution
digital video of partially cluttered environments.
Our method is motivated by biological vision sys-
tems and uses a learned“attentive”interest map
on a low resolution data stream to direct a high
resolution“fovea.”Objects that are recognized in
the fovea can then be tracked using peripheral vi-
sion.Because object recognition is run only on
a small foveal image,our system achieves perfor-
mance in real-time object recognition and tracking
that is well beyond simpler systems.
1Introduction
The human visual system far outperforms any robotic system in terms of object recognition.There are many reasons for this,and in this paper,we focus only on one hypothesis—which has heretofore been little exploited in the computer vi-sion literature—that the division of the retina into the foveal and peripheral regions is of key importance to improving per-formance of computer vision systems on continuous video sequences.Brie?y,the density of photoreceptor rod and cone cells varies across the retina.The small central fovea,with a high density of color sensitive cone cells,is responsible for detailed vision,while the surrounding periphery,contain-ing a signi?cantly lower density of cones and a large num-ber of monochromatic rod cells,is responsible for,among other things,detecting motion.Estimates of the equivalent number of“pixels”in the human eye vary,but based on spatial acuity of1/120degrees[Edelman and Weiss,1995; Fahle and Poggio,1981],appears to be on the order of5×108 pixels.For a more in-depth treatment of human visual per-ception,we refer the reader to one of the many excellent text-books in the literature,for example[Wandell,1995].
Because very little changes in a visual stream from one frame to the next,one expects that it is possible to identify objects or portions of a scene one at a time using the high resolution fovea,while tracking previously identi?ed objects in the peripheral region.The result is that it is possible to use computationally expensive classi?cation algorithms on the relatively limited portion of the scene on which the fovea is?xated,while accumulating successful classi?cations over a series of frames in real-time.
A great deal has happened in recent years in the area of location and identi?cation of speci?c objects or object classes in images.Much of this work,however,has concentrated on single frames with the object of interest taking up a signi?cant proportion of the?eld-of-view.This allows for accurate and robust object recognition,but implies that if we wish to be able to?nd very small objects,we must go to much higher image resolutions.1In the naive approach,there is signi?cant computational cost of going to this higher resolution.
For continuous video sequences,the standard technique is simply to treat each frame as a still image,run a classi?ca-tion algorithm over the frame,and then move onto the next frame,typically using overlap to indicate that an object found in frame n+1is the same object found in frame n,and us-ing a Kalman?lter to stabilize these measurements over time. The primary dif?culty that this simple method presents is that a tremendous amount of effort is misdirected at the vast ma-jority of the scene which does not contain the objects we are attempting to locate.Even if the Kalman?lter is used to pre-dict the new location of an object of interest,there is no way to detect the appearance of new objects which either enter the scene,or which are approached by a camera in motion.
The solution we propose is to introduce a peripheral-foveal model in which attention is directed to a small portion of the visual?eld.We propose using a low-resolution,wide-angle video camera for peripheral vision and a pan-tilt-zoom(PTZ) camera for foveal vision.The PTZ allows very high resolu-tion on the small portion of the scene in which we are inter-ested at any given time.We refer to the image of the scene supplied by the low-resolution,wide-angle camera as the pe-ripheral view and the image supplied by the pan-tilt-zoom 1Consider,for example,a typical coffee mug at a distance of?ve meters appearing in a standard320×240resolution,42??eld-of-view camera.The95mm tall coffee mug is reduced to a mere8 pixels high,rendering it extremely dif?cult to recognize.
camera as the foveal view or simply fovea.
We use an attentive model to determine regions from the peripheral view on which we wish to perform object classi-?cation.The attentive model is learned from labeled data, and can be interpreted as a map indicating the probability that any particular pixel is part of an unidenti?ed object. The fovea is then repeatedly directed by choosing a region to examine based on the expected reduction in our uncer-tainty about the location of objects in the scene.Identi?ca-tion of objects in the foveal view can be performed using any of the state-of-the-art object classi?cation technologies (for example see[Viola and Jones,2004;Serre et al.,2004; Brubaker et al.,2005]).
The PTZ camera used in real-world robotic applications can be modeled for study(with less peak magni?cation)using a high-de?nition video(HDV)camera,with the“foveal”re-gion selected from the video stream synthetically by extract-ing a small region of the HDV image.The advantage of this second method is that the input data is exactly reproducible. Thus we are able to evaluate different foveal camera motions on the same recorded video stream.Figure1shows our robot mounted with our two different camera
setups.
Figure1:STAIR platform(left)includes a low-resolution periph-eral camera and high-resolution PTZ camera(top-right),and alter-native setup with HDV camera replacing the PTZ(bottom-right).In our experiments we mount the cameras on a standard tripod instead of using the robotic platform.
The robot is part of the STAIR(STanford Arti?cial Intelli-gence Robot)project,which has the long-term goal of build-ing an intelligent home/of?ce assistant that can carry out tasks such as cleaning a room after a dinner party,and?nding and fetching items from around the home or of?ce.The ability to visually scan its environment and quickly identify objects is of one of the key elements needed for a robot to accomplish such tasks.
The rest of this paper is organized as follows.Section2outlines related work.In Section3we present the probabilis-tic framework for our attention model which directs the foveal gaze.Experimental results from a sample HDV video stream as well as real dual-camera hardware are presented in Sec-tion4.The video streams are of a typical of?ce environment and contain distractors,as well as objects of various types for which we had previously trained classi?ers.We show that the performance of our attention driven system is signi?cantly better than naive scanning approaches.
2Related Work
An early computational architecture for explaining the visual attention system was proposed by Koch and Ullman(1985). In their model,a scene is analyzed to produce multiple fea-ture maps which are combined to form a saliency map.The single saliency map is then used to bias attention to regions of highest activity.Many other researchers,for example[Hu et al.,2004;Privitera and Stark,2000;Itti et al.,1998],have suggested adjustments and improvements to Koch and Ull-man’s model.A review of the various techniques is presented in[Itti and Koch,2001].
A number of systems,inspired by both physiological mod-els and available technologies,have been proposed for?nd-ing objects in a scene based on the idea of visual attention,for example[Tagare et al.,2001]propose a maximum-likelihood decision strategy for?nding a known object in an image. More recently,active vision systems have been proposed for robotic platforms.Instead of viewing a static image of the world,these systems actively change the?eld-of-view to obtain more accurate results.A humanoid system that de-tects and then follows a single known object using peripheral and foveal vision is presented in[Ude et al.,2003].A sim-ilar hardware system to theirs is proposed in[Bj¨o rkman and Kragic,2004]for the task of object recognition and pose esti-mation.And in[Orabona et al.,2005]an attention driven sys-tem is described that directs visual exploration for the most salient object in a static scene.
An analysis of the relationship between corresponding points in the peripheral and foveal views is presented in[Ude et al.,2006],which also describes how to physically control foveal gaze for rigidly connected binocular cameras.
3Visual Attention Model
Our visual system comprises two separate video cameras. A?xed wide-angle camera provides a continuous low-resolution video stream that constitutes the robot’s peripheral view of a scene,and a controllable pan-tilt-zoom(PTZ)cam-era provides the robot’s foveal view.The PTZ camera can be commanded to focus on any region within the peripheral view to obtain a detailed high-resolution image of that area of the scene.As outlined above,we conduct some of our ex-periments on recorded high-resolution video streams to allow for repeatability in comparing different algorithms.
Figure2shows an example of a peripheral and foveal view from a typical of?ce scene.The attention system selects a foveal window from the peripheral view.The corresponding region from the high-resolution video stream is then used for
Figure 2:Illustration of the peripheral (middle)and foveal (right)views of a scene in our system with attentive map showing regions of high interest (left).In our system it takes approximately 0.5seconds for the PTZ camera to move to a new location and acquire an image.
classi?cation.The attentive interest map,generated from fea-tures extracted from the peripheral view,is used to determine where to direct the foveal gaze,as we will discuss below.The primary goal of our system is to identify and track ob-jects over time.In particular,we would like to minimize our uncertainty in the location of all identi?able objects in the scene in a computationally ef?cient way.Therefore,the at-tention system should select regions of the scene which are most informative for understanding the robot’s visual envi-ronment.
Our system is able to track previously identi?ed objects over consecutive frames using peripheral vision,but can only classify new objects when they appear in the (high-resolution)fovea,F .Our uncertainty in a tracked object’s position grows with time.Directing the fovea over the expected position of the object and re-classifying allows us to update our estimate of its location.Alternatively,directing the fovea to a different part of the scene allows us to ?nd new objects.Thus,since the fovea cannot move instantaneously,the attention system needs to periodically decide between the following actions:A1.Con?rmation of a tracked object by ?xating the fovea
over the predicted location of the object to con?rm its presence and update our estimate of its position;A2.Search for unidenti?ed objects by moving the fovea to
some new part of the scene and running the object clas-si?ers over that region.Once the decision is made,it takes approximately 0.5seconds—limited by the speed of the PTZ camera—for the fovea to move to and acquire an image of the new region.2During this time we track already identi?ed objects using pe-ripheral vision.After the fovea is in position we search for new and tracked objects by scanning over all scales and shifts within the foveal view as is standard practice for many state-of-the-art object classi?ers.
More formally,let ξk (t )denote the state of the k -th object,o k ,at time t .We assume independence of all objects in the scene,so our uncertainty is simply the sum of entropy terms
2
In our experiments with single high-resolution video stream,we simulate the time required to move the fovea by delaying the image selected from the HDV camera by the required number of frames.
over all objects,
H =
o k ∈T
H tracked (ξk (t ))+ o k /
∈T H unidenti?ed (ξk (t ))(1)
where the ?rst summation is over objects being tracked,T ,
and the second summation is over objects in the scene that have not yet been identi?ed (and therefore are not being tracked).
Since our system cannot know the total number of objects in the scene,we cannot directly compute the entropy over unidenti?ed objects, o k /∈T H (ξk (t )).Instead,we learn the probability P (o k |F )of ?nding a previously-unknown ob-ject in a given foveal region F of the scene based on features extracted from the peripheral view (see the description of our interest model in section 3.2below).If we detect a new ob-ject,then one of the terms in the rightmost sum of Eq.(1)is reduced;the expected reduction in our entropy upon taking action A2(and ?xating on a region F )is
ΔH A 2=P (o k |F )
H unidenti?ed (ξk (t ))
?H tracked (ξk (t +1))
.(2)Here the term H unidenti?ed (ξk (t ))is a constant that re?ects our uncertainty in the state of untracked objects,3and H tracked (ξk (t +1))is the entropy associated with the Kalman ?lter that is attached to the object once it has been detected (see Section 3.1).
As objects are tracked in peripheral vision,our uncertainty in their location grows.We can reduce this uncertainty by observing the object’s position through re-classi?cation of the area around its expected location.We use a Kalman ?lter to track the object,so we can easily compute the reduction in entropy from taking action A1for any object o k ∈T ,
ΔH A 1=
12
(ln |Σk (t )|?ln |Σk (t +1)|)(3)
where Σk (t )and Σk (t +1)are the covariance matrices as-sociated with the Kalman ?lter for the k -th object before and after re-classifying the object,respectively.
3
Our experiments estimated this term assuming a large-variance distribution over the peripheral view of possible object locations;in practice,the algorithm’s performance appeared very insensitive to this parameter.
In this formalism,we see that by taking action A1we re-duce our uncertainty in a tracked object’s position,whereas by taking action A2we may reduce our uncertainty over unidenti?ed objects.We assume equal costs for each action and therefore we choose the action which maximizes the ex-pected reduction of the entropy H de?ned in Eq.(1).
We now describe our tracking and interest models in detail.
3.1Object Tracking
We track identi?ed objects using a Kalman ?lter.Because we assume independence of objects,we associate a separate Kalman ?lter with each object.An accurate dynamic model of objects is outside the current scope of our system,and we use simplifying assumptions to track each object’s coordi-nates in the 2-d image plane.The state of the k -th tracked object is
ξk =[x k
y k
˙x k
˙y k ]
T
(4)
and represents the (x,y )-coordinates and (x,y )-velocity of the object.
On each frame we perform a Kalman ?lter motion update step using the current estimate of the object’s state (position and velocity).The update step is
ξk (t +1)=???10ΔT 0
01
0ΔT 00100001???ξk (t )+ηm
(5)where ηm ～N (0,Σm )is the motion model noise,and ΔT
is the duration of one timestep.
We compute optical ?ow vectors in the peripheral view generated by the Lucas and Kanade (1981)sparse optical ?ow algorithm.From these ?ow vectors we measure the ve-locity of each tracked object,z v (t ),by averaging the optical ?ow within the object’s bounding box.We then perform a Kalman ?lter observation update,assuming a velocity obser-vation measurement model
z v (t )= 0010
0001 ξk (t )+ηv
(6)where ηv ～N (0,Σv )is the velocity measurement model noise.
When the object appears in the foveal view,i.e.,after tak-ing action A1,the classi?cation system returns an estimate,z p (t ),of the object’s position,which we use to perform a cor-responding Kalman ?lter observation update to greatly reduce our uncertainty in its location accumulated during tracking.Our position measurement observation model is given by
z p (t )=
1000
0100 ξk (t )+ηp (7)where ηp ～N (0,Σp )is the position measurement model noise.
We incorporate into the position measurement model the special case of not being able to re-classify the object even though the object is expected to appear in the foveal view.For example,this may be due to misclassi?cation error,poor
estimation of the object’s state,or occlusion by another ob-ject.In this case we assume that we have lost track of the
object.
Since objects are recognized in the foveal view,but are tracked in the peripheral view,we need to transform coor-dinates between the two https://www.doczj.com/doc/7116004458.html,ing the well-known stereo calibration technique of [Zhang,2000],we compute the ex-trinsic parameters of the PTZ camera with respect to the wide-angle camera.Given that our objects are far away rel-ative to the baseline between the cameras,we found that the disparity between corresponding pixels of the objects in each view is small and can be corrected by local correlation.4This approach provides adequate accuracy for controlling the fovea and ?nding the corresponding location of objects be-tween views.
Finally,the entropy of a tracked object’s state is required for deciding between actions.Since the state,ξ(t ),is a Gaus-sian random vector,it has differential entropy
H tracked (ξk (t ))=
12
ln(2πe )4|Σk (t )|(8)
where Σk (t )∈R 4×4is the covariance matrix associated with our estimate of the k -th object at time t .
3.2Interest Model
To choose which foveal region to examine next (under action A2),we need a way to estimate the probability of detecting a previously unknown object in any region F in the scene.To do so,we de?ne an “interest model”that rapidly identi-?es pixels which have a high probability of containing objects that we can classify.A useful consequence of this de?nition is that our model automatically encodes the biological phe-nomena of saliency and inhibition of return —the processes by which regions of high visual stimuli are selected for fur-ther investigation,and by which recently attended regions are prevented from being attended again (see [Klein,2000]for a detailed review).
In our approach,for each pixel in our peripheral view,we estimate the probability of it belonging to an unidenti?ed ob-ject.We then build up a map of regions in the scene where the density of interest is high.From this interest map our atten-tion system determines where to direct the fovea to achieve maximum bene?t as described above.
More formally,we de?ne a pixel to be interesting if it is part of an unknown,yet classi?able object .Here classi?able ,means that the object belongs to one of the object classes that our classi?cation system has been trained to recognize.We model interestingness,y (t ),of every pixel in the periph-eral view,x (t ),at time t using a dynamic Bayesian network (DBN),a fragment of which is shown in Figure 3.For the (i,j )-th pixel,y i,j (t )=1if that pixel is interesting at time t ,and 0otherwise.Each pixel also has associated with it a vector of observed features φi,j (x (t ))∈R n .
4
We resize the image from the PTZ camera to the size it would appear in the peripheral view.We then compute the cross-correlation of the resized PTZ image with the peripheral image,and take the actual location of the fovea to be the location with maximum cross-correlation in a small area around its expected location in the periph-eral view.
y i,j (t -1)y i-1,j (t -1)y i +1,j (t -1)i,j (x (t ))
y i,j -1(t y i-1,j -1(t y i +1,j -1(t y i,j +1(t -1)y i +1,j +1(t -1)
y i-1,j +1(t -1)Figure 3:Fragment of dynamic Bayesian network for modeling at-tentive interest.
The interest belief state over the entire frame at time t is P (y (t )|y (0),x (0),...,x (t )).Exact inference in this graphical model is intractable,so we apply approximate infer-ence to estimate this probability.Space constraints preclude a full discussion,but brie?y,we applied Assumed Density Fil-tering/the Boyen-Koller (1998)algorithm,and approximate this distribution using a factored representation over individ-ual pixels,P (y (t ))≈ i,j P (y i,j (t )).In our experiments,this inference algorithm appeared to perform well.
In our model,the interest for a pixel depends both on the interest in the pixel’s (motion compensated)neighbor-hood,N (i,j ),at the previous time-step and features ex-tracted from the current frame.The parameterizations for both P y i,j (t )|y N (i,j )(t ?1) and P (y i,j (t )|φi,j (x (t )))are given by logistic functions (with learned parameters),and we use Bayes rule to compute P (φi,j (x (t ))|y i,j (t ))needed in the DBN belief https://www.doczj.com/doc/7116004458.html,ing this model of the interest map,we estimate the probability of ?nding an object in a given foveal region,P (o k |F ),by computing the mean of the probability of every pixel in the foveal region being inter-esting.5
The features φi,j used to predict interest in a pixel are ex-tracted over a local image patch and include Harris corner features,horizontal and vertical edge density,saturation and color values,duration that the pixel has been part of a tracked object,and weighted decay of the number of times the fovea had previously ?xated on a region containing the pixel.
3.3Learning the Interest Model
Recall that we de?ne a pixel to be interesting if it is part of an as yet unclassi?ed object.In order to generate training data so that we can learn the parameters of our interest model,we ?rst hand label all classi?able objects in a low-resolution training video sequence.Now as our system begins to recognize and track objects,the pixels associated with those objects are no longer interesting,by our de?nition.Thus,we generate our training data by annotating as interesting only those pixels marked interesting in our hand labeled training video but that are not part of objects currently being https://www.doczj.com/doc/7116004458.html,ing this
5
Although one can envisage signi?cantly better estimates,for ex-ample using logistic regression to recalibrate these probabilities,the algorithm’s performance appeared fairly insensitive
to this choice.
procedure we can hand label a single training video and auto-matically generate data for learning the interest model given any speci?c combination of classi?ers and foveal movement policies.
We adapt the parameters of our probabilistic models so as to maximize the likelihood of the training data described above.Our interest model is trained over a 450frame video sequence,i.e.,30seconds at 15frames per second.Figure 4shows an example of a learned interest map,in which the algorithm automatically selected the mugs as the most inter-esting region.
Figure 4:Learned interest.The righthand panel shows the proba-bility of interest for each pixel in the peripheral image (left).
4Experimental Results
We evaluate the performance of the visual attention system by measuring the percentage of times that a classi?able object appearing in the scene is correctly identi?ed.We also count the number of false-positives being tracked per frame.We compare our attention driven method for directing the fovea to three naive approaches:
(i)?xing the foveal gaze to the center of view,
(ii)linearly scanning over the scene from top-left to bottom-right,and,(iii)randomly moving the fovea around the scene.
Our classi?ers are trained on image patches of roughly 100×100pixels depending on the object class being trained.For each object class we use between 200and 500positive training examples and 10,000negative examples.Our set of negative examples contained examples of other objects,as well as random images.We extract a subset of C1fea-tures (see [Serre et al.,2004])from the images and learn a boosted decision tree classi?er for each object.This seemed to give good performance (comparable to state-of-the-art sys-tems)for recognizing a variety of objects.The image patch size was chosen for each object so as to achieve good classi-?cation accuracy while not being so large so as to prevent us from classifying small objects in the scene.
We conduct two experiments using recorded high-de?nition video streams,the ?rst assuming perfect classi?-cation (i.e.,no false-positives and no false-negatives,so that every time the fovea ?xates on an object we correctly classify the object),and the second using actual trained state-of-the-art classi?ers.In addition to the different fovea control algo-rithms,we also compare our method against performing ob-ject classi?cation on full frames for both low-resolution and
high-resolution video streams.The throughput (in terms of frames per second)of the system is inversely proportional to the area scanned by the object classi?ers.The low resolution was chosen to exhibit the equivalent computational require-ments of the foveal methods and is therefore directly com-parable.The high resolution,however,being of much larger size than the foveal region,requires additional computation time to complete each frame classi?cation.Thus,to meet real-time operating requirements,the classi?ers were run ten times less often.
Figure 5shows example timelines for recognizing and tracking a number of different objects using our method (bot-tom),randomly moving the fovea (middle),and using a ?xed fovea (top).Each object is shown on a different line.A dot-ted line indicates that the object has appeared in the scene but has not yet been recognized (or that our tracking has lost it).A solid line indicates that the object is being tracked by our system.The times when the fovea ?xates on each object are marked by a diamond ( ).It is clear that our method recog-nizes and starts tracking objects more quickly than the other methods,and does not waste effort re-classifying objects in regions already explored (most apparent when the fovea is
?xed).
jects in a continuous video sequence using different methods for controlling the fovea:?xed (top),random (middle),and our inter-est driven method (bottom).
In our experiments we use a ?xed size fovea covering ap-proximately 10%of the peripheral ?eld-of-view.6All recog-nizable objects in our test videos were smaller than the size of the foveal window.We record peripheral and foveal video streams of an of?ce environment with classi?ers trained on commonly found objects:coffee mugs,disposable cups,sta-plers,scissors,etc.We hand label every frame of the video stream.The video sequence consists of a total of 700frames recorded at 15frames per second—resulting in 100foveal
6
Although direct comparison is not possible,for interest we note that the human fovea covers roughly 0.05%of the visual ?eld.
?xations (since it takes approximately seven frames for the fovea to move).
A comparison between our method and the other ap-proaches is shown in Table 1.We also include the F 1-score when running state-of-the-art object classi?ers in Table 2.Fovea control
Perfect Actual classi?cation
classi?ers Full image (low-resolution)n/a 0.0%a Full image (high-resolution)n/a 62.2%Fixed at center 16.1%15.9%Linear scanning 61.9%43.1%Random scanning 62.0%60.3%Our (attentive)method
85.1%82.9%
a
Our object classi?ers are trained on large images samples (of ap-proximately 100×100pixels).This result illustrates that objects in the low-resolution view occupy too few pixels for reliable classi?cation.Table 1:Percentage of objects in a frame that are correctly identi?ed using different fovea control algorithms.
Fovea control
Recall Precision F 1-Score Full image (low-res)0.0%0.0%0.0%Full image (high-res)62.2%95.0%75.2%Fixed at center 15.9%100.0%27.4%Linear scanning 43.1%97.2%59.7%Random scanning 60.3%98.9%74.9%Our method 82.9%99.0%90.3%
2:Recall,precision and F 1-score for objects appearing the The low-resolution run results in an F 1-score of zero since appearing in the scene occupy too few pixels for reliable The results show that our attention driven method performs better than the other foveal control strategies.method even performs better than scanning the entire image.This is because our method runs much since it only needs to classify objects in a small region.running over the entire high-resolution image,the sys-is forced to skip frames so that it can keep up with real-time.It is therefore less able to detect objects entering and leaving the scene than our method.Furthermore,because we only direct attention to interesting regions of the scene,our method results in signi?cantly higher precision,and hence better F 1-score,than scanning the whole image.
Finally,we conduct experiments with a dual wide-angle and PTZ camera system.In order to compare results be-tween the different foveal control strategies,we ?x the robot and scene and run system for 100frames (approximately 15foveal ?xations)for each strategy.We then change the scene by moving both the objects and the robot pose,and repeat the experiment,averaging our results across the different scenes.The results are shown in Table 3.Again our attention driven method performs better than the other approaches.Videos demonstrating our results are provided at
https://www.doczj.com/doc/7116004458.html,/～sgould/vision/
Fovea control Recall Precision F1-Score
Fixed at center a9.49%97.4%17.3%
Linear scanning13.6%100.0%24.0%
Random scanning27.7%84.1%41.6%
Our method62.2%83.9%71.5%
In some of the trials a single object happened to appear in the center of the?eld of view and hence was detected by the stationary foveal gaze.
Table3:Recall,precision and F1-score of recognizable objects for hardware executing different foveal control strategies over a number of stationary scenes.
5Discussion
In this paper,we have presented a novel method for im-proving performance of visual perception on robotic plat-forms and in digital video.Our method,motivated from biological vision systems,minimizes the uncertainty in the location of objects in the environment using a with con-trollable pan-tilt-zoom camera and?xed wide-angle camera. The pan-tilt-zoom camera captures high-resolution images for improved classi?cation.Our attention system controls the gaze of the pan-tilt-zoom camera by extracting interesting re-gions(learned as locations where we have a high probabil-ity of?nding recognizable objects)from a?xed-gaze low-resolution peripheral view of the same scene.
Our method also works on a single high-resolution digital video stream,and is signi?cantly faster than naively scanning the entire high-resolution image.By experimentally compar-ing our method to other approaches for both high-de?nition video and on real hardware,we showed that our method?x-ates on and identi?es objects in the scene faster than all of the other methods tried.In the case of digital video,we even perform better than the brute-force approach of scanning the entire frame,because the increased computational cost of do-ing so results in missed frames when run in real-time. Acknowledgments
We give warm thanks to Morgan Quigley for help with the STAIR hardware,and to Pieter Abbeel for helpful discus-sions.This work was supported by DARPA under contract number FA8750-05-2-0249.
References
[Bj¨o rkman and Kragic,2004]Marten Bj¨o rkman and Danica https://www.doczj.com/doc/7116004458.html,bination of foveal amd peripheral vision for object recognition and pose estimation.In Proceedings of the2004 IEEE International Conference on Robotics and Automation, pages5135–5140,2004.
[Boyen and Koller,1998]Xavier Boyen and Daphne Koller.
Tractable inference for complex stochastic processes.In Pro-ceedings of the14th Conference on Uncertainty in Arti?cial Intelligence—UAI1998,pages33–42.San Francisco:Morgan Kaufmann,1998.
[Brubaker et al.,2005]S.Charles Brubaker,Jianxin Wu,Jie Sun, Matthew D.Mullin,and James M.Rehg.On the design of cas-cades of boosted ensembles for face detection.In Tech report GIT-GVU-05-28,2005.[Edelman and Weiss,1995]Shimon Edelman and Yair Weiss.Vi-sion,hyperacuity,1995.
[Fahle and Poggio,1981]M.Fahle and T.Poggio.Visual hyperacu-ity:Spatiotemporal interpolation in human vision.In Proceed-ings of the Royal Society of London.The Royal Society,1981. [Hu et al.,2004]Yiqun Hu,Xing Xie,Wei-Ying Ma,Liang-Tien Chia,and Deepu Rajan.Salient region detection using weighted feature maps based on the human visual attention model.In Fifth IEEE Pacifc-Rim Conference on Multimedia,Tokyo Waterfront City,Japan,November2004.
[Itti and Koch,2001]L.Itti and https://www.doczj.com/doc/7116004458.html,putational modeling of visual attention.Nature Reviews Neuroscience,2(3):194–203, Mar2001.
[Itti et al.,1998]Laurent Itti,Christof Koch,and Ernst Niebur.A model of saliency-based visual attention for rapid scene analy-sis.IEEE Transactions on Pattern Analysis and Machine Intelli-gence,20(11):1254–1259,1998.
[Klein,2000]Raymond M.Klein.Inhibition of return.Trends in Cognitive Sciences,4(4),April2000.
[Koch and Ullman,1985] C.Koch and S.Ullman.Shifts in selec-tive visual attention:Towards the underlying neural circuitry.Hu-man Neurobiology,4:219–227,1985.
[Lucas and Kanade,1981]Bruce D.Lucas and Takeo Kanade.An iterative image registration technique with an application to stereo vision(ijcai).In Proceedings of the7th International Joint Conference on Arti?cial Intelligence(IJCAI’81),pages674–679, April1981.
[Orabona et al.,2005]Francesco Orabona,Giorgio Metta,and Giulio Sandini.Object-based visual attention:a model for a be-having robot.IEEE Conference on Computer Vision and Pattern Recognition,3:89,2005.
[Privitera and Stark,2000]Claudio M.Privitera and Lawrence W.
Stark.Algorithms for de?ning visual regions-of-interest:Com-parison with eye?xations.IEEE Transactions on Pattern Analy-sis and Machine Intelligence,22(9):970–982,2000.
[Serre et al.,2004]Thomas Serre,Lior Wolf,and Tomaso Poggio.
A new biologically motivated framework for robust object recog-
nition.In Al Memo2004-026,November2004.
[Tagare et al.,2001]Hermant D.Tagare,Kentaro Toyama,and Jonathan G.Wang.A maximum-likelihood strategy for direct-ing attention during visual search.IEEE Transactions on Pattern Analysis and Machine Intelligence,23(5):490–500,2001. [Ude et al.,2003] A.Ude,C.G.Atkeson,and https://www.doczj.com/doc/7116004458.html,bin-ing peripheral and foveal humanoid vision to detect,pursue,rec-ognize and act.In Proc.IEEE/RSJ Int.Conf.Intelligent Robots and Systems,pages2173–2178,October2003.
[Ude et al.,2006] A.Ude,C.Gaskett,and G.Cheng.Foveated vi-sion systems with two cameras per eye.In Proceedings2006 IEEE International Conference on Robotics and Automation, 2006.ICRA2006.,pages3457–3462,May2006.
[Viola and Jones,2004]Paul Viola and Michael J.Jones.Robust real-time face detection.International Journal of Computer Vi-sion,57(2):137–154,2004.
[Wandell,1995]Brian A.Wandell.Foundations of Vision.Sinauer Associates,Inc.,1995.
[Zhang,2000]Zhengyou Zhang.A?exible new technique for cam-era calibration.IEEE Transactions on Pattern Analysis and Ma-chine Intelligence,22(11):1330–1334,2000.
毕业设计（论文）开题报告 题目电动汽车充电对配电系统的影响学院自动化学院 专业电气工程与自动化专业 姓名 班级 学号 指导教师
一、综述本课题选题的依据和意义 2015年9月，关于充电桩的好消息频频传出，如国务院常务会议上提出加快电动汽车充电基础设施、电动汽车充电接口国家标准修订稿通过审查，以及国务院办公厅日前发布《关于加快电动汽车充电基础设施建设的指导意见》，到2020年，我国将基本建成适度超前、车桩相随、智能高效的充电基础设施体系，满足超过500万辆电动汽车的充电需求。尤为引人关注的是，北京通信展举行期间，国务院副总理马凯称：国家准备将充电桩普及的任务交由铁塔公司负责。 随着时代的发展，汽车的普及，石油的消耗也日益增加，作为不可再生的石油资源，总有枯竭的一天，而且汽车尾气的排放对环境有很大的污染，所以利用电能替代石油，实现汽车尾气的零排放，将是未来汽车行业的发展趋势。截至2015年一月，从公开信息显示，目前杭州已经建成了70座充换电站、620个充电桩，其中杭州主城区投入运营的充换电站有27个。在杭州市区里，一辆新能源车要找到最近的充电站，只要开4.5公里。 国内外电动汽车在近几年的发展情况如下：经过2012和2013年的缓慢起步，全球电动汽车销量终于在2014年下半年爆发，全年销售已超过30万辆大关，中国新能源汽车品牌突飞猛进，比亚迪从2013年的第40名跃升至第七位，康迪电动车也挤进第十名。在2015年，全球电动汽车销售量达到40万辆左右。 随着电动汽车规模化应用,电网原有装机和线路容量是否能应对新增充电负荷需求,即在不扩大规模的情况下,如何提高原有电网利用率,增加容纳能力,同时最大限度降低充电负荷对电网的负面影响,在分析充电负荷特性基础上,针对电动汽车充电对电网各方面的影响,提出相应对策,将对电动汽车产业化进程具有重要的研究意义。 二、国内外研究动态 我国电动汽车起步较发达国家晚,但研究发展快。各汽车生产商进入研发电动汽车的行列,已经在促进电动汽车各方面技术发展方面取得有益的成见。由于配套设施等主要原因,家用电动汽车并没有广泛的推广开来，现有的电动汽车充电站主要对电动公交车和工程车辆进行充电。同时针对电动汽车充电站运行和电动汽车充电对配电网的研究仍然非常缺乏。电动汽车规模预测，用户充电行为的随机性,以及接入电网造成的影响和相应充电控制策略都没有一套成熟的模式和系统，以及对此的全面研究。 电动汽车对电网的影响，主要包括电网负荷特性影响、谐波影响、电压电流
对翻译中异化法与归化法的正确认识 班级：外语学院、075班 学号：074050143 姓名：张学美 摘要：运用异化与归化翻译方法,不仅是为了让读者了解作品的内容,也能让读者通过阅读译作,了解另一种全新的文化,因为进行文化交流才是翻译的根本任务。从文化的角度考虑,采用异化法与归化法,不仅能使译文更加完美,更能使不懂外语的人们通过阅读译文,了解另一种文化,促进各民族人们之间的交流与理解。翻译不仅是语言符号的转换,更是跨文化的交流。有时,从语言的角度所作出的译文可能远不及从文化的角度所作出的译文完美。本文从翻译策略的角度，分别从不同时期来说明人们对异化法与归化法的认识和运用。 关键词:文学翻译；翻译策略；异化；归化；辩证统一 一直以来,无论是在我国还是在西方,直译(literal translation)与意译(liberal translation)是两种在实践中运用最多,也是被讨论研究最多的方法。1995年,美籍意大利学者劳伦斯-韦努蒂(Lawrence Venuti)提出了归化(domestication)与异化(foreignization)之说,将有关直译与意译的争辩转向了对于归化与异化的思考。归化与异化之争是直译与意译之争的延伸,是两对不能等同的概念。直译和意译主要集中于语言层面,而异化和归化则突破语言的范畴,将视野扩展到语言、文化、思维、美学等更多更广阔的领域。 一、归化翻译法 Lawrwnce Venuti对归化的定义是,遵守译入语语言文化和当前的主流价值观,对原文采用保守的同化手段,使其迎合本土的典律,出版潮流和政治潮流。采用归化方法就是尽可能不去打扰读者,而让作者向读者靠拢(the translator leaves the reader in peace, as much as possible, and moves the author towards him)。归化翻译法的目的在于向读者传递原作的基本精神和语义内容,不在于语言形式或个别细节的一一再现。它的优点在于其流利通顺的语言易为读者所接受,译文不会对读者造成理解上的障碍,其缺点则是译作往往仅停留在内容、情节或主要精神意旨方面,而无法进入沉淀在语言内核的文化本质深处。 有时归化翻译法的采用也是出于一种不得已,翻译活动不是在真空中进行的,它受源语文化和译语文化两种不同文化语境的制约,还要考虑到两种文化之间的
机器视觉技术发展现状 人类认识外界信息的80%来自于视觉，而机器视觉就是用机器代替人眼来做 测量和判断，机器视觉的最终目标就是使计算机像人一样，通过视觉观察和理解 世界，具有自主适应环境的能力。作为一个新兴学科，同时也是一个交叉学科，取“信息”的人工智能系统，其特点是可提高生产的柔性和自动化程度。目前机器视觉技术已经在很多工业制造领域得到了应用，并逐渐进入我们的日常生活。 机器视觉是通过对相关的理论和技术进行研究，从而建立由图像或多维数据中获机器视觉简介 机器视觉就是用机器代替人眼来做测量和判断。机器视觉主要利用计算机来模拟人的视觉功能，再现于人类视觉有关的某些智能行为，从客观事物的图像中提取信息进行处理，并加以理解，最终用于实际检测和控制。机器视觉是一项综合技术，其包括数字处理、机械工程技术、控制、光源照明技术、光学成像、传感器技术、模拟与数字视频技术、计算机软硬件技术和人机接口技术等，这些技术相互协调才能构成一个完整的工业机器视觉系统[1]。 机器视觉强调实用性，要能适应工业现场恶劣的环境，并要有合理的性价比、通用的通讯接口、较高的容错能力和安全性、较强的通用性和可移植性。其更强调的是实时性，要求高速度和高精度，且具有非接触性、实时性、自动化和智能 高等优点，有着广泛的应用前景[1]。 一个典型的工业机器人视觉应用系统包括光源、光学成像系统、图像捕捉系统、图像采集与数字化模块、智能图像处理与决策模块以及控制执行模块。通过 CCD或CMOS摄像机将被测目标转换为图像信号，然后通过A/D转换成数字信号传送给专用的图像处理系统，并根据像素分布、亮度和颜色等信息，将其转换成数字化信息。图像系统对这些信号进行各种运算来抽取目标的特征，如面积、 数量、位置和长度等，进而根据判别的结果来控制现场的设备动作[1]。 机器视觉一般都包括下面四个过程：
1) With the rapid improvement in.../growing awareness of..., more and more.../sth.... 1 随着…的飞速发展/越来越多的关注，越来越… (e.g. With the considerable improvement in building industry, more and more structures are being erected to set the people's minds at ease.) 例：随着建筑业的大力推进，人们对建立越来越多的房屋建筑感到宽心。 2) Recently, sth./the problem of...has been brought to popular attention/ has become the focus of public concern. 2 近来，某事/某问题引起了人们的普遍关注/成了公众关注的焦点。 (e.g. Recently, the problem of unemployment has been brought to such popular attention that governments at all levels place it on the agenda as the first matter.) 例：近来失业问题引起了人们的普遍关注，各级政府已把它列为首要议程。 3) One of the universal issues we are faced with/that cause increasing concern is that... 3 我们面临的一个普遍问题是… /一个越来越引人关注的普遍问题 是… (e.g. One of the universal issues that draw (cause) growing concern is whether it is wise of man to have invented the automobile.) 例：一个越来越引人关注的普遍问题是，发明汽车是否为人 4) In the past few years, there has been a boom/sharp growth/decline in.. . 4 在过去的几年里，…经历了突飞猛进/迅猛增长/下降。 (e.g. In the past ten years, there has been a sharp decline in the number of species.) 例：在过去的几年里，物种数量骤然下降。 5) Nowadays, more/most important/dangerous for our society is... 5 如今对我们社会更（最）为重要的（危险的）事情是… (e.g. Nowadays, most dangerous for our society is the tendency to take advantage of each other in political circles.) 例：对我们社会最为危险的事情是政界倾向于互相利用。 6) According to the information given in the table/graph, we can find that... 6 根据图表资料，我们可以发现… 7) As can be seen from the table/graph/figure, there is a marked increase /decline/favorable (an unfavorable) change in... 7 根据图表（数字）显示，…明显增长（下降）/发生了有利（不利）变化。 8) As we can see from the table/graph/figure above, drastic/considerable/ great changes have taken place in...over the period of time from...(年份)to...( 年份) 8 据上面图表（数字）所示，从某年到某年某方面发生了剧烈的（相当大的；巨大的）变化。
基于机器视觉的工件识别和定位文献综述 1.前言 1.1工业机器人的现状与发展趋势 机器人作为一种最典型的应用范围广、技术附加值高的数字控制装备，在现代先进生产制造业中发挥的作用越来越重要，机器人技术的发展将会对未来生产和社会发展起到强有力的推动作用。《2l 世纪日本创建机器人社会技术发展战略报告》指出，“机器人技术与信息技术一样，在强化产业竞争力方面是极为重要的战略高技术领域。培育未来机器人产业是支撑2l 世纪日本产业竞争力的产业战略之一，具有非常重要的意义。” 研发工业机器人的初衷是为了使工人能够从单调重复作业、危险恶劣环境作业中解脱出来，但近些年来，工厂和企业引进工业机器人的主要目的则更多地是为了提高生产效率和保证产品质量。因为机器人的使用寿命很长，大都在10 年以上，并且可以全天后不间断的保持连续、高效地工作状态，因此被广泛应用于各行各业，主要进行焊接、装配、搬运、加工、喷涂、码垛等复杂作业。伴随着工业机器人研究技术的成熟和现代制造业对自动生产的需要，工业机器人越来越被广泛的应用到现代化的生产中。 现在机器人的价格相比过去已经下降很多，并且以后还会继续下降，但目前全世界范围的劳动力成本都有所上涨，个别国家和地区劳动力成本又很高，这就给工业机器人的需求提供了广阔的市场空间，工业机器人销量的保持着较快速度的增长。工业机器人在生产中主要有机器人工作单元和机器人工作生产线这两种应用方式，并且在国外，机器人工作生产线已经成为工业机器人主要的应用方式。以机器人为核心的自动化生产线适应了现代制造业多品种、少批量的柔性生产发展方向，具有广阔的市场发展前景和强劲生命力，已开发出多种面向汽车、电气机械等行业的自动化成套装备和生产线产品。在发达国家，机器人自动化生产线已经应用到了各行各业，并且已经形成一个庞大的产业链。像日本的FANUC、MOTOMAN，瑞典的ABB、德国的KUKA、意大利的COMAU 等都是国际上知名的被广泛用于自动化生产线的工业机器人。这些产品代表着当今世界工业机器人的最高水平。 我国的工业机器人前期发展比较缓慢。当将被研发列入国家有关计划后，发展速度就明显加快。特别是在每次国家的五年规划和“863”计划的重点支持下，我国机器人技术的研究取得了重大发展。在机器人基础技术和关键技术方面都取得了巨大进展，科技成果已经在实际工作中得到转化。以沈阳新松机器人为代表的国内机器人自主品牌已迅速崛起并逐步缩小与国际品牌的技术差距。 机器人涉及到多学科的交叉融合，涉及到机械、电子、计算机、通讯、控制等多个方面。在现代制造业中，伴随着工业机器人应用范围的扩大和机器人技术的发展，机器人的自动化、智能化和网络化的程度也越来越高，所能实现的功能也越来越多，性能越来越好。机器人技术的内涵已变为“灵活应用机器人技术的、具有实在动作功能的智能化系统。”目前，工业机器人技术正在向智能机器和智能系统的方向发展，其发展趋势主要为：结构的模块化和可重构化；控制技术的开放化、PC 化和网络化；伺服驱动技术的数字化和分散化；多传感器融合技术的实用化；工作环境设计的优化和作业的柔性化以及系统的网络化和智能化等方面。 1.2机器视觉在工业机器人中的应用 工业机器人是FMS（柔性加工）加工单元的主要组成部分，它的灵活性和柔性使其成为自动化物流系统中必不可少的设备，主要用于物料、工件的装卸、分捡和贮运。目前在全世界有数以百万的各种类型的工业机器人应用在机械制造、零件加工和装配及运输等领域，
电信学院毕业设计（论文）任务书 题目分布式发电并网系统潮流计算与分析 学生姓名张吉俊班级电气11级4班学号11230429 题目类型科学研究指导教师张晓英系主任 一、毕业设计（论文）的技术背景和设计依据： 分布式发电（DG-Distributed Generation）指的是直接接入配电网或用户侧的发电系统，功率等级一般在几十KW到几十MW之间，经济、高效、可靠地发电。根据使用的技术进行分类，DG主要可分为以下几类：（1）以天然气为常用燃料的燃气轮机、内燃机和微燃机等为基本核心的发电系统；（2）燃料电池发电系统；（3）太阳能光伏电池发电系统；（4）风力发电系统；（5）生物质能发电系统。分布式发电接入电网后，必然会改变电网的潮流分布，从而对节点电压，网络损耗产生重大的影响，其电压、潮流分布不仅取决于负荷，而且取决于分布式发电,而潮流计算是对其影响进行量化的主要手段。因此，研究分布式发电并网系统潮流计算与分析具有重要的实际意义。 1.原始资料 图1 20节点配电网络图 沿馈线将每一集中负荷视为一个节点并加以编号，从地区性变电站母线开始依次编为1，2，…，N（N=20），每一段线路的阻抗、各节点负荷大小如表1所示，配电网中节点1为平衡节点，电压取为1.02，基准值为10KV，功率基准值为100MV A。 节点电阻电抗有功无功 1 0.000 2 0.0005 0.0497 0.0261 2 0.0082 0.0246 0.0034 0.0065 3 0.0006 0.0017 0.0487 0.0002 4 0.0009 0.0027 0.0461 0.0163 5 0.0070 0.0209 0.0479 0.0180 6 0.0015 0.0044 0.0282 0.0497 7 0.0006 0.0017 0.0475 0.0217 8 0.0072 0.0219 0.0159 0.0217 9 0.0072 0.0217 0.0133 0.0468
基于机器视觉的产品检测技术研究 1、机器视觉 1.1机器视觉的概念 机器视觉被定义为用计算机来模拟人的视觉功能，从客观事物的图像中提取信息，进行处理并加以理解，最终用于实际检测、测量和控制。一个典型的工业机器视觉应用系统包括光源、光学系统、图像采集系统、数字图像处理与智能判断决策模块和机械控制执行模块。系统首先通过CCD相机或其它图像拍摄装置将目标转换成图像信号,然后转变成数字化信号传送给专用的图像处理系统，根据像素分布!亮度和颜色等信息，进行各种运算来抽取目标的特征,根据预设的容许度和其他条件输出判断结果。 值得一提的是，广义的机器视觉的概念与计算机视觉没有多大区别，泛指使用计算机和数字图像处理技术达到对客观事物图像的识别、理解。而工业应用中的机器视觉概念与普通计算机视觉、模式识别、数字图像处理有着明显区别，其特点是： 1、机器视觉是一项综合技术，其中包括数字图像处理技术、机械工程技术、控制技术、电光源照明技术，光学成像技术、传感器技术、模拟与数字视频技术、计算机软硬件技术、人机接口技术等。这些技术在机器视觉中是并列关系。相互协调应用才能构成一个成功的工业机器视觉应用系统。 2、机器视觉更强调实用性，要求能够适应工业生产中恶劣的环境，要有合理的性价比，要有通用的工业接口，能够由普通工作者来操作，有较高的容错能力和安全性，不会破坏工业产品，必须有较强的通用性和可移植性。 3、对机器视觉工程师来说，不仅要具有研究数学理论和编制计算机软件的能力，更需要光、机、电一体化的综合能力。 4、机器视觉更强调实时性,要求高速度和高精度，因而计算机视觉和数字图像处理中的许多技术目前还难以应用于机器视觉，它们的发展速度远远超过其在工业生产中的实际应用速度。 1.2机器视觉的研究范畴 从应用的层面看，机器视觉研究包括工件的自动检测与识别、产品质量的自动检测、食品的自动分类、智能车的自主导航与辅助驾驶、签字的自动验证、目标跟踪与制导、交通流的监测、关键地域的保安监视等等。从处理过程看,机器视觉分为低层视觉和高层视觉两阶段。低层视觉包括边缘检测、特征提取、图像分割等，高层视觉包括特征匹配、三维建模、形状分析与识别、景物分析与理解等。从方法层面看，有被动视觉与主动视觉之，又有基于特征的方法与基于模型的方法之分。从总体上来看，也称作计算机视觉。可以说，计算机视觉侧重于学术研究方面，而机器视觉则侧重于应用方面。 机器人视觉是机器视觉研究的一个重要方向，它的任务是为机器人建立视觉系统，使得机器人能更灵活、更自主地适应所处的环境，以满足诸如航天、军事、工业生产中日益增长的需要（例如，在航天及军事领域对于局部自主性的需要，在柔性生产方式中对于自动定位与装配的需要，在微电子工业中对于显微结构的检测及精密加工的需要等）。机器视觉作为一门工程学科，正如其它工程学科一样，是建立在对基本过程的科学理解之上的。机器视觉系统的设计依赖于具体的问题，必须考虑一系列诸如噪声、照明、遮掩、背景等复杂因素，折中地处理信噪比、分辨率、精度、计算量等关键问题。 1.3机器视觉的研究现状 机器视觉研究出现于60年代初期，电视摄像技术的成熟与计算机技术的发展使得机器视觉研究成为可能。它作为早期人工智能研究的一部分，由于技术条件的限制，进展缓慢。80年代初，在D·Marr提出的计算视觉理论指导下，机器视觉研究得到了迅速发展，成为
北语14秋《高级英语I》导学资料一 Unit1, Unit2& Unit3 一、本阶段学习内容概述 各位同学，大家好，本课程第一阶段学习的主要内容为Unit1：Party Politics, Unit2：The New Singles, Unit3：教材课文：Doctor’s Dilemma: Treat or Let Die? 网络课件课文：Computer Violence中包括课前练习（Warm-up）、单词和词组（New words and phrases）、课文（Text）、课后练习（Exercises）及补充阅读（Supplementary readings）中的指定内容。 课前练习：大家应先了解课前练习的要求，根据已有的知识思考其中问题，或者利用网络与同学开展一些讨论，争取在阅读课文前了解文章主要论述的问题，有利于更好的了解作者的思想观点和思维过程，从而了解文章所反映的思想文化，这样既能提高阅读理解能力又能获取知识和信息。 单词和词组:名词、动词和形容词是词汇练习和记忆中的重要部分。Unit 1、2、3中所列出的新单词绝大部分都是这三类，因此，大家一定要掌握好新单词。要开发利用多种方法记单词，如联想法、音节法、构词法等。单词和词组基本上给出了英语直接释义，可以培养大家英语思维的习惯。如果在阅读释义后还有疑问，一定要查阅英语词典，寻找一些相关的解释来加深对单词的理解和记忆。许多词的释义中给出了若干同义词或近义词，能帮助大家迅速扩大词汇量，同时，课后练习的词汇（Vocabulary Study）部分又给出了一些相关的练习，大家可以在学习完单词和词组后，乘热打铁，立即做词汇练习的第一小题（选择填空）和第二小题（找出意义相近的替代词）这样可以及时巩固所学单词，大概了解这些词的用法，为正确阅读课文打下基础，也能使得练习题做起来不是那么难。 （特别说明：高级英语阶段的学习，提高词汇量和词汇运用能力是一个很大，并且很重要，同时又是比较难的问题，这是我们必须面对的问题，所以，请大家务必花时间多熟悉单词。所谓的磨刀不误砍材功，先熟悉单词和短语，才能流畅的阅读文章。更关键的是，单词和短语在考试中所占比例不少！） 课文：每单元有一篇课文（Unit1：Party Politics, Unit2：The New Singles, Unit3：教材课文：Doctor’s Dilemma: Treat or Let Die? 网络课件课文：Computer Violence）。在掌握单词和词组之后，阅读课后注释，学习课文的背景资料、作者介绍和相关内容，如人物、事件、地点等的解释，这能帮助大家准确快速的理解文章的内容。在课件资源的帮助下，认真学习课文，包括课文中常用词语和句型的用法。文章比较长，大家一定要有耐心和毅力，坚持就是胜利。在学习完整篇文章后，及时完成课文理解（Comprehension Check）的练习。其中第一小题是根据课文内容选择最佳答案，第二小题是将部分文中的句子用英语注释。只要认真学习了课件资料，相信能很快准确地完成。同时也考察大家对课文理解的程度，督促大家很好的阅读课文。 课后练习：共有四个大题。 1.课文理解（Comprehension Check）：有两个题型。在借助课件学习完课文之后，大家可以先自己做这一部分的练习，然后再看课件，对正答案的同时，再重温课文的大意。 2.词汇（Vocabulary Study）：有两个题型。在学完课文和单词后，大家可以自己先做这一部分的词汇练习，不会做得可以看课件，并牢固掌握，对于补充的单词也要掌握。这样有利于快速扩充词汇量。 3.翻译（Translation）：有的单元是汉译英，有的单元是英译汉。都是一段与课文内容相近的短文。先认真思考，仔细看文章，如果有一定的难度，可以参考课件的解说，然后再组织语言完成翻译练习。
“异化”与“归化”之间的关系并评述 1、什么是归化与异化 归化”与“异化”是翻译中常面临的两种选择。钱锺书相应地称这两种情形叫“汉化”与“欧化”。A．归化 所谓“归化”（domestication 或target-language-orientedness），是指在翻译过程中尽可能用本民族的方式去表现外来的作品；归化翻译法旨在尽量减少译文中的异国情调，为目的语读者提供一种自然流畅的译文。Venuti 认为，归化法源于这一著名翻译论说，“尽量不干扰读者，请作者向读者靠近” 归化翻译法通常包含以下几个步骤：（1）谨慎地选择适合于归化翻译的文本；（2）有意识地采取一种自然流畅的目的语文体；（3）把译文调整成目的语篇体裁；（4）插入解释性资料；（5）删去原文中的实观材料；（6）调协译文和原文中的观念与特征。 B．“异化”（foreignization或source-language-orientedness）则相反，认为既然是翻译，就得译出外国的味儿。异化是根据既定的语法规则按字面意思将和源语文化紧密相连的短语或句子译成目标语。例如，将“九牛二虎之力”译为“the strength of nine bulls and two tigers”。异化能够很好地保留和传递原文的文化内涵，使译文具有异国情调，有利于各国文化的交流。但对于不熟悉源语及其文化的读者来说，存在一定的理解困难。随着各国文化交流愈来愈紧密，原先对于目标语读者很陌生的词句也会变得越来越普遍，即异化的程度会逐步降低。 Rome was not built in a day. 归化:冰冻三尺,非一日之寒. 异化:罗马不是一天建成的. 冰冻三尺,非一日之寒 异化:Rome was not built in a day. 归化:the thick ice is not formed in a day. 2、归化异化与直译意译 归化和异化，一个要求“接近读者”，一个要求“接近作者”，具有较强的界定性；相比之下，直译和意译则比较偏重“形式”上的自由与不自由。有的文中把归化等同于意译，异化等同于直译，这样做其实不够科学。归化和异化其实是在忠实地传达原作“说了什么”的基础之上，对是否尽可能展示原作是“怎么说”，是否最大限度地再现原作在语言文化上的特有风味上采取的不同态度。两对术语相比，归化和异化更多地是有关文化的问题，即是否要保持原作洋味的问题。 3、不同层面上的归化与异化 1、句式 翻译中“归化”表现在把原文的句式（syntactical structure）按照中文的习惯句式译出。
高级英语写作1-10课翻译
The Delicate Art of the Forest 库珀的创造天分并不怎么样；但是他似乎热衷于此并沾沾自喜。确实，他做了一些令人感到愉快的事情。在小小的道具箱内，他为笔下的森林猎人和土人准备了七八种诡计或圈套，这些人以此诱骗对方。利用这些幼稚的技巧达到了预期的效果，没有什么更让他高兴得了。其中一个就是他最喜欢的，就是让一个穿着鹿皮靴的人踩着穿着鹿皮靴敌人的脚印，借以隐藏了自己行踪。这么做使库珀磨烂不知多少桶鹿皮靴。他常用的另一个道具是断树枝。他认为断树枝效果最好，因此不遗余力地使用。在他的小说中，如果哪章中没有人踩到断树枝惊着两百码外的印第安人和白人，那么这一节则非常平静/那就谢天谢地了。每次库珀笔下的人物陷入危险，每分钟绝对安静的价格是4美元/一分静一分金，这个人肯定会踩到断树枝。尽管附近有上百种东西可以踩，但这都不足以使库珀称心。他会让这个人找一根干树枝；如果找不到，就去借一根。事实上，《皮袜子故事系列丛书》应该叫做《断树枝故事集》。 很遗憾，我没有足够的篇幅，写上几十个例子，看看奈迪·班波和其他库伯专家们是怎样运用他的森林中的高招。大概我们可以试着斗胆举它两三个例子。库伯曾经航过海—当过海军军官。但是他却一本正经/煞有介事地告诉我们，一条被风刮向海岸遇险的船，被船长驶向一个有离岸暗流的地点而得救。因为暗流顶着风，把船冲了回来。看看这森林术，这行船术，或者叫别的什么术，很高明吧？库珀在炮兵部队里待过几年，他应该注意到炮弹落到地上时，要么爆炸，要么弹起来，跳起百英尺，再弹再跳，直到跳不动了滚几下。现在某个地方他让几个女性—他总是这么称呼女的—在一个迷雾重重的夜晚，迷失在平原附近一片树林边上—目的是让班波有机会向读者展示他在森林中的本事。这些迷路的人正在寻找一个城堡。他们听到一声炮响，接着一发炮弹就滚进树林，停在他们脚下。对女性，这毫无价值。但对可敬的班波就完全不同了。我想，如果班波要是不马上冲出来，跟着弹痕，穿过浓雾，跨过平原，找到要塞，我就再也不知道什么是“和平”了。是不是非常聪明？如果库伯不是对自然规律一无所知，他就是故意隐瞒事实。比方说，他的精明的印地安专家之一，名叫芝稼哥（我想，该读作芝加哥）的，跟踪一个人，在穿过树林的时候，脚印就找不到了。很明显，脚印是再也没法找到了。无论你还是我，都猜不出，怎么会
万方数据
万方数据
万方数据
万方数据
翻译的归化与异化 作者：熊启煦 作者单位：西南民族大学,四川,成都,610041 刊名： 西南民族大学学报（人文社科版） 英文刊名：JOURNAL OF SOUTHWEST UNIVERSITY FOR NATIONALITIES（HUMANITIES AND SOCIAL SCIENCE） 年，卷(期)：2005,26(8) 被引用次数：14次 参考文献(3条) 1.鲁迅且介亭杂文二集·题未定草 2.刘英凯归化--翻译的歧路 3.钱钟书林纾的翻译 引证文献(15条) 1.郭锋一小议英语翻译当中的信达雅[期刊论文]-青春岁月 2011(4) 2.许丽红论汉英语言中的文化差异与翻译策略[期刊论文]-考试周刊 2010(7) 3.王笑东浅谈汉英语言中的差异与翻译方法[期刊论文]-中国校外教育（理论） 2010(6) 4.王宁中西语言中的文化差异与翻译[期刊论文]-中国科技纵横 2010(12) 5.鲍勤.陈利平英语隐喻类型及翻译策略[期刊论文]-云南农业大学学报(社会科学版) 2010(2) 6.罗琴.宋海林浅谈汉英语言中的文化差异及翻译策略[期刊论文]-内江师范学院学报 2010(z2) 7.白蓝跨文化视野下文学作品的英译策略[期刊论文]-湖南社会科学 2009(5) 8.王梦颖探析汉英语言中的文化差异与翻译策略[期刊论文]-中国校外教育（理论） 2009(8) 9.常晖英汉成语跨文化翻译策略[期刊论文]-河北理工大学学报（社会科学版） 2009(1) 10.常晖对翻译文化建构的几点思考[期刊论文]-牡丹江师范学院学报（哲学社会科学版） 2009(4) 11.常晖认知——功能视角下隐喻的汉译策略[期刊论文]-外语与外语教学 2008(11) 12.赵勇刚汉英语言中的文化差异与翻译策略[期刊论文]-时代文学 2008(6) 13.常晖.胡渝镛从文化角度看文学作品的翻译[期刊论文]-重庆工学院学报(社会科学版) 2008(7) 14.曾凤英从文化认知的视角谈英语隐喻的翻译[期刊论文]-各界 2007(6) 15.罗琴.宋海林浅谈汉英语言中的文化差异及翻译策略[期刊论文]-内江师范学院学报 2010(z2) 本文链接：https://www.doczj.com/doc/7116004458.html,/Periodical_xnmzxyxb-zxshkxb200508090.aspx
英语专业(高级翻译)简介 ·日期：2007-06-07 ·点击次数： 1446 培养目标：本专业培养具有扎实的英语语言基础，较强的语言交际能力和跨文化交际能力，掌握多方面的翻译知识和技巧，能熟练地运用英、汉语在外事、外交、经贸、文化、科技等部门从事口译、笔译的专门人才。 知识能力 要求：1．语音、语调准确，清楚自然。 2．词法、句法、章法(包括遣词造句与谋篇布局)规范、表达得体。 3．认知词汇10000~12000，且能正确而熟练地使用其中的5000~6000个单词及其最常用的搭配。 4．听、说、读、写、译技能熟练，具有较强的英语综合运用能力。毕业时达到英语专业八级水平。 听的能力：听懂真实交际场合中各种英语会话；听懂英 语国家广播电台以及电视台(如VOA, BBC, CNN)有关政治、经济、文化、教育、科技 等方面的新闻、现场报道、专题报道、综合 评述以及与此类题材相关的演讲和演讲后 的问答；听懂电视时事报道和电视短剧中的 对话。语速为每分钟150~180个单词，理解 准确率不低于60%。 说的能力：能就国内外重大问题与外宾进行流利而得体 的交流；能系统、深入、连贯地发表自己的 见解。 读的能力：能读懂一般英美报刊杂志上的社论和书评、 英语国家出版的有一定难度的历史传记和 文学作品；能分析上述题材文章的思想观 点、语篇结构、语言特点和修辞手法。能在
5分钟内速读1600词左右的文章，掌握文章 的主旨和大意，理解事实和细节。 写的能力：能写各类体裁的文章，做到内容充实，语言 通顺，用词恰当，表达得体。写作速度为30 分钟300~400个单词。能撰写长度为 5000~6000个单词的毕业论文，要求思路清 晰、内容充实、语言通顺。 译的能力：1)笔译：能运用翻译的理论和技巧，将英美报 刊上的文章、文学、科技、文化、经贸方面 的原文译成汉语，或将我国报刊、杂志上的 文章、一般文学作品、科技、文化、经贸方 面的原文译成英语，速度为每小时250~300 个单词。译文要求忠实原意，语言流畅。 2)口译：掌握连续或同步口译所需的技巧及基 本要求，能担任外经外贸、外事、外交、国 际文化、科技交流活动的口译和国际会议的 同声传译任务。 5．具有宽广的英语专业知识和相关学科知识。英语专业知识包括文学、语言学和对象国社会与文化的知识。相关学科的知识涉及教育学、教育心理学、语言习得理论、英语教学法、英语测试理论与实践、课程设置等领域。6．具有获取知识的能力、运用知识的能力、分析问题的能力、独立提出见解的能力和创新的能力。 7．熟悉中、英两种文字计算机基本技术并能实际应用。 主干课程：基础英语、高级英语、散文选读、英语泛读、英语口语、基础英语写作、高级英语写作、英汉/汉英口译、英汉笔译、英语视听说、英语报刊选读、英美文学，同声传译、连续传译、经贸法律翻译等。 修业年4年
翻译作业10 Nov 15 一、请按归化法(Domestication)翻译下列习语。 Kill two birds with one stone a wolf in sheep’s clothing strike while the iron is hot. go through fire and water add fuel to the flames / pour oil on the flames spring up like mushrooms every dog has his day keep one’s head above water live a dog’s life as poor as a church mouse a lucky dog an ass in a lion’s skin a wolf in sheep’s clothing Love me, love my dog. a lion in the way lick one’s boots as timid as a hare at a stone’s throw as stupid as a goose wet like a drown rat as dumb as an oyster lead a dog’s life talk horse One boy is a boy, two boys half a boy, and three boys nobody. Man proposes, God disposes. Cry up wine and sell vinegar (cry up, to praise; extol: to cry up one's profession) Once bitten, twice shy. An hour in the morning is worth two in the evening． New booms sweep clean. take French leave seek a hare in a hen’s nest have an old head on young shoulder Justice has long arms You can’t teach an old dog Rome was not built in a day. He that lives with cripples learns to limp. Everybody’s business is nobody’s business. The more you get, the more you want. 二、请按异化法(foreignization)翻译下列习语。 Kill two birds with one stone a wolf in sheep’s clothing
机器视觉概念/研究现状/应用/检测 内容来源网络，由“深圳机械展（11万㎡，1100多家展商，超10万观众）”收集整理！ 更多cnc加工中心、车铣磨钻床、线切割、数控刀具工具、工业机器人、非标自动化、数字化无人工厂、精密测量、3D打印、激光切割、钣金冲压折弯、精密零件加工等展示，就在深圳机械展. 1、机器视觉 1.1机器视觉的概念 机器视觉被定义为用计算机来模拟人的视觉功能，从客观事物的图像中提取信息，进行处理并加以理解，最终用于实际检测、测量和控制。一个典型的工业机器视觉应用系统包括光源、光学系统、图像采集系统、数字图像处理与智能判断决策模块和机械控制执行模块。系统首先通过CCD相机或其它图像拍摄装置将目标转换成图像信号,然后转变成数字化信号传送给专用的图像处理系统，根据像素分布!亮度和颜色等信息，进行各种运算来抽取目标的特征,根据预设的容许度和其他条件输出判断结果。 值得一提的是，广义的机器视觉的概念与计算机视觉没有多大区别，泛指使用计算机和数字图像处理技术达到对客观事物图像的识别、理解。而工业应用中的机器视觉概念与普通计算机视觉、模式识别、数字图像处理有着明显区别，其特点是： 1、机器视觉是一项综合技术，其中包括数字图像处理技术、机械工程技术、控制技术、电光源照明技术，光学成像技术、传感器技术、模拟与数字视频技术、计算机软硬件技术、人机接口技术等。这些技术在机器视觉中是并列关系。相互协调应用才能构成一个成功的工业机器视觉应用系统。 2、机器视觉更强调实用性，要求能够适应工业生产中恶劣的环境，要有合理的性价比，
要有通用的工业接口，能够由普通工作者来操作，有较高的容错能力和安全性，不会破坏工业产品，必须有较强的通用性和可移植性。 3、对机器视觉工程师来说，不仅要具有研究数学理论和编制计算机软件的能力，更需要光、机、电一体化的综合能力。 4、机器视觉更强调实时性,要求高速度和高精度，因而计算机视觉和数字图像处理中的许多技术目前还难以应用于机器视觉，它们的发展速度远远超过其在工业生产中的实际应用速度。 1.2机器视觉的研究范畴 从应用的层面看，机器视觉研究包括工件的自动检测与识别、产品质量的自动检测、食品的自动分类、智能车的自主导航与辅助驾驶、签字的自动验证、目标跟踪与制导、交通流的监测、关键地域的保安监视等等。从处理过程看,机器视觉分为低层视觉和高层视觉两阶段。低层视觉包括边缘检测、特征提取、图像分割等，高层视觉包括特征匹配、三维建模、形状分析与识别、景物分析与理解等。从方法层面看，有被动视觉与主动视觉之，又有基于特征的方法与基于模型的方法之分。从总体上来看，也称作计算机视觉。可以说，计算机视觉侧重于学术研究方面，而机器视觉则侧重于应用方面。 机器人视觉是机器视觉研究的一个重要方向，它的任务是为机器人建立视觉系统，使得机器人能更灵活、更自主地适应所处的环境，以满足诸如航天、军事、工业生产中日益增长的需要（例如，在航天及军事领域对于局部自主性的需要，在柔性生产方式中对于自动定位与装配的需要，在微电子工业中对于显微结构的检测及精密加工的需要等）。机器视觉作为一门工程学科，正如其它工程学科一样，是建立在对基本过程的科学理解之上的。机器视觉系统的设计依赖于具体的问题，必须考虑一系列诸如噪声、照明、遮掩、背景等复杂因素，折中地处理信噪比、分辨率、精度、计算量等关键问题。
青岛大学 毕业论文(设计)开题报告 院系: 自动化工程学院电气系 专业: 电气工程及其自动化 班级: 电气一班 学生姓名: 祝昆 指导教师: 马力 2012年3月18 日
1.文献综述 1.1 电力系统仿真技术的发展 随着社会对能源需求的增长和发电技术的进步, 现代电力系统发展很快, 大容量发电机组不断涌现, 自动化程度更趋提高, 计算、监视及控制问题日益复杂, 这就需要运行人员具有更强的应变能力和更熟练地操作。新的电气研究也需做各种试验, 但无论从现有技术上还是从供电的可靠性及设备的安全性考虑, 直接在实际的电力系统及厂矿企业变电所中进行操作人员的培训和科学研究, 可能性很小, 因此运行电力系统仿真技术脱离现场对运行人员培训及电气研究成了迫切的需要。 电力系统仿真技术的研究可追塑到50 年代, 最早的电力系统仿真设备可认为是直流计算台, 以后出现支流计算台, 主要用做短路、潮流及稳定计算。真正成为实用技术是从60 年代末70 年代初用于电气培训。美、日、英等国推出了核电、火电仿真系统, 培训运行人员效果良好, 各国纷纷仿效。1975 年美国编制了第一个全复制型培训仿真器国家标准。目前国外电厂培训仿真技术的研制和开发已日趋成熟。 我国清华大学1983 年研制成功了国产200MW 燃煤机组原理型培训系统, 同年又以哈尔滨电厂机组为原理, 研制成功200MW 全复制型培训系统。目前电力部已作出规定, 大型火电机组操作人员应经过仿真培训取得合格证方可录用。部分地区(如重庆) 的厂矿企业变电所值班人员也分批参加仿真培训。现在国内各大电业局几乎都已经或准备购置培训用仿真装置, 成立仿真培训中心。 电力仿真技术在电力系统中的发展和应用与电力发展对仿真的要求、计算机硬件、软件技术发展密切相关, 某些电力仿真装置就象一个缩小的电力系统, 凡是与电力相关部门的技术人员, 通过对仿真技术的学习, 了解及参加仿真培训, 就能迅速地熟悉电力系统, 掌握电力运行操作的一些基本方法。 电力系统是一个大规模、时变的复杂系统,而且在国民经济中有非常重要的作用,电力系统数字仿真技术已成为电力系统研究、规划和设计的重要手段。因此电力系统仿真软件的功能特性对于提高研究、设计的效率和可信性有重要作用。随着现代电力系统网络规模的不断扩大和电网电压等级的不断升高,电力系统规划、运行和控制的复杂性亦日益增加,因此在电力系统的生产和研究中仿真软件的应用越来越广泛。 1.2 PSASP仿真软件简介 《电力系统分析综合程序》（Power System Analysis Software Package，PSASP）是由中国电力科学研究院研发的电力系统分析程序。主要用于电力系统规划设计人员确定经济合理、技术可行的规划设计方案；运行调度人员确定系统运行方式、分析系统事故、寻求反事故措施；科研人员研究新设备、新元件投入系统等新问题以及高等院校用于教学和研究。 基于电网基础数据库、固定模型库以及用户自定义模型库的支持，PSASP可进行电力系统（输电、供电和配电系统）的各种计算分析。包括：稳态分析的潮流计算、网损分析、最优潮流和无功优化、静态安全分析、谐波分析、静态等值等；