下载文档原格式
Computer Knowledge and Technology 电脑知识与技术本栏目责任编辑：冯蕾第9卷第01期(2013年01月)物联网安全关键技术研究马卫（中南民族大学，湖北武汉430074）摘要：物联网的安全问题直接关系到物联网技术的发展和应用，在讨论了物联网体系架构的基础上，分析了物联网的安全需求，并对物联网安全的关键技术进行了研究，希望为建立可靠安全的物联网体系提供一定的参考作用。
关键词：物联网;安全需求;感知层中图分类号：TP393文献标识码：A 文章编号：1009-3044(2013)01-0029-03Research on the Key Technology of the Security of the Internet of ThingsMA Wei(South-Central University for Nationalities,Wuhan 430074,China)Abstract:Security issue of the IoT is directly related to the development and application.On the basis of discussion on the archi⁃tecture of the IoT,the paper analyze the security needs of it,and give the further study on the key technology,expect to providecertain reference to the establishment of reliable and secure networking system.Key words:internet of things;security needs;sensing layer随着计算机网络和通信技术的发展，一种新的网络——物联网应运而生，物联网是通过射频识别(RFID)装置、红外感应器、全球定位系统、激光扫描器、传感器节点等信息传感设备，按约定的协议，把任何物品与互联网相连接，进行信息交换和通信，以实现智能化识别、定位跟踪、监控和管理等功能的一种网络[1]。
本栏目责任编辑：唐一东人工智能及识别技术Computer Knowledge and Technology 电脑知识与技术第5卷第13期(2009年5月)一种新的基于图像边缘的插值算法刘艳琪，罗扬（南华大学计算机科学与技术学院，湖南衡阳421001）摘要：灰度图像放大时，插值所具有的平滑作用会退化图像的高频部分，使放大图像轮廓变得模糊，该文提出了一种新的图像插值算法，先用一阶微分运算分离出图像的平坦区域和边缘区域，对图像的平坦区域采用双线性插值法，对图像的边缘部分采用本文的插值算法，实验证明该算法有效地保持了边缘信息，得到了视觉信息较好的插值图像。
关键词：图像插值；边缘保持；距离密次反比中图分类号：TP391文献标识码：A 文章编号：1009-3044(2009)13-3470-02A New Edge-directed Image Interpolation AlgorithmLIU Yan-qi,LUO Yang(College of Computer Science and Technology,University of South China,Hengyang 421001,China)Abstract:When gray images are magnified,the smoothing effect of gray level interpolation may degrade the fine details in images and make contours become blurred .A new image interpolation algorithm is proposed in this paper ,which is divided by a first-order differential op -erator and interpolates flat and edge regions respectively.In flat regions the bilinear interpolation is applied,and the interpolation algorithm is employed in edge regions.Experimental results demonstrate that the algorithm not only preserves the edge effectively,but also improve the visual quality of the interpolated images.Key words:image interpolation;edge-preserving;inverse distance weighted1引言成像系统所记录的分辨率往往无法达到实际图像处理应用的要求，因而本文围绕基于低分辨率图像重建超分辨率的图像这一问题展开讨论，图象插值在其中起着重要的作用，传统的插值算法有最近邻插值[1]，双线性插值[2]，三次样条插值[3]等，这些方法侧重于图像的平滑，从而取的更好的视觉效果，但这类方法在保持图像平滑的同时常常导致图像边缘的模糊，而图像的边缘信息是影响视觉效果的重要因素，因而基于边缘方向的插值技术成为近年来研究的热点。
计算机专业副高职称评定承认期刊《电脑与电信》（省级）一.期刊简介：《电脑与电信》杂志创刊于1995年，是经国家新闻出版署正式批准、由广东省科技厅主管、广东省对外科技交流中心主办，全国发行的省级期刊，国内外公开发行，原名为《广东电脑与电讯》，2003年改名为《电脑与电信》。
从2006年起，本刊主要刊登计算机、互联网、通讯、电子等相关领域的理论与实践研究文章，以及相关理论研究综述和介绍外国先进技术的论文。
欢迎广大读者踊跃投稿和订阅！本刊为“中国学术期刊综合评价数据统计源期刊”、“ 中国期刊全文数据库全文收录期刊”、“万方数据库收录期刊”和“中文科技期刊数据库收录期刊”。
二.主要栏目:基金项目、算法研究、安全技术、系统开发、数据库技术、网络与通信、硬件与软件、电子工程、经验交流，①全球IT新浪潮：通过对国际上发展较快的初创企业进行全面的跟踪和报道，从而窥探出世界科技的发展潮流，把握行业动态，为我国企业的发展提供一些启发和经验，有较强的前瞻性和开拓性；投稿咨询QQ：869702190②国际合作项目：本刊与中国驻外使领事馆以及国外相关机构建立信息共享机制，定期发布国外最新的项目研发动态，为国内相关领域的企业提供合作项目的信息，进一步加强电脑与电信领域的对外科技合作与交流；③新闻链接：主要是报道最新的IT类新闻，让读者掌握一手的IT类资讯；④学术探讨：开拓学术新思路，探讨在理论研究中有创新内容的研究成果，或者在实践研究中取得创新成绩的有应用价值的实践成果，要求有方法、观点、比较和实验分析。
⑤读者对象：IT技术人员、大专院校师生、企事业单位、科研院所从事电脑与电信领域开发应用人员等，是启迪思维、开拓进取、更新知识的良师益友。
《电脑知识与技术》征稿（省级）刊名：电脑知识与技术（Computer Knowledge and Technology(Academic Exchange) 主办：安徽科技情报协会;中国计算机函授学院周期：旬刊出版地：安徽省合肥市语种：中文;开本：大16开ISSN：1009-3044CN：34-1205/TP邮发代号：26-188历史沿革：现用刊名：电脑知识与技术曾用刊名：电脑知识与技术.学术交流;电脑知识& 计算机文汇创刊时间：1994一.期刊简介：全国唯一的专业IT 类旬刊刊物是由安徽省科技情报学会，中国计算机函授学院主办的一本融知识，技术，信息于一体的电脑类杂志，杂志内容被中国学术期刊网、万方数据、维普资讯、QQ 网络杂志全文收录。
《大学计算机基础（Windows 7+WPS Office 2019 ）（微课版）》试卷班级：________________姓名：________________一、填空题(共5题，每题1分。
)1．WPS文档中段落的对齐方式包括5种，分别是左对齐、右对齐、两端对齐、__________和__________。
2．要清除单元格中的内容，只需选中单元格或单元格区域后按__________键。
3．WPS表格界面中的名称框用于____________________；编辑栏用于____________________。
4．选定多个连续的文件或文件夹，操作步骤为:单击所要选定的第一个文件或文件夹，然后按住__________键，单击最后一个文件或文件夹。
5．单击“改写”按钮或按__________键，可在文本的改写状态和插入状态间转换。
二、单项选择题(共25题，每题1分。
)1．对单元格中的公式进行复制时，会发生变化的是（）。
A．相对地址中的偏移量B．相对地址所引用的单元格C．绝对地址中的地址表达式D．绝对地址所引用的单元格2．E2单元格对应于工作表的（）行、列。
A．5,2B．4,3C．2,5D．5,33．在WPS文档中，若想删除单个字符，并且光标在该字符后，可使用（）方法。
A．按Backspace键B．按Delete键C．按Tab键D．按空格键4．用户可以通过（）来进行窗口切换。
A．单击任务栏中的按钮B．按“Alt+Tab”组合键C．按“Win+Tab”组合键D5．计算机软件总体分为系统软件和（）。
A．非系统软件B．重要软件C．应用软件D．工具软件6．在Window s7中，下列叙述错误的是（）。
A．可支持鼠标操作B．可同时运行多个程序C．不支持即插即用D．桌面上可同时容纳多个窗口7．在WPS表格中存储和处理数据的文件是（）。
A．工作簿B．工作表C．单元格D．活动单元格8．在WPS文档中，打印预览和实际打印效果（）A．完全相同B．完全不同C．部分相同D．绝大部分相同9．在工作表区域使用复制或剪切命令后会出现虚线边框，如何将其去掉（）。
I.J. Information Technology and Computer Science, 2019, 9, 48-54Published Online September 2019 in MECS (/)DOI: 10.5815/ijitcs.2019.09.06Text Extraction from Natural Scene Images usingOpenCV and CNNVaibhav Goel, Vaibhav Kumar, Amandeep Singh Jaggi, Preeti NagrathComputer Science Department, Bharati Vidyapeeth’s College of Engineering, New Delhi, IndiaE-mail: vaibhavgoel17@, kvaibhavs077@, amandeep998877@,preetinagrath1@Received: 06 June 2019; Accepted: 24 June 2019; Published: 08 September 2019Abstract—The influence of exponentially increasing camera-embedded smartphones all around the world has magnified the importance of computer vision tasks, and gives rise to a vast number of opportunities in the field. One of the major research areas in this field is the extraction of text embedded in natural scene images. Natural scene images are the images taken from a camera, where the background is random, and the variety of colors used in the image may be diverse. When text is present in such type of images, it is usually difficult for a machine to detect and extract this text due to a number of parameters. This paper presents a technique that uses a combination of the Open Source Computer Vision Library (OpenCV) and the Convolutional Neural Networks (CNN), to extract English text from images efficiently. The CNN model is based on a two-stage pipeline that uses a single neural network to directly detect the characters in the scene images. It eliminates the unnecessary intermediate steps that are present in the previous approaches to this task making them slower and inaccurate, thereby improving the time complexity and the performance of the algorithm.Index Terms—Text Extraction, Deep Learning, OpenCV, natural scene images, CNN, Optical Character Recognition.I.I NTRODUCTIONThe extraction of text from natural scene images is a demanding task owing to the wide variations in the text properties such as font, alignment, orientation, size, style, lightning conditions, color, and a bunch of other parameters. Also, the natural scene images usually suffer from low resolution and low quality, perspective distortion, non-uniformity in illumination, and complex background [1]. Moreover, there is a lack of prior knowledge of the text features and the location of the text regions. All these factors contribute to the difficulty in recognizing text from such images.The primary reason behind performing this task is the fact that the natural scene images may contain important written information that may not be able to reach certain people due to a number of problems like visual impairment, language barrier, etc., or that may be needed to feed to a computer without explicitly typing the information [2]. There may be situations when we want to write down some information present in the form of text in a hard document, an image, or a video. We can just take a snap of the document from the smartphone’s camera and in a couple of seconds, the algorithm will extract all the text present in the document. This will save time as well as effort to write down the whole information manually. Thus, autonomous text extraction from an image could prove to be beneficial. Some major applications of this task include document retrieving, vehicle number plate detection, road signs detection (by smart cars), page segmentation, video content summary, and intelligent driving assistance, to name a few [3]. Self-driving cars’ efficiency can be much improved if it knows more than just what is around it, like what is being written on destination boards or traffic signs of curves and diversion. It is also useful for providing assistance to visually impaired persons, thus improving magnificence of life for them. These are some of the greatest benefits of this task that can be beneficial for all.Now, the present OCR (Optical Character Recognition) techniques are able to attain exemplary accuracy on scanned images, i.e., images having text on a monotonous background. But, they still cannot extract text information accurately from images that are directly taken from a camera (i.e. natural scene images). In other words, the current OCR systems are able to handle text only with a monochrome background, where the text can easily be separated from the background, which is definitely not the case with natural scene images. On the other hand, the current techniques which can actually extract text from natural scene images still lack satisfactory accuracy. Thus, this research-based project tries to fill this gap by proposing a technique that uses a combination of the Open Source Computer Vision Library (OpenCV) and the Convolutional Neural Networks (CNN) for efficient text extraction. It also involves the study of different methods of text extraction from a given image, and a thorough comparison of their performances.The next section lists some of the past work done in this particular area, followed by our research and proposed methodology in section III, results in section IV, and finally conclusion in section V.Fig.1. Some examples of images containing text information taken from the Total-Text-Dataset [18]II.R ELATED W ORKSIn layman’s terms, at the highest level, this process of extraction of text from scene images is divided into two sub-processes or steps: the first step is Scene text detection, followed by the second step, which is Scene text recognition [4] (Fig.2).Fig.2. The two most basic steps of text extraction process.In the text detection process, the locations where the text is present in the image are identified. This process is also known as text localization [5]. It is the most difficult task because of the aforementioned challenges. The pre-processing of the image is done so as to make it suitable for feeding it to the feature detector model. The pre-processing involves tasks such as converting the image to grayscale, defining a specific width and height for the image that is optimum for the feature detector in which it has to be sent into, detecting the edges in the image using edge-detectors such as the Canny edge detector, etc. After the regions in which the text is present have been identified, the next step is text recognition, i.e., recognizing the text information present in those regions. These regions of image are sent into deep learning models that are vigorously trained to recognize characters, and consequently words. Different models are present to implement this task, which will be discussed later. It requires a lot of data to train these models to recognize different styles, strokes, and spacing between the letters or numbers.The output of these models will have the words or characters that were present in the image. This can help us to know and record data in low memory (text instead of image) and share the main point in the random image that is selected.Now, the present OCR (Optical Character Recognition) [6] techniques are able to attain exemplary accuracy on scanned images, i.e. images having text on a monotonous background. But, they still cannot extract text information accurately from images that are directly taken from a camera, i.e. natural scene images. In other words, the current OCR systems are able to handle text only with a monochrome background, where the text can easily be separated from the background, which is definitely not the case with natural scene images. Many researchers have done intensive research work in finding the methods for improving the accuracy of existing text recognition algorithms, but because of the immense variations seen in the natural scene images and the lack of prior knowledge of any kind of text features, it is still a challenging task to achieve almost perfect recognition rate. The next paragraph discusses some of the past work proposed or implemented in this area.There can be two types of approaches to this task of text detection: conventional approach and deep neural network based approach. The existing methods based on either approach are time consuming and not optimal due to the use of multiple stages and components in the process. The conventional approaches include techniques like Maximally Stable Extremal Regions (MSER) [15, 22] and Stroke Width Transform (SWT). These techniques generally use extremal region detection or edge detection to seek candidate characters in an image. The authors in [7] proposed an approach that employed a clustering method based on density values for the segmentation of candidate characters by combining color featuresofcharacter pixels and spatial connectivity. JiSoo Kim [8] proposed three text extraction techniques for natural scene images that were based on the intensity information of the input image. The 1st technique is “Gray Value Stretching” that involves binarization of the image by calculating the average intensity of all the pixels. The 2nd technique introduces the “Split and Merge” approach, which is a popular algorithm for image segmentation. The 3rd technique is derived from the amalgamation of the first two techniques. Busta et al. [9] made FASText (a fast scene text detector) by modifying the famous FAST key point detector which is used for stroke extraction. But, this technique was not as efficient as the ones based on deep neural networks. It lagged behind both in terms of accuracy and flexibility, especially in scenarios with low resolution. Coates et al. [10] introduced an unsupervised learning model that was integrated with a variation of the famous clustering algorithm, the K-means clustering. This model extracts the local features of character patches, followed by pooling them on the basis of cascading sub-patch features. Lu et al. [11] proposed a method that describes a dictionary of basic shape codes, modeling the inner character structure, to perform word and consequently character retrieval on scanned documents, without OCR. Huang et al. [12] proposed a technique with a two pipeline structure. The first step was to deduce the candidate features using the concept of Maximally Stable Extremal Regions (MSER). The second step included the use of a convolutional neural network as a strong classifier. It was employed to suppress false positives, i.e., the regions that does not contain any text but are still positively detected by the algorithm. Mishra et al. [13] acquired an approach that involves the concept of conditional random field, to combine bottom-up character recognition and top-down word-level recognition. Based on the Scale-Invariant Feature Transform (SIFT), Smith et al. [14] built a model of text detection that amplified the posterior probability of similarity constraints by making use of integer programming. SIFT is a corner detection technique invented by D.Lowe in 2004 to overcome the limitations of Harris Corner Detector. Neumann et al. [15] used the idea of extremal regions to propose a real-time text detection and extraction technique. Liu et al. [16] merged three models for the text recognition in scene images. The three models are: 1). the Gabor-based appearance model for feature detection, 2). the similarity model, and, 3). the Lexicon model.These were some of the major contributions to this field of text extraction that have set the bar high for new researchers and researches in the field. But, the best accuracy achieved till date is still not that great, and has a scope for improvement.III.M ETHODOLOGYThe authors have used a pre-trained deep learning model, the EAST text detector, for reference. EAST is short for “Efficient and Accuracy Scene Text” detection pipeline [17]. It is a fully-convolutional neural network proposed by Zhou et al. in 2017 in his paper “EAST: An Efficient and Accurate Scene Text Detector”. It is a model for determining the locations of text in natural scene images. The output of this model provides per-pixel predictions for characters or words. The model has the ability of running in real-time at 13 frames per second on images with 720p resolution. Also, it has achieved state-of-the-art text localization accuracy. The structure of the EAST text detector is shown in Fig.3 as per Zhou et al. [17].Fig.3. Structure of the EAST text detector Fully ConvolutionalNetwork [17].The CNN model is based on a two-stage pipeline that uses a single fully convolutional neural network to directly detect the characters in the scene images. The first stage includes feeding the image to the neural network. It directly detects the text regions in the image, eliminating the unnecessary intermediate steps that are present in the previous approaches [9, 23, 24, 25, 26] to this task making them slower and inaccurate, thereby improving the time complexity and the performance of the algorithm. The second stage is the Non-maximum suppression (NMS) stage which suppresses the multiple bounding boxes created around the same text region to just one box. This is the high level overview of thepipeline used.Fig.4. (a). Original Image (b). Converted to grayscale (c). After applying Canny Edge DetectionNow, when it comes to computer vision tasks, there is nothing that can beat OpenCV as of now. It is a specialized open source library specifically designed for developing Computer Vision applications.The methodology presented in the paper combines both the state-of-the-art technologies for image processing, the Convolutional Neural Networks (CNN) and the OpenCV library. The aforementioned trained deep learning model is exploited using OpenCV 3.4.2. The already trained tensorflow model named “frozen_east_text_detection” is used as a reference for the new model. There are two output layers in the model. The first layer is a sigmoid convolution layer named “feature_fusion/Conv_7/Sigmoid”. This layer outputs the regional probability, specifying whether that region is containing text information. The second output layer is named “feature_fusion/concat_3” which determines and outputs the coordinates of the bounding boxes containing the text information.The trained model is loaded into memory by feeding it to the OpenCV using the cv2.dnn.readnet() method. Now, before feeding the image to the model, it is essential to preprocess the image in order to obtain accurate results. OpenCV provides different preprocessing features via its dnn module’s blobFromImage function, which performs Mean subtraction, scaling by some factor and channel swapping.Now, the above blob produced out of the image is set as the input to the model that was previously loaded. On performing a forward pass on the CNN model, it generates as output two feature maps from the two output layers that have been discussed above. The first tells the probability of a particular region containing text, while the second determines the coordinates of bounding boxes containing text in the input image. The model might have detected regions that are having very less probability of containing text. Such regions that does not have sufficiently high probabilities should be ignored. To perform this task, the algorithm loops over all the probability scores that were obtained as output from the model. The regions that have probability less than 0.5 are ignored, while the regions having probability more than 0.5 are considered and the corresponding bounding box coordinates are plotted so as to obtain a box (rectangle) around the text.Now comes the process of text recognition. The regions having text in the image have been identified successfully at this stage. Those regions are now required to be fed into an OCR system. OCR stands for “Optical Character Recognition”. It i s a system that converts the images containing text into machine-encoded text, thus recognizing the text present in the images. For the OCR to produce accurate results, instead of sending the natural scene images as it is, only those portions of the image are sent that contains the text information. This is because OCR can’t handle images with complex background, and struggles to differentiate text regions from non-text regions. Thus, preprocessing is done on the regions detected by the above deep learning model, in order to alienate the text elements from the background, thereby maximizing the efficiency and accuracy of OCR. The preprocessing includes techniques like BGR to grayscale conversion, and canny edge detection.In OCR, a CNN model is employed that analyze image for contrast of light and dark areas to recognize characters or numeric digits. The lines are segmented into words and then into characters. On getting the characters, the algorithm compares those characters with the data of predefined pattern images set. As a character is recognized, it is then converted into an ASCII code.This is the complete algorithm that the authors have adopted for the efficient text extraction. The following steps summarize the complete process in a high-level manner.Step#1: Input Image.Step#2: Preprocessing of the input image.Step#3: Text detection and localization using a fully convolutional neural network.Step#4: Preprocessing of the detected text regions.Step#5: Text recognition using another CNN model (OCR).Now, after going through the complete process, the authors realized that this segmentation of work in two segment pipeline of detection and recognition can be achieved under a single network where: (1). end to end recognition is converted into a spatial transformer network in a semi supervised way which is a part of a major deep neural network (DNN), and, (2). the later part of the network is then trained for recognition of thecharacters in the text. A CNN model is built that takes aninput feature map A and perform spatial recognition on the image and produce two output map that gives the probability and location of the text in the image, respectively. The output maps are than fed to other algorithms for image sampling. The sampled image is then passed into another CNN model for recognition of the text from the predefined data set. This is known as OCR. The two models can be named as localization networks and recognition networks for the sake of ease of understanding how the networks are going to behave and in what fashion.IV. P ERFORMANCE E VALUATIONThere are two quantities, namely Precision Rate and Recall Rate, which are generally used to measure the performance of the algorithm in the task of text extraction. The precision rate is the ratio of the number of correctly detected words to the sum of the number of correctly detected words and false positives. False positives are the regions of the image that do not contain any text but still have been recognized by the algorithm.correctly detected words% Precision Rate 100correctly detected words false positives =⨯+The recall rate is the ratio of the number of correctly detected words to the sum of the number of correctly detected words and false negatives. False negatives are the regions of the image that do actually contain text but haven’t been recognized by the algorithm [3]. correctly detected words% Recall Rate 100correctly detected words false negatives =⨯+This approach can also be used for individual characters rather than words, if the image doesn't contain complete words. These quantities are calculated to determine the efficiency of the algorithm.Table 1. Performance Comparison of different text extractiontechniquesThe proposed OpenCV implementation achieved a precision rate of 88.3% and a recall rate of 76.8%. This implies that the model is capable of not producing considerable false positives, but may sometimes encounter false negatives, i.e., may not detect regions actually containing text. Table 1 gives the performance comparison of different text extraction techniques. Some of the images that the algorithm produced are shown inFig.5. The bounding boxes around the text are accurately generated, and the algorithm also performed well on noisy images.Fig.5. The original images fed to the model on the left side and theoutput images with bounding text boxes on right side.The accuracy obtained is quite satisfactory, considering the wide variety of images present in the used dataset. However, it has some basic limitations. The first limitation is that the algorithm cannot detect text that is not horizontally aligned as it cannot produce rotated bounding boxes. Another limitation is that it cannot detect text that is embedded in a circular manner. But these limitations can be easily resolved in future.V. C ONCLUSIONThe paper presented an efficient technique of text extraction from natural scene images, i.e., the images directly taken from a camera, having a random background and a diverse variety of colors. There are a number of significant applications of this task, including document retrieving, vehicle number plate detection, road signs detection (by smart cars), page segmentation, assistance to visually impaired persons, video content summary, and intelligent driving assistance, to name a few. There exists a number of techniques for text extraction, each having its own set of strengths and limitations. There is no single algorithm that works for all the applications due to the wide variations in the type of natural images.The proposed technique is efficient in producing text predictions on images having horizontally-aligned text. A single convolutional neural network based on the EASTtext detector [17] is employed for the detection task. For the recognition part, another CNN model is trained which produces excellent results when proper preprocessing of the image is done before feeding it to the network.The future research may possibly include overcoming the limitations of this implementation, i.e. its inability to detect text in images that are not horizontally aligned or images where the text is not in a straight line. Also, this approach can be extended from images to videos, which are nothing but a series of images. Moreover, a smartphone application can be built that can detect and extract text from camera-captured images (or videos) in real-time.R EFERENCES[1] B. Epshtein, E. Ofek, and Y. Wexler, “Detecting text innatural scenes with stroke width transform”, Proc. CVPR, 2010.[2]H. Raj, R. Ghosh, “Devanagari text extraction fromnatural scene images”, IEEE 2014.[3]T. Kumuda and L. Basavaraj, “Text extraction fromnatural scene images using region based methods-a survey”, Proc. of Int. Conf. on Recent Trends in Signal Processing, Image Processing and VLSI, ACEEE 2014. [4]M. Prabaharan and K. Radha, “Text extraction fromnatural scene images and conversion to audio in smartphone applications”, IJIRCCE, 2015.[5] C. Bartz, H. Yang, and C. Meinel, “STN-OCR: A singleneural network for text detection and text recognition”,arXiv:1707.08831v1 [cs.CV] 27 Jul 2017.[6]S. Mori, H. Nis hida, and H. Yamada, “Book opticalcharacter recognition”, John Wiley & Sons, Inc. New York, NY, USA, 1999.[7] F. Liu, X. Peng, T. Wang, and S. Lu, “A density-basedapproach for text extraction in images”, IEEE 2008. [8]J. Kim, S. Park, and S. Kim, “Text locating from naturalscene images using image intensities”, IEEE 2005.[9]M. Busta, L. Neumann, and J. Matas, “Fastext: Efficientunconstrained scene text detector”, in Proc. of ICCV, 2015.[10]Coates et al., “Text detection an d character recognition inscene images with unsupervised feature learning”, Proc.ICDAR 2011, pp. 440–445.[11]S. Lu, T. Chen, S. Tian, J. H. Lim, and C. L. Tan, “Scenetext extraction based on edges and support vectorregression”, IJDAR, 2015.[12]W. Huang, Y. Qiao, and X. Tang, “Robust scene textdetection with convolution neural network induced msertrees”, in Proc. of ECCV, 2014.[13]N. Mishra, C. Patvardhan, Lakshimi, C. Vasantha, and S.Singh, “Shirorekha chopping integrated tesseract ocrengine for enhanced hi ndi language recognition”, International Journal of Computer Applications, Vol. 39,No. 6, February 2012.[14]R. Smith, “An overview of the tesseract OCR engine”, inProc. Int. Conf. Document Anal. Recognition, pp. 629–633 2007.[15]L. Neumann and J. Matas, “Real-time scene textlocalization and recognition”, 25th IEEE Conference on Computer Vision and Pattern Recognition, 2012.[16]X. Liu and J. Samarabandu, “Multiscale edge-based textextraction from complex images”, 2006 IEEEInternational Conference on Multimedia and Expo. [17]X. Zhou, C. Yao, H. Wen, Y. Wang, S. Zhou, W. He, andJ. Liang, “EAST: an efficient and accurate scene textdetector”, arXiv:1704.03155v2 [cs.CV] 10 Jul 2017. [18]Chee Kheng Ch’ng & Chee Seng Chan, “Total-Text: acomprehensive dataset for scene text detection and recognition”, 14th IAPR International Conference on Document Analysis and Recognition {ICDAR}, 2017, pp.935-942.[19]X. C. Yin, X. Yin, K. Huang, and H. W. Hao, “Robusttext detection in natural scene images”, IEEE transactions on pattern analysis and machine intelligence, 0162-8828/13, 2013, IEEE.[20]X. Liu, K. Lu, and W. Wang, “Effectively localize text innatural scene images”, 21st international co nference on pattern recognition(ICPR), November 11-15, 2012, Tsukuba, Japan.[21]Y. F. Pan, X. Hou, C. L. Liu, “A robust system to detectand localize texts in natural scene images”, unpublished. [22]L. Neumann and J. Matas, “A method for text localizationand recognition in real-world images”, in Proc. of ACCV, 2010.[23]S. Tian, Y. Pan, C. Huang, S. Lu, K. Yu, and C. L. Tan,“Text flow: A unified text detection system in natural scene images”, in Proc. of ICCV, 2015.[24]M. Jaderberg, K. Simonyan, A. Vedaldi, and A.Zisserman, “Reading text in the wild with convolutional neural networks”, International Journal of Computer Vision, 116(1):1– 20, jan 2016.[25] A. Gupta, A. Vedaldi, and A. Zisserman, “Synthetic datafor text localisation in natural images”, arXiv preprint arXiv:1604.06646, 2016.[26]Z. Zhang, C. Zhang, W. Shen, C. Yao, W. Liu, and X. Bai,“Multi-oriented text detection with fully convolutional networks”, in Proc. of CVPR, 2015.Authors’ ProfilesVaibhav Goel: Under-graduate studentpursuing Bachelor of Technology (B.Tech)at Bharati Vidyapeeth’s College ofEngineering, New Delhi, India, major inComputer Science.Vaibhav Kumar: Under-graduate studentpursuing Bachelor of Technology (B.Tech)at Bharati Vidyapeeth’s College ofEngineering, New Delhi, India, major inComputer Science.Amandeep Singh Jaggi: Under-graduatestudent pursuing Bachelor of Technology(B.Tech) at Bharati Vidyapeeth’s College ofEngineering, New Delhi, India, major inComputer Science.Vidyapeeth’s College of Engineering, NewDelhi, India, major in Computer Science.How to cite this paper: Vaibhav Goel, Vaibhav Kumar, Amandeep Singh Jaggi, Preeti Nagrath, "Text Extraction from Natural Scene Images using OpenCV and CNN", International Journal of Information Technology and Computer Science(IJITCS), Vol.11, No.9, pp.48-54, 2019. DOI:10.5815/ijitcs.2019.09.06。
《电脑知识与技术》征稿启事期刊简介：《电脑知识与技术》杂志创刊于1994年，是经国家批准的旬刊杂志，主管单位：安徽省科技厅，主办单位：安徽省科技情报学会、中国计算机函授学院。
国内统一刊号：CN34-1205/TP，国际标准刊号：ISSN1009-3044。
《电脑知识与技术》邮发代号：26-188，系中国学术期刊综合评价数据库统计源期刊，中国期刊全文数据库全文收录期刊，中国核心期刊（遴选）数据库收录期刊，中国学术期刊（光盘版）收录期刊。
《电脑知识与技术》是一本面向计算机全行业的综合性学术刊物。
编委委员均来自国内各著名高校的教授和博士生导师，他们是我国计算机科学领域方面的专家、学者和权威人士；稿源来自全国各高等院校，相关专业研究机构以及国内大型信息通讯、软件研发企业设置的专业研究所。
学术交流版刊登的论文均有2名国内同行专家审稿通过。
学术交流版的作者有教授、博士生导师、研究员、博士研究生、和硕士研究生以及从事该研究领域的工作人员、专家和学者。
主要面向广大的科技工作者、高等院校、各公共图书馆、情报所（室）、研究所以及厂矿，它对科技工作者、科学研究人员、广大教师、研究生、博士生都具有重要的参考价值。
栏目设置：数据库与信息管理：数据库理论与系统，各种专用数据库，数据挖掘的实用技术,各类信息系统及其管理等；网络通讯与安全：计算机网络理论及应用，计算机网络安全，计算机数据和设备安全，移动通信技术等；软件设计开发：软件工程，程序设计，基于计算机、电子、自动化某个领域的理论与应用研究等；人工智能及识别技术：人工智能技术，图像识别技术等；多媒体技术及其应用：多媒体技术设计、开发、应用技术等；计算机教学与教育信息化：计算机化教学，计算机教学应用研究，计算机人才的培养等。
信息技术与课程整合：教育信息化、信息技术在教学中的应用、网络远程教育、数字化教学资源建设等；征稿对象：从事计算机、电子、通讯、教育方面学习或工作的人员均可向本刊投稿。
Computer Knowledge and Technology 电脑知识与技术本栏目责任编辑：唐一东多媒体技术及其应用第8卷第22期(2012年8月)一种基于提升小波变换的边缘检测方法何伟（湖北大学数学与计算机科学学院，湖北武汉430062）摘要：为了更好的提取图像的边缘信息，该文利用小波变换的性质，提出了一种基于提升小波变换的边缘检测方法。
通过提升小波变换将图像分解为一个低频系数和三个高频系数。
首先对低频系数提取边缘，再分别对高频系数进行平滑处理，最后对边缘图和高频系数进行提升小波逆变换获得源图像的最大边缘信息。
关键词：提升小波变换；平滑处理；边缘检测中图分类号：TP391文献标识码：A 文章编号：1009-3044(2012)22-5446-02An Image Edge Detection Method based on Lifting Wavelet TransformHE Wei(Collage of Mathematics and Computer Science,Hubei University,Wuhan 430062,China)Abstract:In order to extract the image edge information,this paper uses the properties of wavelet transform,proposed one kind of edge detection method based on the lifting wavelet transform .Through the lifting wavelet transform to image is decomposed into one low fre ⁃quency coefficients and three high frequency coefficients.The low frequency coefficients of the edge extraction,respectively of the high fre ⁃quency coefficients smoothing,the edges of the image and the high frequency coefficients are inverse transform of the lifting wavelet to ob ⁃tain maximum edge information of source images.Key words:lifting wavelet transform;smooth processing;edge detection图像边缘检测是指检测出图像局部变化不连续的区域，然后根据这一细则勾勒出图像的具体轮廓，起到图像识别的作用。
