What areas of the deep neural network are related to the human brain neural network?

matlab中BP神经网络OCR识别？

Just looking at the error rate analysis can't tell anything,Maybe the sample size is too small,也可能是别的原因.Misidentified samples can be taken out,See what caused the error,targeted improvement.

It may also be that the feature engineering is not in place,bad feature selection,does not satisfy scale invariance、旋转不变性、Three elements of affine invariance,说白了就是,大小变了,The angle of rotation has changed,When taking pictures, the position of the station is different, which causes the perspective of the license plate to change.,Then it may not be recognized.

So you can consider finding a better way to describe features,比如HoG（方向梯度直方图）.

HoG,简单说就是,Subtract two adjacent pixel values,the change in color,then a pixel around,上下、There are two pixels left and right,can do two subtractions,得到两个值,Just like in mechanics two forces can be combined,The two values ​​can also be combined,得到方向,和大小（就是梯度）,In this way, there is a feature of a pixel.

But too many features are too computationally intensive,Use statistical methods to reduce features,First divide the image into grids,It's like drawing a Go line on an image,Then count each square separately,方向在0-20what is the sum of the gradients of the pixels within the angle,依次类推,就得到了直方图,如果以20If the degree is a square,那么180degree can be divided into9histogram,也就是9个特征,The number of features in such a square is independent of the number of pixels,but fixed.

然后就是关于HoGother means,For example, in order to eliminate lighting changes,The feature vector can be normalized, etc..

另外还可以对HoG可视化,in each square,The direction of the thread and length instead of direction and gradient features,最后呈现的效果是,There are several squares,Inside each square there seems to be a star symmetrical about the origin,This will help to analyze the effect of the algorithm.

HoGIt is a relatively common feature descriptor.,It is more used in pedestrian detection.除了HoG,还有SIFT、SURFEqual feature descriptor,These are all in computer vision,It belongs to the category of feature detection.

Computer vision mainly includes binarization、滤波器、特征检测、Some basic means such as feature matching,Then there are the image filters、图像分割、图像识别、Image generation and other specific application algorithms.

The resurgence of neural networks due to the reduction in computing costs in recent years,The research focus of computer vision has turned to various improvements and performance optimizations of deep neural networks,像HoG已经是05年的事情了.

About license plate recognition（LPR）,如果环境不复杂,can be approached100%的准确率的,If the environment is more complex,95%The above accuracy should be achievable.总的来说,The basic application is implemented fall to the ground and commercial.

The current methods are basically deep learning,End-to-end in one go,No need to specifically extract features,Traditional pattern recognition methods have beenGG.说的比较细.

If you only care about the results,GithubSome open source projects on license plate recognition can be found on,比如openalpr之类的,Of course, deep learning is also used.,Blast!,就是这么直接.

深度学习ocrIdentity and Traditionocr的区别

AI爱发猫.

Business card recognition software comes fromOCRIdentify development in depth study2016-12-1617:42Business card recognition software refers to the identification of business cards through mobile phone photos.,After importing phone contacts,Using software is、SyncMLStandard sync to the cloud,便可进行WEB/WAPCloud People Management,Easily get rid of the trouble of difficult management of paper business cards.

When the phone address book is synchronized to the network cloud,People management will achieve a qualitative leap.Whether it is a business office worker,or government leaders,All kinds of convenient and inexpensive network communication and convenient management of contacts can be easily realized in the cloud.

In order to improve the input card information on the mobile terminal speed and accuracy,Beijing Zhongan Future Launches Huicard Business Card Recognition Software,To meet the needs of various industries for automatic entry of business card information,只需在APPIntegration of card card identificationSDK,Users can take pictures with their mobile phones,Automatically enter and identify business card information.

尤其是在crm系统中,After the business card recognition software is introduced, the business card recognition development kit is embedded in thecrmAfter the system can automatically extract the fields on the paper business card,Import into address book,Significantly reduce the time spent entering business cards,提升效率.

Beijing Zhongan Future Needle In order to meet the needs of different users,Build your own cloud platform,You can experience business card recognition on the cloud platform,还可以通过APIThe form of the interface is linked to the WeChat public account,全面支持微信H5的挂接,The business card recognition software launched by Beijing Zhongan Future has fully opened up the mobile terminalAndroid和iOS,WEB网页,微信公众号H5,Is all netcom.

Huicard business card identification originates fromOCR识别技术,The future of Beijing Zhong AnOCR技术来源于TH-OCR识别核心,经过20多年的OCRIdentification technology and the accumulation of experience,Take advantage of the recent boom in artificial intelligence technology,通过引入深度学习算法,Now the recognition speed of business card recognition,The recognition rate is in the leading position in the industry.

In the social process, you will encounter a variety of business cards,If you manually enter it, it looks like a headache,also enter manually,烦呀,Now we have Zhongan Future Business Card Recognition Software,Just use your mobile phone to scan the business card or take a photo to identify the fields on the business card, eliminating the need for manual entry.

OCRWhat is a deep learning platform？

Is bank card identification used?OCRIdentification technology!!How can improve the effect of bank card recognition of recognition,Can you do it with deep learning??

What are the career development directions of deep learning？

当前,The development of artificial intelligence has received comprehensive attention and promotion with the help of deep learning technology breakthroughs,各国政府高度重视、The capital boom is still increasing,All walks of life have also reached a consensus that it has become a development hotspot.

This paper aims to analyze the current state of deep learning technology,Research and judge the development trend of deep learning,And put forward development suggestions according to the technical level of our country.一、Current Situation of Deep Learning Technology Deep learning is the key technology in this round of artificial intelligence outbreak.

Breakthrough progress of artificial intelligence technology in the fields of computer vision and natural language processing,Make artificial intelligence usher in a new round of explosive development.And deep learning is the key technology to achieve these breakthroughs.

其中,Image classification technology based on deep convolutional networks has surpassed the accuracy of the human eye,Speech recognition technology based on deep neural network has reached95%的准确率,Machine translation technology based on deep neural network has approached the average human translation level.

The substantial increase in accuracy has brought computer vision and natural language processing into the industrialization stage,带来新产业的兴起.Deep learning is an algorithm tool in the era of big data,become a research hotspot in recent years.compared to traditional machine learning algorithms,Deep learning technology has two advantages.

First, deep learning technology can continuously improve its performance with the increase of data scale,It is difficult for traditional machine learning algorithms to use massive data to continuously improve their performance..

Second, deep learning technology can directly extract features from data,削减了对每一个问题设计特征提取器的工作,Traditional machine learning algorithms require manual feature extraction..

因此,Deep learning has become a hot technology in the era of big data,Of depth study carried out by academia and industry have a lot of research and practice.All kinds of deep learning models fully empower basic applications.Convolution neural network and loop neural network is the depth of the two types of being widely applied neural network model.

Computer vision and natural language processing are the two basic applications of artificial intelligence.Convolutional Neural Networks are widely used in the field of computer vision,在图像分类、目标检测、The performance on tasks such as semantic segmentation greatly surpasses traditional methods.

Recurrent neural network is suitable for solving sequence information related problems,Has been widely used in the field of natural language processing,如语音识别、机器翻译、对话系统等.Deep learning technology is still not perfect,to be further improved.

First, the model complexity of deep neural network is high,Huge number of parameters lead to large model size,Difficult to deploy to mobile end devices.Second, the amount of data required for model training is large,And the training data samples are obtained、标注成本高,Samples of some of the scenes are difficult to obtain.

Third, the application threshold is high,Algorithm modeling and the process too complicated、Algorithm design cycle is long、Difficulty in implementing and maintaining the system.Fourth, lack of causal reasoning ability,图灵奖得主、贝叶斯网络之父JudeaPearlPoint out that current deep learning is only“曲线拟合”.

Fifth, there are interpretability problems,Due to internal parameter sharing and complex feature extraction and combination,It is difficult to explain what exactly the model has learned,But for security reasons as well as the needs of the ethical and legal,The interpretability of the algorithm is also very necessary.因此,Deep learning still needs to solve the above problems.

二、The development trend of deep learning Deep neural network presents deeper and deeper layers,The development trend of more and more complex structure.In order to continuously improve the performance of deep neural networks,The industry continues to explore in terms of network depth and network structure.

The number of layers of the neural network has been extended to hundreds or even thousands of layers,随着网络层数的不断加深,Its learning effect is getting better and better,2015年微软提出的ResNet以152For the first time, the accuracy of the network depth of the layer exceeds that of the human eye on the image classification task.

New network design structures are constantly being proposed,Make the structure of neural network more and more complex.

如：2014年谷歌提出了Inception网络结构、2015In 2008, Microsoft proposed the residual network structure、2016Nian Huang et al. proposed a densely connected network structure,These network structure designs continue to improve the performance of deep neural networks.

Deep neural network node functions continue to enrich.In order to overcome the limitations of current neural networks,The industry explores and proposes a new type of neural network node,Make the function of neural network more and more rich.

2017年,Jeffrey Hinton came up with the concept of capsule networks,Using capsules as network nodes,In theory is closer to the brain,Designed to overcome the limitations of convolutional neural networks without spatial layering and reasoning capabilities.

2018年,DeepMind、谷歌大脑、MITScholars combined puts forward the concept of network diagram,defines a new class of modules,With relational induction bias,Aims to empower deep learning with causal reasoning.Deep neural network engineering application technology continues to deepen.

Most deep neural network models have hundreds of millions of parameters and hundreds of megabytes of space.,运算量大,Difficult to deploy to smartphones、Performance and resource-constrained terminal devices such as cameras and wearables.

为了解决这个问题,The industry adopts model compression technology to reduce the quantity and size of model parameters,减少运算量.The current model compression method includes pruning the trained model（如剪枝、Weight sharing and quantization, etc.）and design more elaborate models（如MobileNet等）两类.

The process of deep learning algorithm modeling and parameter adjustment is cumbersome,应用门槛高.In order to reduce the application threshold of deep learning,Industry proposes automated machine learning（AutoML）技术,Enables automated design of deep neural networks,简化使用流程.

Deep learning and various machine learning technologies continue to integrate and develop.

Deep reinforcement learning technology born from the fusion of deep learning and reinforcement learning,Combining the perception ability of deep learning and the decision-making ability of reinforcement learning,Overcome the defect that reinforcement learning is only suitable for discrete and low-dimensional states,Control policies can be learned directly from high-dimensional raw data.

In order to reduce the amount of data required for deep neural network model training,The industry has introduced the idea of ​​transfer learning,Thus was born the deep transfer learning technology.迁移学习是指利用数据、任务或模型之间的相似性,将在旧领域学习过的模型,应用于新领域的一种学习过程.

By transferring the trained model to similar scenarios,Achieve better results with only a small amount of training data.三、Suggestions for future development to strengthen the graph network、Research on cutting-edge technologies such as deep reinforcement learning and generative adversarial networks.

Due to the lack of major original research results in the field of deep learning in my country,The basic theory research contribution,such as capsule networks、innovations such as graph networks、The concept of originality was proposed by American experts,China's contribution to the study.

in deep reinforcement learning,Most of the latest research results are based onDeepMind和OpenAIresearchers from foreign companies,There are no breakthrough research results in my country.

Generative Adversarial Networks, a Research Hotspot in Recent Years（GAN）by American researchersGoodfellow提出,并且谷歌、facebook、twitterand Apple and other companies have proposed various improvements and application models,有力推动了GAN技术的发展,However, there are few research results in this field in our country..

因此,Research institutes and enterprises should be encouraged to strengthen the combination of deep neural networks and causal inference models、Research on cutting-edge technologies such as generative adversarial networks and deep reinforcement learning,Present more original research results,Enhance global academic research impact.

Accelerate automated machine learning、Research on deep learning application technologies such as model compression.Relying on domestic market advantages and enterprise growth advantages,For personalized application requirements with the characteristic of our country,To speed up the study of deep learning application technology.

Enhancing machine learning for automation、Research on technologies such as model compression,Accelerate the engineering application of deep learning.Strengthen the application research of deep learning in the field of computer vision,Further improve the accuracy of visual tasks such as target recognition,and performance in practical application scenarios.

Strengthen the application research of deep learning in the field of natural language processing,Propose an algorithm model with better performance,Improve machine translation、Performance of applications such as dialog systems.

来源：产业智能官ENDFor more exciting content, please visit the official website▼1.Yinlu Network2018-2019China's top 100 artificial intelligence industry innovation list released！2.Yinlu Network2018-2019China's artificial intelligence industryTop20List of investment institutions released！

3.Yinlu Network2018-20192019 China Big Data Industry Innovation Top 100 List Released！4.Yinlu Network2018-2019China's big data industryTop20List of investment institutions released！

5.Yinlu Network2018-20192019 China Internet of Things Industry Innovation Top 100 List Released！6.Yinlu Network2018-2019年中国5GIndustry with IoTTOP20List of investment institutions released！

7.Yinlu Network2018-2019Announcement of China's Top 100 Integrated Circuit Industry Innovation Top 100 List！8.Yinlu Network2018-2019In China's integrated circuit industryTop20List of investment institutions released！

9.Yinlu Network2018-2019Announcement of the Top 100 List of China's Enterprise Service Industry Innovations！10.Yinlu Network2018-2019China's enterprise service industryTOP20List of investment institutions released！

At home is better, a professor of what deep learning field

.

在过去的三十年,The deep learning movement was once considered an anomaly in academia,但是现在,GeoffHinton（如图1）and his deep learning colleagues,Including the New York university,YannLeCun和蒙特利尔大学的YoshuaBengio,Unprecedented attention in the Internet world.

Hintonis a professor and researcher at the University of Toronto, Canada,目前就职于Google,He uses deep learning techniques to improve speech recognition、User experience for image tags and countless other online tools,LeCun在Facebook做类似的工作.

Artificial intelligence at Microsoft today、IBMas well as Baidu and many other companies are getting a lot of attention.我非常兴奋,We found a method of neural network can be made better,Especially this method can reveal how the brain works——GeoffHinton.

GeoffHintonet al created the renaissance of deep learningHintonStudy Psychology at Cambridge at undergraduate level,He realized that scientists didn't really understand the brain——Inability to fully grasp the interactions between billions of neurons and how to improve intelligence.

These scientists can interpret electrical signals along one axon connecting one neuron to another,But they couldn't explain how these neurons learned or calculated.HintonThink that these are big problem,The answer may let us realize in the end1950The dream of an artificial intelligence researcher.

图1：GeoffHinton（AIleader,目前就职于Google）他也没有答案,But he will try his best to find the answer,At least an improved artificial neural network could mimic some aspects of the human brain.

“我非常兴奋,We found a method of neural network can be made better,Especially this method can reveal how the brain works,”Hinton说,full of youthful enthusiasm.

These artificial neural networks can gather information,and able to respond,They can understand what things look or sound like.

when you combine words,They become smarter when making decisions,while completing these processes does not require a human to provide the object or the label of the object,This is not possible with traditional machine learning tools.

随着人工智能的发展,The neural network will be more quickly、灵活、高效,They get smarter as the size of the machine increases,Will be able to solve more and more complex tasks over time.

早在80年代初,当HintonWhen starting the idea with a colleague,Computers at the time were far from being able to handle the huge datasets needed by neural networks.,success is limited,Then the AI ​​community turned their backs on them,Instead, look for shortcuts to human-like brains,Rather than try to imitate the operation of the brain.

But there are still some researchers who firmly supportHinton的工作.

根据Hinton和LeCun回忆,it's extremely difficult,甚至直到2004年——已经是Hinton和LeCun第一次开发“反向传播”Algorithm neural network20年之后了——Academia has no interest in these.

但是那一年,from the Canadian Institute for Advanced Projects（CIFAR）received very little money,并在LeCun以及Bengio的支持下,Hinton建立了神经计算和自适应感知项目,This project invites only a few computer scientists、生物学家、电气工程师、神经科学家、physicist and psychologist.

by recruiting these researchers,HintonAims to create a world-class team,Working on creating simulations that mimic biological intelligence——Modeling how the brain sifts through vast amounts of vision、auditory as well as written cues to understand and respond to its environment.

Hintonbelieves that establishing such an organization would spur innovation in the field of artificial intelligence,甚至改变世界,事实证明,他是对的.

GeoffreyHintonZeng felt that his academic career was likeANN（人工神经网络）ups and downs,所幸的是,这位GatsbyThe founders have never given upANN的研究.

to realize their early ideas,Gather together regularly for seminars,To build a more powerful deep learning algorithms,Manipulate larger datasets.Win the global artificial intelligence competition during,And then the giants of the internet started noticing them.

2011年,一位NCAPFellow and StanfordAndrewNg在GoogleDeep learning project established,今天,Companies use neural networks toAndroidMobile phones and social networks as wellGoogle+mark up image.

去年,Hinton加入Google公司,The purpose is to further take this work to a more in-depth.

less than a million dollars per yearCIFAR投资,HintonAnd the rewards for his mates are plentiful,这不仅发生在Googlealso occurs in some countries,包括加拿大.

在这个过程中,Hinton和NCAPHas changed the face of the once abandoned their communities,The phenomenon of college students switching from traditional machine learning projects to deep learning is everywhere.毫无疑问,Deep learning is now mainstream.

“We are no longer extremists”Hinton说,“We are now a hot core technology.”HintonAlso travels the world and preaches actively for deep learning,HintonI have a habit of shouting suddenly：“I now understand how the brain works！

”It's infectious,He will do it every week,you are hard to imitate.通过NCAP和CIFAR,HintonOpened a summer school,Committed to training a new generation of artificial intelligence researchers.

There are so many commercial companies entering this field,This is more important than ever.It's not just tech giants joining the field,We also see a large number of deep learning startups includingErsatz,、ExpectLabs以及Declara.

“我们希望把AI和CIFARTake to a wonderful new territory,”Hinton说,“A realm that no one or program has yet reached.

”和GeoffHintonThe great gods who jointly created the deep learning renaissance also includeYoshuaBengio（如图2）和YannLeCun（图3）教授,他们是Hinton坚定的支持者.

YoshuaBengio（如图2）Professor is also one of the gods of machine learning,他的研究工作主要聚焦在高级机器学习方面,致力于用其解决人工智能问题.

He is one of the few professors who is still fully committed to deep learning academia,Many other professors are already involved in industry,加入了Google或Facebook公司.

图2：MontrealUniversity professors andAI研究者YoshuaBengioYannLeCun和YoshuaBengio不同,He is currently employed inFacebook,任FacebookDirector of Artificial Intelligence Research Institute,He is also one of the most well-known scholars in the field of artificial intelligence, especially deep learning.,at the University of TorontoHintonJoin Bell Labs as a postdoctoral fellow,During the development of convolutional neural network（ConvolutionalNeuralNetworks）and has been widely used for handwriting recognition andOCRThe graph transformation network method of.

2003年加入纽约大学,Engage in a wide variety of research with both breadth and depth,涉及机器学习、计算机视觉、Mobile Robots and Computational Neurology.

图3：纽约大学AIresearchers andFacebookThe director of the artificial intelligence researchYannLeCun毋庸置疑的是,Deep learning, and the whole field of artificial intelligence, has become a focus of competition among Internet giants.

Talents in deep learning are extremely scarceMontrealUniversity full professorYoshuaBengio表示：“Deep learning is hot right now,The current dilemma is the lack of experts,A PhD takes about five years to develop,But five years ago, no PhD student started working on deep learning,This means that there are now particularly few experts in the field,To say the precious、极度稀缺.

”It is said that the current top talents in the field of deep learning are no more than50人,AndrewNgThe main reason for the lack of talent in the deep learning field is first and foremost data,for solving problems in certain areas,To get the data will not be easy;followed by computing infrastructure tools,including computer hardware and software;Finally, the training time of engineers in this field is very long.

So tech giants includeGoogle、Facebook、Twitter、Like baidu have by acquiring start-up companies to recruit talent in the field of deep learning.

Google2013年3acquired a company calledDNNresearch的初创公司,The company is affiliated with the University of Toronto's School of Computer Science.,只有三个人——GeoffreyHintonwith his graduate studentsAlexKrizhevsky和IlyaSutskever.

之后,Google今年1月份斥资4$100 million to acquire artificial intelligence startupsDeepMind,DeepMindby artificial intelligence programmer and neuroscientistDemisHassabis等人联合创立,Is the forefront of artificial intelligence enterprise,It combines state-of-the-art techniques in machine learning and systems neuroscience,建立强大的通用学习算法.

另外,GoogleDevelopers also bought Ukrainian facial recognition technologyViewdle.

GoogleUnceasing acquisition of the company in the field of deep learning is the main purpose“抢购”A group of the world's best experts,In a rapidly growing field of artificial intelligence,These experts are all outstanding.

Facebook也在2012年以近6000Thousands of dollars to buy facial recognition of Israel company.

人事方面,Appoint a computer scientistYannLeCun（图3）Director of the Institute of Artificial Intelligence,Leverage deep learning expertise to help create solutions,On a day to betterFacebook上的3.5Recognize faces and objects in billions of photos and videos.

去年8月13日FacebookAnnounced the acquisition of speech recognition and machine translation companyMobileTechnologies,The latter will help us expand from image recognition to speech recognition.

假以时日,FacebookMay develop services that interact more naturally,and relative to any prior art,It will also help with far more problems.Twitter今年7月29Japan acquires deep learning-based computer vision startupMadbits.

Madbits这家公司是由Facebook人工智能实验室主任YannLeCunFounded by two former students,automatic understanding、Visual Intelligence Technology for Organizing and Extracting Media Content Information.

The computer vision techniques have been developed based on depth of learning,正在测试.TwitterCountless pictures appear every day.

收购Madbits可以帮助TwitterIntroducing features like image search,Improve search rankings based on image content,Even analyzing images to better understand what people are tweeting.其他公司.

Yahoo acquires deep learning companyLookFlowand Image Annotation CompanyIQEngine;QualCommAcquired Image Recognition CompanyKooaba;PinterestAcquisition of Object Recognition CompanyVisualGraph;DropboxAcquired Image Annotation CompanyAnchoviLabs;Baidu established a deep learning research institute headed by Robin Li personally,有AndrewNg、Yu Kai and other technical giants joined;至此,In the field of deep learning several Daniel basically all have belongs to.

The misunderstanding of deep learning and the wave of productization Chief scientist of BaiduAndrewNg表示：“目前围绕DeepLearningsome degree of exaggeration,It doesn't just appear between the lines in the media,also exist in some researchers.It's an unhealthy atmosphere.

将DeepLearningDepicted as a simulation of the human brain,This statement is very attractive,but an oversimplified imitation,它距离真正的AIor what people call‘奇点’还相当遥远.

”At present, the technology mainly learns from massive data,理解数据,This is also about todayDeepLearningDriving force for technical research and product development.

and capable of commensurate with humanAIneed to be all-encompassing,For example, human beings have rich emotions,These are the presentDeepLearningresearch not yet covered.

今天,AIThe biggest challenges and shortcomings in the field arePerception,How to make machines better understand human intent;而这正是"深度学习"The category that can emit light and heat.

A technology can quickly become mainstream,One of the main reasons is the ability to quickly launch mature products,深度学习也不例外,So the deep study the transition is a big trend,pursuit of the unrealistic“天网”Or the high-tech of the movie plot is too quick for quick success、不切实际.

目前"深度学习"让Googleproduct in voice,Smarter recognition of text and images,More accurate insight into our input,A more human understanding of our intentions.

现在,Voice recognition for every Android phone andGoogleImage processing in Street View has"深度学习"的影子.笔者认为,As deep learning grows and tech companies invest more,More and more products will be introduced to the market.

基于pythonIs the graduation design of printed Chinese character recognition system deep learning?

什么是OCR技术?(专业术语解释)

要谈OCR的发展,早在60、70年代,世界各国就开始有OCR的研究,而研究的初期,多以文字的识别方法研究为主,且识别的文字仅为0至9的数字.

以同样拥有方块文字的日本为例,1960年左右开始研究OCR的基本识别理论,初期以数字为对象,直至1965至1970年之间开始有一些简单的产品,如印刷文字的邮政编码识别系统,识别邮件上的邮政编码,帮助邮局作区域分信的作业;也因此至今邮政编码一直是各国所倡导的地址书写方式.

OCRIt can be said that it is an uncertain technical research,The accuracy rate is like an infinite approaching function,know its approach value,but can only get close,always with100%tug of war.

Because there are too many factors involved,Writing habits or file printing quality、Scan quality of the scanner、识别的方法、Study and test samples……等等,How much will affect its accuracy,也因此,OCRof products in addition to having a strong core of identity,Ease of use of the product、Debugging functions and methods provided,It is also an important factor in determining the quality of the product.

一个OCR识别系统,其目的很简单,只是要把影像作一个转换,使影像内的图形继续保存、有表格则表格内资料及影像内的文字,一律变成计算机文字,使能达到影像资料的储存量减少、识别出的文字可再使用及分析,当然也可节省因键盘输入的人力与时间.

从影像到结果输出,须经过影像输入、影像前处理、文字特征抽取、比对识别、最后经人工校正将认错的文字更正,将结果输出.

Introduce one by one here：image input：欲经过OCR处理的标的物须透过光学仪器,如影像扫描仪、传真机或任何摄影器材,将影像转入计算机.

科技的进步,扫描仪等的输入装置已制作的愈来愈精致,轻薄短小、品质也高,对OCR有相当大的帮助,扫描仪的分辨率使影像更清晰、扫除速度更增进OCR处理的效率.

Before the image processing：Image pretreatment isOCR系统中,须解决问题最多的一个模块,from a binary image that is either black or white,or grayscale、color image,to the process of independently creating text images one by one,are image preprocessing.

Contains the image normalization、去除噪声、Image processing such as image correction,And graphic analysis、File preprocessing with text line and word separation.

在影像处理方面,Both academically and technically have reached a mature stage,Therefore, there are many link libraries available on the market or on the website;In the file before treatment,It's up to everyone's skills;影像须先将图片、表格及文字区域分离出来,甚至可将文章的编排方向、The title of the article and the main body of the content are separated,而文字的大小及文字的字体亦可如原始文件一样的判断出来.

文字特征抽取：单以识别率而言,特征抽取可说是OCR的核心,用什么特征、怎么抽取,直接影响识别的好坏,也所以在OCR研究初期,特征抽取的研究报告特别的多.

而特征可说是识别的筹码,简易的区分可分为两类：a statistical feature,如文字区域内的黑/白点数比,当文字区分成好几个区域时,这一个个区域黑/白点数比之联合,就成了空间的一个数值向量,在比对时,基本的数学理论就足以应付了.

The other type of feature is the feature of the structure,如文字影像细线化后,取得字的笔划端点、交叉点之数量及位置,或以笔划段为特征,配合特殊的比对方法,进行比对,市面上的线上手写输入软件的识别方法多以此种结构的方法为主.

对比数据库：当输入文字算完特征后,不管是用统计或结构的特征,都须有一比对数据库或特征数据库来进行比对,数据库的内容应包含所有欲识别的字集文字,根据与输入文字一样的特征抽取方法所得的特征群组.

对比识别：这是可充分发挥数学运算理论的一个模块,根据不同的特征特性,选用不同的数学距离函数,较有名的比对方法有,欧式空间的比对方法、松弛比对法（Relaxation）、动态程序比对法（DynamicProgramming,DP）,And the database establishment and comparison of neural network-like、HMM（HiddenMarkovModel）…等著名的方法,为了使识别的结果更稳定,也有所谓的专家系统（ExpertsSystem）被提出,Exploiting the Distinct Complementarity of Various Feature Alignment Methods,使识别出的结果,其信心度特别的高.

字词后处理：由于OCR的识别率并无法达到百分之百,或想加强比对的正确性及信心值,一些除错或甚至帮忙更正的功能,也成为OCR系统中必要的一个模块.

字词后处理就是一例,利用比对后的识别文字与其可能的相似候选字群中,根据前后的识别文字找出最合乎逻辑的词,做更正的功能.字词数据库：为字词后处理所建立的词库.

人工校正：OCR最后的关卡,在此之前,使用者可能只是拿支鼠标,跟着软件设计的节奏操作或仅是观看,而在此有可能须特别花使用者的精神及时间,去更正甚至找寻可能是OCR出错的地方.

一个好的OCR软件,除了有一个稳定的影像处理及识别核心,以降低错误率外,人工校正的操作流程及其功能,亦影响OCR的处理效率,因此,文字影像与识别文字的对照,及其屏幕信息摆放的位置、还有每一识别文字的候选字功能、拒认字的功能、及字词后处理后特意标示出可能有问题的字词,都是为使用者设计尽量少使用键盘的一种功能,当然,不是说系统没显示出的文字就一定正确,就像完全由键盘输入的工作人员也会有出错的时候,这时要重新校正一次或能允许些许的错,就完全看使用单位的需求了.

结果输出：In fact, the output is a simple matter,But it depends on the userOCR到底为了什么？

有人只要文本文件作部份文字的再使用之用,所以只要一般的文字文件、有人要漂漂亮亮的和输入文件一模一样,所以有原文重现的功能、有人注重表格内的文字,所以要和Excel等软件结合.

无论怎么变化,都只是输出档案格式的变化而已.

基于深度卷积神经网络进行人脸识别的原理是什么？

本质上是模式识别,把现实的东西抽象成计算机能够理解的数字.如果一个图片是256色的,那么图像的每一个像素点,都是0到255中间的一个值,这样你可以把一个图像转换成一个矩阵.如何去识别这个矩阵中的模式？

用一个相对来讲很小的矩阵在这个大的矩阵中从左到右,从上到下扫一遍,每一个小矩阵区块内,你可以统计0到255每种颜色出现的次数,以此来表达这一个区块的特征.

这样通过这一次“扫描”,你得到了另一个由很多小矩阵区块特征组成的矩阵.这一个矩阵比原始的矩阵要小吧？那就对了！

然后对这个小一点的矩阵,再进行一次上面的步骤,进行一次特征“浓缩”,用另一个意思来讲,就是把它抽象化.最后经过很多次的抽象化,你会将原始的矩阵变成一个1维乘1维的矩阵,这就是一个数字.

而不同的图片,比如一个猫,或者一个狗,一个熊,它们最后得到的这个数字会不同.

于是你把一个猫,一个狗,一个熊都抽象成了一个数字,比如0.34,0.75,0.23,这就达到让计算机来直接辨别的目的了.

人脸,表情,年龄,这些原理都是类似的,只是初始的样本数量会很大,最终都是通过矩阵将具体的图像抽象成了数字,因为计算机只认识数字.但是抽象的函数,会有所不同,达到的效果也会不同.

OCRSeveral deep learning methods for recognition