Talk about the product manager the threshold value of image recognition technology control

Source：Updated：2017-04-05 08:09:34

Products meet the needs of the user has a threshold, the product value is lower than the threshold value users will feel dull, namely product so-so, namely product manager to do the functional manager. Product value is equal to the threshold function basically met the needs of users, and only product manager to manage the demand, the products make it work, the product value can be higher than the threshold, a product manager at any time should learn to demand higher than the threshold value of methodology. AI + era image recognition technology is the starting point! 
 
The purpose of writing this: 
 
Now a day to one word: AI, flood is AI stage, we should how to understand the product manager AI context of technology and market demand trend. AI is not a new concept, up again because a new breakthrough. 
 
Dr Lee said innovation works is now a practitioner of technology innovation era, then our product manager actually know what AI technology, this paper emphatically analyze the AI + image recognition technology of The Times. 
 
In the field of AI, image recognition technology occupies a very important position, and with the continuous development of computer technology and information technology, the image recognition technology of AI to expanding the application scope of, such as LineLian have seen the IBM Watson, medical diagnosis, all kinds of fingerprint identification, and the commonly used alipay facial recognition and baidu map view satellite cloud image recognition belongs to typical of this application, AI this technology has been used in daily life, image recognition technology in the future will have become more widely used, and in order to guarantee the AI in the image recognition technology to better serve the AI + the several key areas of product of The Times, it is this is AI image recognition technology in the concrete discussion. 
 
Profile image recognition: 
 
In order to better complete product managers understanding the origin of the image recognition technology, the AI we first need to understand the image recognition technology. As an important part in the field of intelligent, has the development of the image recognition of the character recognition, digital image processing and recognition, object recognition, three development stages, and in the era of AI + image recognition technology, its itself has the function of the already exceeded the limits of human, it is also the AI of image recognition technology can achieve better in the field of vertical product application and almost become the standard. 
 
Initially product managers to understand the principle of image recognition technology itself is not too complicated, the processing of information is a key point of this technology, because of the application of computer image recognition technology has nothing to do with the difference of human recognition does not exist in nature, this makes the image recognition technology also needs to be done according to their own memories of image identification of a specific job. 
 
In humans in the process of image recognition, the human brain will image features extraction, and combining previous cognitive judgment of all kinds of images in the brain itself is to image existed impression, this is what people can quickly after watching a picture on the recognition of the cause. Combined with the principle of human recognition image, in the computer image recognition, the computer will be able to complete the image classification and first select important information, eliminate redundant information, according to the classification of the computer can be combined with its own memory storage requirements for image recognition, the process itself and the essential difference between the human brain images recognition does not exist. 
 
For image recognition technology, its direct relationship with the image recognition of image features extracted can achieve more satisfactory results. It is important to note that due to the computer at the end of the day is different from the human brain, so the extracted image features exist instability, this instability will often because the computer to extract image features and common to significantly affect the efficiency and the accuracy of image recognition, thus the image characteristics for the significance of image recognition technology in AI. 
 
Image recognition analysis: 
 
For the AI + era of image recognition technology, neural network image recognition technology and nonlinear dimension reduction of image recognition technology is the most common of the two kinds of image recognition technology, LineLian to two common AI image recognition technology for detailed analysis. 
 
An image recognition, neural network technology 
 
Want to have a thorough understanding of the nonlinear dimension reduction image recognition technology, we must understand what is a neural network, the full name of artificial neural network, neural network itself is in the modern neurobiology research based on simulation of biological processes to reflect the features of the human brain structure calculation, although we use the simulation this term, but in fact the neural network itself has not completely imitate human neural network, its itself only through abstraction of human neural network, simplify and simulation implementation related to promote the efficiency of computing structure. 
 
For neural network image recognition technology, it can realize image recognition is mainly thanks to the use of neural network learning algorithm, and in the application of neural network image recognition, we first need to relevant image preprocessing, the pretreatment including true color image is converted into a gray, gray image rotation and zoom, gray image normalization, etc. In order to ensure that the neural network can well realize the image recognition, we also need for image recognition and accomplish specific objects in the field of neural network design, this design mainly includes the following five aspects: 
 
Input layer design 
 
Hidden layer design 
 
Output layer design 
 
The selection of initial weights 
 
The selection of expectation error 
 
In the design of input layer, we need to be determined according to the needs of image object recognition to solve the problem with the data representation, and in this study, to understand our product manager, LineLian input layer and unified design sample size is 16 * 16 image zooming, 256 d input needs; In the design of the hidden layer, we need to determine the number of hidden layer and hidden layer unit number choice, the industry has identified the increase of the number of hidden layer neurons can guarantee accuracy of error is reduced, so the appropriate time to increase the number of hidden layer can be better to complete the design of the neural network, and in the number of hidden layer units, we can refer to M + N + a empirical formula L =), L = log2N, so that it can effectively avoid neural network generalization ability is weak, for training sample recognition rate reduce problem, formula of M represents the number of output layer neurons, and N represents the number of input layer neurons. 
 
It is worth noting that less affected by removing the hidden layer unit can improve the performance of the neural network itself, but the structure of the selected it takes long time to the defect of this method; In the design of output layer, and will generally choose multiple output model as the design of the neural network; In the selection of initial weights, in order to satisfy the neural network has good convergence in the learning process, the initial weights are generally selected as the random number between (1, 1); In the selection of expectation error, its itself needs to refer to the training time and forecast error values, here LineLian choose 0.001 as expected error value. 
 
After finished the design of the neural network, we also need for neural network training to ensure its better meet the demand of image recognition, in order to ensure that the design of better implementation, LineLian selection used in MATLAB7.0 newff function to create a two layers of the network, the network includes 1 output neuron, 16 * 16 input, 26 units of the hidden layer, learning function, chose learngdm, initial learning rate is 0.01 ~ 0.6 "mse", training, training, performance function index of 0.001, training the largest circulation of 2500. 
 
At the completion of the above mentioned design and training of neural network, we can start on the application and experiment in this experiment the author applied the neural network to 26 handwritten English letters recognition, the image of the table below for the recognition of recognition results, combined with the table we can found that different number of nodes will directly affect the recognition rate of neural network image recognition, and 26 number of hidden layer nodes can better meet the needs of image recognition, image below for the hidden layer of 26 when the error of the neural network training performance curve and the training time. 
 
Combined with the results we can conclude that the neural network recognition technology can meet the recognition of handwritten letters, its itself in the process of the identification embodies the accurate, fast, strong anti-interference ability and other characteristics, these characteristics make itself with learning algorithm is applied to the more complex well image recognition, to better provide service for our products in the field of the vertical. 
 
AI + times, talk about the product manager the threshold value of image recognition technology control 
 
AI + times, talk about the product manager the threshold value of image recognition technology control 
 
Second, the nonlinear dimension reduction of image recognition technology 
 
Besides neural network image recognition technology, nonlinear dimension reduction of image recognition technology is the present era of AI is more commonly used form of image recognition technology. For traditional application of computer image recognition technology, its itself belongs to the relatively high dimensional recognition technology, the high dimensional feature makes the computer often in the process of image recognition under a lot of unnecessary burden, the burden will naturally affect the speed of image recognition and quality, nonlinear dimension reduction of image recognition technology is to be able to better achieve the image recognition technology form of dimension reduction. 
 
In nonlinear dimensionality reduction before image recognition technology, the industry the most commonly used is linear dimension reduction of image recognition technology, the technology itself has the advantages of simple easy to understand, but in practice it has been found that the image recognition technology of linear dimension reduction exist in high computational complexity and takes more time and space features, also makes the image recognition technology of linear dimension reduction will not be able to better satisfy the needs of each product image recognition field. For nonlinear dimension reduction of image recognition technology, its itself can under the premise that does not destroy the structure of the image to achieve its own dimension reduction this makes the image recognition technology in the recognition speed and accuracy can be achieved good promotion. 
 
In face recognition system, for example, in the past by the image dimension is higher, the influence of human recognition system often requires a lot of time, the computer system also tend to get larger "destroyed", this is mainly due to the face in the high latitudes space caused by uneven distribution of properties, and the nonlinear dimension reduction in the application of image recognition technology, face graphics can better realize its own compact, this makes the human face recognition system to enhance the work efficiency, overall nonlinear dimension reduction of image recognition technology can be well for the image 
 
Don't provide auxiliary, above mentioned LineLian neural network image recognition technology, also can in nonlinear dimension reduction of image recognition technology with the support of complete their work better. 
 
The products in the field of application of image recognition technology 
 
As the AI technology of intelligent network development, its itself will be in the product data security, AI, AI live + + medical products, AI + social products, and other areas of the vertical have important applications. 
 
Man-machine game before the war, and finally in human's top player lee se-dol 4-1 defeat to Google's AlphaGo artificial intelligence. Its core principle of multi-layer neural network is adopted to image analysis, summing rules at the same time using the deep learning algorithm, finally it is concluded that overcome the human player of chess. 
 
Image identification, 10000 field control do not to open the Internet not only leads to the freedom, is also a hotbed of spam. Best known for a job is called "jian huang", representative of nature is "tang Ma Ru", but in fact "tang Ma Ru" now can meet more demand for image identification and mining. The best example is the recent burst of live "made man", the rise of video and broadcast content makes demand geometric multiples for the identification of the contents. 
 
Live for audit real-time demand is too high, the number of live online at the same time, accidentally bad things online. The traditional solution is done by the human, the number of needed will with the host as a percentage. Are usually hundreds of people sitting in front of the screen continues to flash across the screen to filter, if found that do not conform to the provisions for manual processing. The products of interest to is to use the AI AI + era image recognition technology. 
 
Also in the field of public safety, the application of face recognition products can better improve the security of market society and convenience; In the medical field, electrocardiogram (ecg) and ultrasound recognition will greatly promote the medical enterprise users convenient; And in the field of agriculture, seed identification technology and the application of food quality testing technology will greatly improve the quality of agricultural production, such as my kind of dozens of mu grapes Grape fruit need pruning Picking leaf needs a lot of labor, feeling a lot of links can be through the machine to process the image recognition, size of different size, different sizes of blades, highly different branches, I always feel can distinguish by image contrast, filtering, image recognition that AI robots is rigid demand. 
 
Image recognition technology in our daily life in the refrigerator will greatly improve the application user convenience of life, the application can realize automatic refrigerator food list generation and preservation of the state of the display, the judgment of the best storage temperature, and other functions, which will greatly improve the user's quality of life. The continuous development of science and technology in the future, the AI of the image recognition technology will achieve more rapid progress, and the development also will be better able to accept service brought by the image recognition technology products, eventually improve the quality of life of users. 
 
As an emerging technology with high content of science and technology, image recognition technology of AI has been together with the user's life, but to ensure that it can provide users with better services, for science and technology network working closely related products technical personnel must vigorously promote continuous learning and innovation of AI image recognition technology products, the product manager in the future for us to create a lot of products is closely related and will improve the product efficiency and the rigid demand for the product to users. 
 
Thinking is to seize the opportunities, demand once washed out or be competitors beyond products to win is very difficult, can only follow the pace of the market to make track for the demand of the market and iteration. And AI + era product manager should some thinking mode is work thinking, not just follow the requirements, is more of a temper to filter and lead the tide of demand of true gold fears not the fire.