Welcome to Tencent Cloud + community, get more Tencent mass technology practice dry goods oh ~
This post was published by Columneditor on cloud + Community
On July 28, Tencent Cloud held a cloud + community salon in Beijing, inviting five AI technology experts from Tencent and Sichuan Cloud Check Technology to share their AI development experience in professional fields and help developers practice AI technology in specific industry scenarios. Nearly 400 developers participated in the session. During the q&A and tea break, many developers and lecturers discussed the content of the speech, showing their keen interest in AI development.
AI technology is already a household word. Both mobile terminal devices and enterprise system platforms are beginning to integrate AI capabilities. At present, AI has great potential to be integrated into various industries and can play a role in many scenarios, such as cloud computing. In today’s wave of digital transformation, cloud in enterprises has become the new normal. A large amount of data and rich applications on cloud can solve many problems through AI technology, so the integration of cloud and AI is also the new normal.
Due to Tencent’s strong social and game genes, Tencent AI has very rich practice scenarios. AI technology has been used in popular applications in moments of friends, OCR recognition, medical treatment, games and other scenarios, and a large number of new functions and capabilities have been developed. Even so, some of AI’s attempts in these areas are still in their infancy. How to provide matching capabilities in specific business scenarios, using AI to simplify things and unlock productivity, remains very challenging.
Therefore, on July 28, Tencent Cloud held a Cloud + community salon in Beijing, inviting five AI technology experts from Tencent and Sichuan Cloud Inspection Technology to share their AI development experience in professional fields and help developers practice AI technology in specific industry scenarios. Nearly 400 developers participated in the session. During the q&A and tea break, many developers and lecturers discussed the content of the speech, showing their keen interest in AI development. Finally, all the developers took photos with the lecturers, which ended the fun salon full of cutting-edge knowledge in the middle of summer.
The computer vision technology and application behind the explosion in moments of friends
At the present stage, emphasizing the landing of AI in the scene is to hope that AI will enter thousands of households and integrate into the whole society, rather than just a lofty model. This is the trend of AI technology industrialization. For example, interesting interactive activities such as photos of military uniforms and young warriors in the circle of friends are the fastest landing applications that computer vision can reach people’s side. For developers, through the understanding of the specific case process, they can quickly master the ability needed for development, so as to achieve independent innovation development.
Ye Cong, an AI technology expert from Tencent Big Data and Artificial Intelligence Center, helped developers sort out the current application of computer vision and technology by taking some popular activities in wechat moments as a starting point in the salon. Ye Cong said that Tencent cloud is now very focused on AI scenarios, in many applications are integrated with AI capabilities.
The May 4th Youth Day activity “Review May 4th, Which Youth are you most like?” Tencent did it first. Through face detection and analysis technology and face retrieval technology, the photos uploaded by users are retrieved and compared with specific images on the face level, and the photo with the most similar appearance characteristics to the user in the database is found through matching analysis. This idea provides a reference for the design of subsequent AI entertainment products.
From an engineering perspective, “automating the ability of machines to imitate human vision” is much clearer than the academic definition of computer vision — how to enable computers to extract high-level, abstract information from images and videos. According to Ye, computer vision allows machines to partially replace human beings in understanding information in pictures. Computer vision also includes some branches, mainly including object recognition, object detection, semantic segmentation, motion and tracking, 3D reconstruction, visual question and answer, action recognition and so on, while new branches are constantly emerging.
Now the more popular visual applications include face recognition, unmanned driving, semantic segmentation and so on. Semantic segmentation is common in THE field of AI. It generally refers to the segmentation of parts of speech or words. The semantic segmentation in image segmentation generally refers to the different types of objects in the image are marked and distinguished.
In the field of machine recognition, there are many mature feature extraction methods. For example, the edge detection method, the local symmetry of the image object, the scale-invariant feature, the principle of gray scale and so on, these algorithms can achieve extraction, but there is no optimal scheme. There are also some other algorithms in the field of image segmentation and object detection, such as the watershed algorithm, which, as its name implies, uses the gray characteristics of the image to display the entire gray curve of the image. In addition, the commonly used algorithm of object detection is called subjective shape model. For example, taking human face as the standard, there are 68 points on human face. Transform these points to achieve the matching target.
The development of image recognition has gone through several stages such as CNN, R-CNN, Faster R-CNN and so on. Now the image recognition trend will tend to the nature of the development of logic, is a big cycle. So, based on these scientific experiments, how does Tencent cloud support AI applications? Take the May 4th youth activity as an example, it is an activity to solve image matching. First of all, start from the training data, usually old photos of the Republic of China, and extract and label the data. Since every photo has labels, the model will be generated. When users upload photo test data while playing the game, feature extraction and modeling will be carried out, and the model will return a classification. This score is not the execution degree and can not be completely referenced. Finally, a maximum score is returned to the front end to generate a page, and that’s the whole process.
Take face fusion applications, such as military photos. Its process will firstly locate the key points of the face part of the image, extract the features of the face, and then rotate the image to unify it with the template picture. The next step is to extract the face part of the uploaded image based on its eigenvalue, which will be merged with the template image. After fusion, if the light angles of the uploaded photos are not exactly the same, then the pictures are optimized and the light and shadow and curve are adjusted gently, so that a very good effect is presented.
For developers, once you’ve developed an interesting app, how do you commercialize it? Can you just put it on the web and let people download it? In fact, it is far from that simple. It is now common practice to adopt cloud services and try not to deploy on physical sets. Why deploy AI applications on the cloud? Because these apps tend to have very high peaks for a short period of time and then fall off very quickly. The cost of responding to these needs through your OWN IT infrastructure would be prohibitively large, but with the public cloud, you can free up the cost of some of the machines that are now a common business process from application development to revenue realization.
At present, Tencent cloud AI can support more areas. Including a variety of face synthesis, ID identification, intelligent monitoring, face mill and voice synthesis, keyword search and other aspects. Meanwhile, Tencent Cloud’s machine learning platform can help developers quickly implement models.
Finally, how to get THE AI to work in a specific scenario, beyond the technical level, how can developers polish the product to make it easy for users to use complex technologies? Generally speaking, it is a very long process from the idea to the landing, it has many links. First, you have to have experts in AI algorithms, and then you have to have people with practical AI engineering experience, and finally, you have to have more product developers to polish it into a product.
The application of OCR and the technology behind it
OCR is a hot field in recent years. Like id card recognition, license plate recognition, etc., all need to use OCR technology, and there are many scenes like ID card recognition. Therefore, AS a general basic technology, OCR has a very wide range of uses and commercial value. In many tasks requiring human resources to complete text recognition, such as waybill recognition in the express delivery industry and data recognition in the insurance industry, OCR technology has achieved a greater degree of productivity liberation. For developers, with the help of the OPEN API of OCR technology, application development in various life scenarios can be completed.
According to Ji Yongnan, a senior technical expert at Tencent AI, OCR can be traced back to the 1960s and 1970s, when postal codes were the earliest form of OCR service. Current OCR applications can be divided into two dimensions, one is tabular OCR and general OCR; The second is printed and handwritten OCR. At the present stage, tabular form is relatively easy, generic form is relatively difficult; Typography is relatively simple, handwriting is relatively difficult.
Tencent Cloud can now provide print OCR services in multiple scenarios. In the general OCR scenario, drivers license, license plates, bank cards, business cards, and so on can be recognized in addition to certificates. In addition, OCR services require accuracy and completeness, that is, the ability to recognize Chinese and English characters. Tencent cloud is spreading its capabilities beyond common language and characters, such as Chinese and English, to expand its range of recognition.
Currently, OCR is widely used and highly demanding in digitally sensitive industries such as banking. Tencent is not only the first to apply handwritten characters in real scenes, but also has a recognition rate of more than 90% for digits, less than 15 milliseconds for single characters, and more than 80% for complex Chinese characters. Tencent cloud OCR service in the authoritative evaluation performance is also very outstanding. The International Conference on Document Analysis and Recognition (ICDAR), the most authoritative International academic Conference in the field of Document Analysis and Recognition, is held in the International Pattern Recognition Association (IAPR). In the International Document Analysis and Recognition Conference, Tencent OCR recognition won the first place in ICDAR2015 “Focused Scene Text” Scene Text recognition task and ICDAR2015 “Robust Reading Competition” natural Scene Text detection project.
Tencent cloud now provides OCR service interfaces based on various scenarios, and developers can use these services for free to build their own applications. For example, developers can directly call corresponding services on Tencent Cloud to complete application development when they need to use handwriting recognition or do a general OCR recognition for a specific scenario software they actually need to develop.
Tencent Cloud OCR has many application scenarios, and currently has clear target customers in express waybill identification and insurance policy identification. Projects like this are usually customized services that are designed to solve a specific problem. Tencent Cloud develops a system according to the problems in specific scenarios and specific production processes, or develops a process to match the actual business to improve production efficiency.
Take the identification of express waybills as an example. The handwritten waybills must be stored in the warehouse before delivery, and the efficiency of manual identification and input is very low. After using the OCR system of Tencent Cloud, the daily processing capacity can reach 10 million orders, equivalent to the work efficiency of more than 3,000 people in three shifts. Another case is Taikang insurance. In the usual process of purchasing health insurance, the purchaser’s medical history will be reviewed, including the purchaser’s previous physical examination data and the experience data of the designated hospital, and the purchaser’s physical condition will be determined through the analysis of these data. The current solution is to use OCR of Tencent Cloud, and at the same time, to design medical knowledge base together with taikang medical experts and add it into the customized system.
In addition to the above two scenarios, Tencent Cloud has clients such as Bank of Jiangsu and Xiaomi in different OCR applications. The advertising supervision Bureau, including the State Administration of Industry and Commerce, is also using Tencent cloud OCR services.
Smart construction site: application practice of performance attendance system
Digital technology has also been fully applied in the construction of smart cities and smart construction sites. In order to respond to the call of highway quality engineering construction of Ministry of Transport, the application of four new technologies, including new materials, new equipment, new technology and new tools, should be strengthened. In the relatively flexible and unexpected working environment of engineering projects, the performance attendance system often bears a large load. Now, the intelligent performance attendance management can be realized by using AI technology. The average developer can also use AI technology to develop unique products according to the needs of the average enterprise.
Wu Chen, r&d director of Sichuan Yunjian Technology Development Co., LTD., said that the application practice of intelligent site performance attendance system is mainly divided into six parts, including frontier, product analysis, system architecture, main technology, functional analysis and application prospect. The role of the current performance management is to escort the project construction and quality. In practical application, the object of performance attendance includes some important personnel such as project manager, chief engineer, head of safety production, resident design representative of design unit, test and testing engineer, etc. It is aimed at important personnel rather than other general employees.
At present, there are about 7 kinds of mainstream attendance schemes, including fingerprint attendance, ID card attendance, paper card clock attendance, face recognition attendance, iris recognition, finger vein recognition, and camera attendance machine. In practice, these schemes are not ideal in terms of both efficiency and avoiding impostor. Because of the supervision and supervised relationship between the project construction and the participating party, the management department has higher requirements for the authenticity of the information of the implementers.
Wu explained that the added cloud functions are very important by using Tencent Cloud’s AI technology. For example, there is a common situation of personnel change in the construction site. The owner or the project department had to change the personnel before, but now with AI technology, the changed information can be automatically identified. At present, the intelligentization of attendance is mainly considered from six aspects. The second is based on AI; The third is based on cloud computing; The fourth is to connect big data; The fifth is the need to be mobile applications; The sixth is to support smart devices.
At the same time, it also needs to have five characteristics: one is to have the personnel face basic information collection and input function; Second, it can automatically identify, collect and compare face information based on personnel; Three is to support personnel information and personnel ID card information verification; Fourth, support the setting of attendance location and abnormal attendance location judgment; Fifth, it supports early warning of abnormal attendance information of key posts and absence information of key posts, and supports online management of information change of performance personnel. With these capabilities, the result of attendance can be verified in the cloud to achieve intelligent attendance management.
Tencent cloud smart attendance mainly includes four main technical solutions. The first is face comparison, Tencent cloud face comparison is based on facial features to calculate the similarity of two faces, automatic identification; The second is the function of identity verification. Tencent cloud’s identity verification is also a human face. It verifies the user’s identity through a user’s selfie video or a selfie and another user’s pre-retained photo, which is mainly used to help improve business efficiency and reduce labor costs. The third is in vivo detection, Tencent cloud in vivo detection is through the face feature point positioning tracking recognition 3D face reconstruction model, judge whether it is a real person, it supports multiple platforms, including CPU, GPU computing mode, flexible deployment; The fourth is based on LBS positioning service, through the radio communication network of telecom mobile operators, GSM, CDMA network to obtain the location information of mobile terminal users, including geographic coordinates or geodetic coordinates.
Wu explained that the cloud inspection intelligent performance attendance system is currently capable of dynamic analysis of attendance data, including real-time warning of data with abnormal job vacancies. There are four main functions: one is face recognition attendance machine, the main role is front-end data collection, including personnel information registration, face attendance and data upload; The second is the use of mobile APP, mainly used for face attendance, identity verification; Third, the performance attendance management cloud platform, mainly for attendance management, approval management, statistical analysis, system setting, etc. Fourth, API is used to provide the interface for mobile APP to call applications, mainly Tencent cloud technology services and artificial intelligence in the face comparison, such applications of human identification.
In addition to the intelligent construction field, face recognition solutions are widely used in many scenarios. With the development of technology, market expansion, face recognition technology in real life to play a more and more important role.
Application of AI technology in games
The goal of understanding game AI varies from identity to identity. For game developers, using AI is about improving the user experience and increasing player activity. Almost every game these days uses AI. Shooters, for example, have maps, and having a good AI to help players explore the map saves a lot of time, so AI is very important for small game developers.
Wang Liang, a senior researcher at Tencent, said there are three commonly used methods for game AI technology in the industry. One is the behavior tree; The second is based on the search method; The third is the method based on learning. The biggest impact in the gaming industry this year was Dota2, which had AI abilities that exceeded 90% of players with 5 specific heroes and other limitations. It used reinforcement learning, which many other major games are trying to solve.
Taking the popular Honor of Kings as an example, what are the practices and problems of AI in MOBA games? Because King of Glory is a real-time battle game, mainly competitive battle, the complexity of the hero role will be very many, but also bring a lot of complicated AI problems. The first is that the operation sequence and state space of MOBA games are very large. Second, games contain a lot of knowledge, how to express it; The third is the high complexity of MOBA decision problems.
The solutions to these problems are threefold. One is to introduce the framework for stratification, task stratification and scene segmentation; Secondly, the multi-modal feature expression is introduced. The third is the combination of multi-depth learning model.
Knowing the basics, there are still many pitfalls to be encountered in game AI development. What kind of environment is needed during the game development phase? The first problem that requires an environment is the emulator. Currently, tuning is mainly based on this environment. AI access can be based on server architecture, the game engine and algorithm engine are separated, and they are processed through communication. Its advantage is that the game engine is coupled with the model, and supports online learning, so it can be continuously enhanced and updated.
In terms of game AI as a whole, it’s reinforcement learning. The most significant change is that research, which used to be based on rules, is now primarily based on learning. If it’s based on deep learning, at least provide the environment to the developer and how to update the iteration. In addition, game AI is still difficult, but also full of opportunities.
AI technology used in breast cancer recognition
With the progress of The Times and economic development, people’s health awareness is getting higher and higher. At the same time, with the development of AI technology, AI medical treatment has naturally become a hot air in the Internet industry.
According to Jiang Cheng, a senior researcher at Tencent Miying, Tencent recently officially launched its AI breast cancer diagnosis system. There are two considerations in choosing the field of breast cancer. First, for women, breast cancer has the highest incidence of all tumors, with an incidence of about 16%-17%, which seriously harms women’s health. Second, although the incidence of breast cancer is very high, but the cure rate is better. The five-year survival rate is currently 89 percent in the United States and 83 percent in China if the chances of a cure are very high if detected early. This is mainly due to the large population base in China and the lack of experienced film viewing doctors. Currently, the application of AI technology can effectively alleviate this contradiction and help patients and doctors to the maximum extent.
At present, the diagnosis of breast cancer mainly relies on ultrasound, molybdenum target, nuclear magnetic resonance, pathology and genetics, etc. Tencent Cloud aims to organically combine these data modes to form a complete system, so as to improve the diagnosis and treatment technology of breast cancer. At present, the most mainstream and effective screening and diagnosis method is molybdenum target. Tencent’s AI molybdenum target breast cancer diagnosis system has been released and has been put on trial in more than 30 grade A hospitals.
Tencent cloud in the breast molybdenum target mainly achieved three aspects of the function. The first is to achieve the location of suspected lesions; The second is to provide breast benign and malignant judgment; The third is the ability to automatically generate image reports.
How are these three functions realized? It is based on a three-dimensional technical framework. The front end of the structure is the pre-processing layer of molybdenum target image. The middle layer is the AI learning model; The last part is the dynamic update of the first two parts through doctor feedback.
In the middle layer AI learning model, Tencent Cloud independently designed scheme has four outstanding advantages. The first is that the traditional network input is usually a single graph input, but the current scheme can achieve the left and right breast comparison of four graphs input at the same time; The second is the use of multi-scale network, so that the image input network without scaling; The third is the progressive network construction, which is similar to the brain learning process. It decomposes difficult problems into several relatively simple problems and then solves them one by one. In the construction of the network, it is from part to whole, from single image to multiple images. The fourth is the training mode of self-step learning, which is similar to the brain learning from easy to difficult. First, the training samples are classified according to the degree of difficulty. In the training process, the samples are gradually added from easy to difficult, and the model is trained for several rounds, which can achieve the best effect of the model.
In addition to the basic model, according to the feedback of doctors, the model will be transferred and learned for the data newly connected to the hospital to achieve dynamic update. In the process of model training, an important discovery is that the number and type of difficult cases seen by AI largely determines the upper limit of THE AI system. Therefore, valuable difficult cases are regularly mined from the database and online data and labeled. At the same time, some of them will be discussed with the experts of the third grade a hospital, using pathological or other data for cross-confirmation.
Now, Tencent molybdenum target AI model has reached very high accuracy. Among them, tumor detection can reach 90.2%@0.2FP; Calcification detection accuracy is higher, can reach 99%@ 0.2fp; For benign and malignant classification, 87% sensitivity and 96% specificity were achieved. In addition to molybdenum target, the corresponding pathological aspects are also being studied. At present, the pathological study of breast cancer mainly deals with two aspects. The first is functional histological grading, which defines the degree of malignancy of malignant tumor, including mitotic count, nuclear pleiotropy score, and the degree of glandular duct formation. The second is immunohistochemistry, using different staining tablets for molecular typing. For the completed mitotic part, in academic TUPAC professional competition, the previous champion F1 score was 0.73, and now Tencent’s molybdenum target AI system can reach 0.82, which is an obvious improvement.
This increase in accuracy is due to three aspects of technology. The first point is to use the computer’s hard case mining method. After several iterations, the samples are sorted out in each round. Experts confirm difficult annotations and then put them into the samples for further study. The second point is the normalization of the image, which can be improved by using adversarial network to normalize the image. The third point is to improve the speed. By abandoning the segmentation of a whole image in original computer vision and adopting the effect of shared computing and model compression, a pathological film can achieve the processing speed of 0.5 seconds, which is basically close to real-time.
In NMR, Tencent has also carried out corresponding research and developed a semi-automatic and efficient lesion labeling tool. In addition, Tencent is also investing manpower in the direction of ultrasound to expand the research boundary and has completed the data preparation work. In the later stage, we will continue to carry out in-depth research on various modal data and devote ourselves to organically combining multi-modal data for the benefit of patients and doctors.
After Jiang cheng’s speech, it was already evening, and the site developers’ enthusiasm to ask questions and communicate was still undiminished. During the salon, the speeches of five experts on site all focused on the application cases of AI in specific scenarios, which not only brought enlightenment to developers in terms of ideas, but also helped developers realize convenient development and spread AI technology to more application scenarios with Tencent cloud’s open AI capabilities.
Question and answer
Language requirements for AI development?
reading
How do I protect PostgreSQL from attacks
How do I use Alertmanager and Blackbox exporters to monitor Web servers on Ubuntu 16.04
MariaDB Galera Cluster tutorial
Has been authorized by the author tencent cloud + community release, the original link: https://cloud.tencent.com/developer/article/1173931?fromSource=waitui
Welcome to Tencent Cloud + community or follow the wechat public account (QcloudCommunity), the first time to get more mass technology practice dry goods oh ~
Massive technical practice experience, all in the cloud plus community!