first, where is the value of face recognition technology. We regard human face as a biological feature and a commercial application, which is only one of the alternative schemes. In biology, the only criterion for judging is actually the iris, which is the most accurate from the perspective of accuracy and irreplaceable. However, the cost of iris recognition and collection is very high, and the efficiency of recognition is relatively low, which requires waiting time. Therefore, these two conditions restrict the whole industrial application to a relatively small number, and the investment in military industry, national defense, etc. with high identification requirements is very high, which is not suitable for large-scale promotion.
second, fingerprints. We know that the uniqueness of fingerprints is relatively strong, the cost of simultaneous fingerprint collection is relatively low, and the cost of comparison is not high. But why didn't fingerprints become a particularly large alternative scheme for paying and brushing your face? In fact, the main reason is that fingerprints can be copied, which is a comparison between static images. Now we can see Taobao, and a large number of fingerprint stickers, fingerprint films and reproducible features are not suitable for payment. So the fingerprints are now roughly pass.
The third and fourth technologies are face recognition and voice recognition respectively. Compared with the current horizontal, the two are more cost-effective in terms of collection cost, comparison efficiency and uniqueness of life characteristics. So at this stage, face recognition has surfaced for a reason, and this is its value. Where is the application scenario of business characteristics?
The application scenarios of face recognition are very broad. Now there are two main scenarios, one is the financial industry and the other is the security industry. The financial industry has seen the scene from Ma Yun's ant financial service demonstration. It is obvious that you can pay by brushing your face. Why can't you sign for the courier? Next, Taobao should get through the function of signing for the courier. I believe that one day, we will receive a courier from the drone. The drone will take a photo in front of you and compare it, so that you will know that this user is the one you need and complete the whole payment process. In fact, this kind of scene has been discussed and certified in many aspects. Based on this scenario, it is related to the payment authentication of third parties, including the bank of Tencent that we saw. The first remote card opening is to authenticate the person and the card through the technology of face recognition, so that the function of remote account opening and remote card opening should be widely used in our securities firms and our online banks.
For the security industry, at this stage, the application of face recognition should be said to have reached a level that can be commercialized. Let's take an example. Last year, there was a director named Xu Anhua in Hong Kong who lost his purse in the Nanjing subway. It took only five hours to solve this case. A screenshot of a photo was obtained in the video surveillance, and the photo of the suspect was intercepted, which was extremely vague. If the photo of the side face was compared with the naked eye, nothing could be found. However, there is a non-listed company that can't provide its company name here. They use an image restoration technology to restore the photo to the appearance of a possible suspect, take a clear photo, compare it with the photo in the gallery, lock the identity of the suspect, and capture the suspect. It only takes five hours. Now in the field of security monitoring, we can see that all provinces and cities and prefecture-level cities are on a large number of video surveillance and face recognition platforms. In the whole security investment, the security of the previous generation only recorded data statically, but the security of the next generation is a core technology, which is the collection and identification of real-time data. Face recognition plays a great role in this technology.
Let's expand it, whether there is a potential application scenario of the second generation face recognition technology in the future commercial use. We said that in the future, it should be said that the whole identity card should be confirmed, but the comparison between the card and the person needs to be done manually. If we directly define and reach such a standard, in fact, the only ID corresponding to everyone is the biometric characteristics of the face. After this recognition, all places can use the way of brushing their faces, all places can use brushing their faces to open doors and do all kinds of things. The data of your face brushing, including where you go by train, by plane, where to eat, shopping, receiving express delivery, etc., will be mastered in face recognition, and the data of face brushing will replace the current online clicks.
Now the data of credit card and bank card consumption is actually helpful to know the user's consumption habits and data, and do big data marketing and credit investigation, but after the face brushing era comes, this value is even greater. There are many cards, but there is only one face, which is unique. Brush face data is what we focus on in the 2. era.
Why did the technology of face recognition explode in a wide range during this period, and how mature was it when it was widely applied? We must first define that it is a two-stage process for face recognition technology to achieve product application. In the first stage, it is necessary to obtain a large number of sample data, which are used for training. The training is a learning algorithm, and this is a deep learning algorithm. The relationship between these data and each other is extracted and a special comparison is made. The coupling degree is high. After exceeding a certain level, we will assume that these two people are one person, but this model requires a lot of cost, including the cost of optimization, including the cost of data training, including the cost of operation. At that time, an industry company of face recognition, the founder of this company, once said, what does the technology of face recognition mean? Taishang Laojun's blast furnace, with this furnace, big data is the raw material for furnace smelting, which solves the scarcity of computing power resources. Therefore, together, these have formed the era of face recognition explosion, which is what we call a technological breakthrough.
But in terms of industrial application, at present, we can see that the level of face recognition, especially dynamic recognition, in the United States and Israel is internationally advanced. In the real-time monitoring of the whole network, the FBI launched their next-generation electronic identification system last year, with a total investment of more than 1 billion US dollars. In the future, no matter where a crime is committed in the United States, the criminal suspect will be monitored and locked, and the whole network will be pursued.
what is the domestic level? The top academic level represents the stage of domestic industrial development. At present, there are mainly three forces. One is Professor Su Guangda from Tsinghua University, who is the father of face recognition in China. The second is Professor Li from the Institute of Automation of the Chinese Academy of Sciences. He made great achievements in the Asia Research Institute of Microsoft in his early years, and later went to the Institute of Automation of the Chinese Academy of Sciences, specializing in face recognition. In the Olympic Games, and later in many applications of face recognition, it provided better technology. The third team is the team of Professor Tang Xiaoou from the Chinese University of Hong Kong, who holds academic competitions every year. He is the record holder. At present, the recognition rate is higher than the overall level of human face recognition. Professor Tang helped iFLYTEK to establish his own position in the field of face recognition after speech recognition. Therefore, China is basically at this stage of development. Let's deduce the following stages, how do we identify the technology of face recognition, which one is reliable and which one is not, and we can put forward some key points for identification. Where are these points?
first, we should distinguish between dynamic and static cooperative identification or non-cooperative identification. The cooperative type is like Ant Financial, which needs data to cooperate with each other, and can collect two-dimensional data of the front face well. In addition, it is non-cooperative, and there is no way to cooperate with the exclusionary party. It is necessary to compare the pictures collected at random, and the recognition effect will be worse, but the timeliness of recognition will be high.
of these two modes, we pay attention to three points.
first, how many feature points have been extracted from your face modeling for comparison? This is the key node with some features on our face, and everyone is very different. The more data of feature points you select, the higher the accuracy of comparison will be. We also interviewed some experts, and the comparison of feature points they can do at present should be more than 7 points. At present, most companies that do system products such as face-brushing access control have about 5 feature points. So when we do research and communication, we can ask the number of feature points in the face recognition modeling of the whole company.
the second point is the data sample and size of the face recognition database, which is a very important indicator. Samples and sizes are the data sets we can provide. These must be used for human faces. For example, a person has 5 photos, all of which are taken of his face with different angles, positions and light. After these data are reasonably cleaned for machine training, including comparison and recognition, it can tell you whether the recognition is right or wrong. This sample number is very important, which is helpful for training and improving the accuracy of the model. Therefore, the size of the data sample set that can be labeled, which is at least one million at present, will make the recognition rate rise to the world's leading level, which is also one of the key points that can be identified.
the third point is whether your business model can acquire your whole data. We say that the comparison of face data forms a positive cycle model. In fact, the data source, the sample source of the face, comes from two very important channels, Meitu Xiu Xiu and Beauty Camera, which is a commercial exchange. This data, because of a desensitization process, only a few hundred key feature points are left, and the rest are omitted. After desensitization technology is used, the data is obtained, the training model is formed, and then the model is optimized, and the results are continuously fed back to obtain new data. With this, the data of your model will be well obtained, which is a very important indicator in the business model.
if you have these three indicators, it should be said that you have these three indicators at the same time, which may be something with great leading edge in the field of face recognition or potential for future development. At the same time, we analyze the intuitive performance, and there are two very important indicators in the performance of intuitive recognition. One is the accuracy of recognition. We have defined the face recognition contest that is compared every year in the academic circles just now, and now the test level is basically above 95%, but the comparison between people and pictures shows that it is this person. This is one, and then one is compared, and by the way, it is the second. All the people and photos are matched well, and the final correct rate is about 99.2%, which is what we call the current normal comparison method.
There is another very important method. We see that some face recognition technologies, including commercial banks and Taobao, will raise the problem of error rate. At present, this data can achieve an error rate of one in 1,, and others will take my ID card to compare it. If the machine can distinguish it, it will not pass, which is correct. If the machine passes my ID card when giving it to someone else, it may be a mistake, and the error rate should be about one in 1,. At present, there are only a handful of companies that can achieve such an error rate, which is a question of recognition accuracy.
In addition, it is very important to realize such accuracy in how many samples. There are two or three hundred people in a company. Among these people, it is not difficult to select and pass. However, among the big platforms of the Ministry of Public Security and provincial platforms, there are hundreds of millions of people's ID card photos. It is a very important indicator to accurately pick out ten or one hundred candidates. This range is narrowed down to this probability. How accurate can you be?
second, the speed of recognition. Similarly, the size of the sample set we just mentioned determines the speed of recognition. In itself, you don't have a lot of data in the samples that can be compared, for example, there are thousands, and the number of recognition is almost the same, which is reflected within one second. However, if you identify the photos accurately in a sample of hundreds of millions, the requirements for time and efficiency will be improved. So recognition speed is a very important indicator.
We have mentioned five indicators above, and we say that this can really make a comprehensive judgment on the company's specific capabilities and technologies.
Based on what we have said above, the company concerned is a company with recognition technology, and this recognition technology is a face recognition technology. As we said earlier, the strength of several academic circles speaking in China is very clear to everyone, which one comes from, which powerful academic team is backed by, and the strength of the research team makes this company a good position. For example, Iflytek, as we mentioned earlier, with the support of Professor Tang Xiaoou, their team is the first force in the academic field to support them, which is a resource advantage. For example, Sichuan Dazhisheng has close cooperation with Professor Li. At the same time, they also have their own unique technology in the field of image recognition, which undertakes a large number of national research fund projects. At the same time, we also emphasize one of them, namely, Sichuan Dazhisheng's face recognition technology, which is the human-computer interaction we see at present, because it is very different from two-dimensional plane recognition and has obvious advantages. Because it has collected the combination of three-dimensional surfaces between the five senses, it has collected more data. There are more features available for comparison. We found the side face and unclear photos in the video before, and it is difficult to identify who the suspect is, because in our second-generation ID card library, there is only positive data available for comparison. In the process of collecting data from three or four generations of ID cards, biometric features must be extracted, first of all, fingerprints, three-dimensional face recognition will be faster, and three or four generations may be extracted.
Once the 3D face data needs to be extracted, at this time, Sichuan Dazhisheng, as the only company with products and technology in China, is facing a broad market. But at the same time, we should also see that although 3D face recognition has amazing advantages and obvious disadvantages, it is difficult to select feature points, including side faces. At the same time, the factor of expression, in fact, is not as good as the factor of stereoscopic expression, and the efficiency is low when extracting, and the data consumed is also very large. So now, the application scenarios we can see are still small, including prisoners with criminal records in the United States. At present, our country is gradually promoting them in prisons, and all the people will collect them in the future. This is definitely a very huge market. At the same time, among the face recognition companies, the technical characteristics of this company and the ongoing overweight of 3D face recognition, there is a project of 18 million yuan to be invested in research and development, and the National Natural Science Foundation has also continuously supported their academic research project of 3D face recognition, which has been supported for many years. Therefore, in this field, it should be to the point of flowering and bearing fruit. So at this point, we especially remind everyone to pay attention to this company, which is technically scarce.
Iflytek is a typical business model we just talked about, which can realize face recognition data.