Contents I. Introduction
The modern society is a society that is tightly integrated with computers and the Internet. Under the impact of a large amount of data in the information age, the problems that have been left over have been solved by the rapid development of information technology, or they have been solved based on existing computer technologies. It gives many new thinking of solving those problems. As we all know, the issue of public safety has been an eternal topic since ancient times. In 2016, the nation’s procurator organs approved and arrested more than 800,000 suspects in various criminal categories. In response to the country’s call for a “safe city,” many scholars conducted a large number of related studies on personal identification. Among them, the identification of people in the video is the focus of research. Although the current research on face recognition has entered the practical stage, there are still many problems. First of all, for the data sources obtained, including indoor and outdoor surveillance video, traffic surveillance capture, and personal surveillance camera video recording, due to hardware limitations, there are low bit rates, low resolution, unclear video[1], etc. Happening. Under certain circumstances, due to many reasons such as light, scene, and facilities, video blurring often occurs. Although artificial can distinguish, it is difficult for computer recognition. Secondly, because the object of recognition is a character, its image features are very complex. In most cases in the video, the character is in a state of continuous movement, and the regular switching angle, the person’s facial expression is definitely not fixed. This makes it more difficult to extract character characteristics and accurately determine the person’s identity. Moreover, for the target person to be identified, there is often only a small amount of data, and it is not possible to train a large amount of data on it, which also brings a great deal of resistance to the study of machine learning and the like. On this basis, whether or not a method for obtaining relatively accurate person recognition using a relatively small amount of training data has become a practical problem whitch is worthy of study. In the field of deep-learning image research, people are looking more closely at designing more elaborate network models in order to complete more accurate image recognition. Among them, the residual network has emerged as a highly influential network model (ResNet) since 2015[2] and it has quickly attracted the attention of many scholars. Its residual learning method greatly deepens the depth of the network, but it does not make the network more complicated. At the same time, it also solves the problem of "degeneration". To highlight its advantages. In this context, this paper uses a trained ResNet model for face recognition in the dlib library to design a set of processes for character recognition and implement it, analyze the results obtained, and discuss the application prospects of this technology practical value. And look at its development prospects simply.
