Research On Face Recognition Methods Under Partial Occlusions

The concept of artificial intelligence has been put forward for sixty years,and the technology of face recognition as a sub-direction of artificial intelligence,has also experienced a few decades of development and progress.However,most of these developments and progress were made under collaborative conditions,which needs the cooperation of the user.On the other side,under the non-collaborative conditions,face recognition technology will be influenced by light,pose,expression,age,occlusion,long distance and other factors of interference,leading to the performance degradation of existing face recognition technology,and serious block of the practical process of face recognition technology.In view of these disturbances under non-collaborative conditions,it is of great significance to study the effective identification method to solve the problems caused by these disturbances and make the face recognition technology develop from the collaborative condition to the non-collaborative condition.In this thesis,two kinds of methods are investigated for face recognition problem with partial occlusion.One is the method based on the global image.This kind of method takes the whole image as the input of the system,and utilizes the global information of the face image to complete the recognition task.A typical technical scheme is to consider the partial occlusion face recognition problem as a reconstruction problem.The class that results in the minimum reconstruction error is the classification result.The other method is based on the local image.Such methods take into account some properties of the occlusion in the image,such as continuity and block-like distribution.We can utilize the local image information to identify,so that the performance degradation caused by partial occlusion can be reduced as much as possible.In view of the shortcomings of these two methods,this thesis proposes two novel methods and models.One is to actively introduce artificial occlusion into the training data.In face recognition for realastic applications,the training data are always clean while the probe data are partially occluded.We find that introducing artificial occlusions into the training data is helpful in this situation.The incremental training data is decomposed into a class-specific dictionary,a non-class-specific dictionary and a sparse noise or corruption matrix by the sparse and dense hybrid representation framework(SDR).The artificially introduced occlusions play an important role in building the discriminative faces for classification during SDR,which can significantly improve the recognition rate of the algorithm.Aiming at the problem of low recognition rate and slow running speed of the existing occluded face recognition technology,another method of fast collaborative matching is proposed.The method is based on the sparse representation classification to quickly determine the set of suspected objects,and uses the DICW to accurately identify the target,so as to improve the recognition rate and speed of the algorithm,thus promoting the practical application of face recognition technology.