Research On Face Recognition Based On Local Information Obscured Scene

The application of face recognition technology has been widely concerned and studied.At present,the occlusion problem is the main problem faced by face recognition technology in practical applications.The partial occlusion of the face causes the defect of the face information,which has a serious impact on the extraction of the facial features and the final recognition result.This paper studies the face recognition problem of partial occlusion from the following two aspects:(1)For occlusion face recognition of fixed area,this paper proposes a face recognition algorithm based on occlusion dictionary for non-uniform block sparse representation.Firstly,the face is masked for facial features.For the five senses that can be detected,the split ratio is 1:20 for the eyebrows,1:12 for the eyes,1:5 for the nose,and 1:3 for the mouth..Then,the training sample and the test sample are non-uniformly divided in the same proportion,and then the corresponding occlusion dictionary and non-occlusion dictionary are constructed for each of the divided blocks,and the test samples are obtained in the occlusion dictionary and the non-occlusion dictionary.The sparse coefficient,finally,according to the statistics,the most number of classes referred to by each sub-block is used as the final recognition result.In this paper,three kinds of occlusions are added to the sunglasses,the mask and the two occlusion methods.Compared with the Sparse Representation Classification(SRC),this paper has better recognition results.In the case where the mutual opacity of sunglasses and scarves is 60%,the recognition rate of the proposed algorithm is up to 93.46%,which is 50% higher than the original SRC recognition rate.At the same time,the higher the face feature dimension,the higher the recognition rate.(2)For occlusion face recognition in non-fixed areas,this paper proposes a face completion algorithm based on conditional antagonism network.Firstly,the occlusion face is firstly extracted from the facial features such as facial features through the convolutional neural network,and is used as constraint information input generation model and discriminant model.The generated model is a self-encoder that reconstructs the face occlusion region.Content,the results are input into the local information discriminant model and the global information discriminant model and the semantic parsing network,combined with the loss function,and finally generate a complete face.The training on the Celeb A dataset,compared with the Context Encoder(CE)model,the algorithm model proposed in this paper can get a better reconstructed face image,and at the same time,the lateral occlusion of different face occlusion methods,the algorithm model is In the case of occlusion at the mouth and nose,the resulting face is more confident and visually more harmonious.