Research On Occluded Face Detection And De-occlusion Based On Deep Learning

Face detection and recognition task is an important branch in the field of computer vision,and has a wide range of applications in the fields of public security and identity authentication.In recent years,with the improvement of computing device and the technology of big data,the performance of face detection and recognition algorithms based on deep learning has been greatly improved.In the real world application,facial occlusion restricts the further development of face detection and recognition algorithms.Therefore,it is significant to study the face detection and de-occlusion algorithms under occlusion conditions.Based on the deep learning framework,this paper studies and improves the face detection and de-occlusion algorithms under occlusion conditions from three aspects: occluded face detection,facial occlusion segmentation and facial image in-painting.The main contributions are as follows:1.In order to solve the problem of occlusion feature pollution in face detection tasks,we propose a face detection algorithm based on attention mechanism and local features.On the basis of the one-shot face detection algorithm,we add different attention maps to the multi-level feature pyramid.At the same time,in order to solve the problem of the downsample operation in the convolution operation,which causes the occlusion feature is diffused to the face feature,we preserve the spatial information of the local feature in the channels of the feature map.The experimental results prove that the multi-level attention mechanism and the use of local features can effectively improve the accuracy of face detection under occlusion conditions.2.In order to solve the problem that the facial occlusion segmentation task lacks the well-labeled dataset,we propose an unsupervised facial occlusion segmentation algorithm based on deep auto-encoder.The image is encoded and reconstructed by a deep convolutional auto-encoder,and the result of occlusion segmentation is obtained by analyzing the reconstruction error map.At the same time,in order to improve the accuracy of segmentation,we utilize the consistency of the occluded image and the corresponding reconstructed image in the non-occlusion region and propose a novel reconstruction loss function.The experimental results prove that the proposed unsupervised method can effectively segment the occlusion region in the facial image.3.In order to solve the problem that the facial image inpainting task is difficult to preserve the identity information,we propose a facial image inpainting algorithm based on Generative Adversarial Networks.In the traditional method of image restoration based on deep self-encoder,the loss of confrontation is introduced,and the authenticity of the restored image is improved.At the same time,we use a pre-trained face recognition network to extract the identity information,and propose the identity loss to train the whole network.The experiments prove that the proposed algorithm method can repair the realistic facial image,and effectively improve the accuracy of face recognition after inpainting.