Face Detection And Recognition In Complex Environments

With the development of deep learning, the perception of facial image has wide prospects of application.In a crowded environment, there is a problem that faces cannot be detected due to mutual occlusion of faces, and the low resolution causes the face to be blurred. In order to detect the missing faces which were occluded in the dense environments, a kind of model, which based on multi-task cascaded convolutional network, was put forth. We improve the detection scheme by using Soft non-maximum suppression algorithm.Aiming at the problem of faces with proportion of inconsistent caused by the inconsistency of the distance between the face and the camera device, this paper extracts facial features based on the deep res-net. For high and low resolution facial images, a dual-branch extraction feature network is used to make images of different resolutions are clustered in a public feature space. At the same time, the central loss function, parameterized rectification linear unit and batch normalization are used to improve the network training process, which improves the network's feature classification performance and the convergence speed in the training process. WIDER FACE database is used to verify the face detection method. LFW database and YALE face database are used to verify face recognition method. Experiments show that the network proposed in this paper has good robustness against occlusion and low resolution in face detection, and effectively expands the distance between classes in face recognition and improves the recognition accuracy.