Occlusion adaptive human face recognition method

Abstract

The invention discloses an occlusion adaptive human face recognition method. The method includes the following steps that 1, network training is pre-processed, a human face data set is selected, the human face data set is a facial picture under a non-ideal condition, supervised learning is carried out by taking various tags in the data set as an output result of the pre-processing network, an activation function of a nerve cell selects f(x)=max(0,x), an input layer of the network inputs a human face grayscale image with a resolution of 196*196, a first layer convolution uses a convolution kernel of 9*9, the convolution kernel of a first layer polling corresponding to that of the first layer convolution is 5*5, and the output of the full connection layer of the network is passed into the next processing as the face feature completed by the preprocessing. According to the occlusion adaptive human face recognition method, the first step of human recognition is improved, the introduction of adaptive discrimination improves the recognition ability in the condition of occlusion, at the same time, the demand for computing performance under non-occlusion condition is reduced, and meanwhilethe preliminary preprocessing is carried out on the feature information of the human face.

Description

technical field [0001] The invention relates to a recognition method, in particular to an occlusion self-adaptive face recognition method. Background technique [0002] Face recognition is a major focus in the field of biometrics today and has high practical value, but it does face two major problems, small samples and non-ideal acquisition conditions. The present invention intends to solve these two major problems and improve the usability of the system. Wright J et al. proposed a sparse representation (SRC) method in the article Robust face recognition via sparse representation, but this method cannot overcome the influence of small samples and intra-class errors, such as identifying objects wearing sunglasses. Deng W et al improved the SRC method in the article Extended SRC: undersampled face via intraclass variant dictionary, which improved the adaptability of the SRC method. But because they are all based on linear methods, nonlinear intra-class interference cannot be...

AI Technical Summary

Problems solved by technology

A single sparse representation classifier has reduced recognition ability and usability under small sample and occluded conditions

Images