Face representation and similarity calculation method

Abstract

The invention discloses a face representation and similarity calculation method. At an off-line training stage, a face training set with labels is collected, wherein each person contains 50 to 100 facial images; after image normalization in the set, dividing is carried out to obtain 2*2 face blocks with identical sizes. For each face block, a small block with the k*k pixel is extracted densely, mean-value return-to-zero and variance normalization pretreatments are carried out, and then convolution kernel learning is carried out based on a K-means cluster. K convolution images are obtained for each face block, the mean value Pooling operation and ReLU non-linear operation are carried out on K*2*2 convolution images of the whole input face, and then features of all response diagrams are straightened. For the straightened features, principal component analysis (PCA) projection is learned; after PCA dimension reduction, projection determination is learned based on a linear discriminant analysis (LDA) algorithm to obtain compact and robust face representations; and inner product operation is carried out on the face representations of two images to obtain the similarity.

Description

technical field [0001] The invention belongs to the technical field of computer vision and image processing, and in particular relates to the expression of human face and the calculation method of similarity. Background technique [0002] Computer face recognition refers to the use of computer analysis image and pattern recognition technology to identify or verify one or more faces from static or dynamic scenes based on the known face sample library, and to extract multiple possible faces by using feature extraction technology. The feature representation, this technology is widely used in public security, identity verification and other important occasions, the key to effective face recognition lies in fast and accurate face feature representation. There are already some facial feature representation algorithms in the prior art, which use convolution kernels to perform convolution processing on input face images to obtain multi-dimensional feature vector representations of h...

AI Technical Summary

Problems solved by technology

There are already some facial feature representation algorithms in the prior art, which use convolution kernels to perform convolution processing on input face images to obtain multi-dimensional feature vector representations of human faces. The selection of convolution kernels involved in the operation does not make full use of the rich face training data , setting the dimension of the feature vector too small will affect the accuracy, and setting it too large will affect the speed, thus restricting the accuracy of facial feature representation and similarity calculation as a whole

Images