Cascaded neural network-based face recognition method

Abstract

The invention relates to a cascaded neural network-based face recognition method. The method comprises the following steps that: a) a training-used face images set is established; b) key points of each face image are detected; c) key point information is utilized to normalized the images into standard gray-value images; d) the standard face images are subjected to gray value statistics normalization; E) a depth neural network is trained, and input is a pair of images, and outputted 0 indicates that the two images are judged as a negative case pair, and outputted 1 indicates that the two images are a positive case pair; f) all image pairs are inputted into the neural network, and the proportion of negative case pairs which are wrongly judged as 1 is made to be lower than a given proportion; g); the step e) is repeatedly used; f), the total number of networks achieves a predetermined number, the method is terminated; and h) any pair of new face images is given, and an obtained cascaded neural network is utilized to judge whether the one pair of new face images is a negative case pair or a positive case pair. With the cascaded neural network-based face recognition method of the invention adopted, the recognition rate of face recognition can be effectively improved, and wrong recognition rate can be greatly reduced, and the reliability of a security monitoring system can be greatly improved.

Description

technical field [0001] The invention belongs to the technical field of image processing and face recognition, and in particular relates to a face recognition method based on a cascaded neural network. Background technique [0002] Face recognition (face recognition), that is, given a pair of face pictures, it is necessary to determine whether the two face pictures belong to the same person. Face recognition is the core module of the face analysis system, and its performance determines the accuracy and reliability of the system. There are two main performance indicators, the recognition rate (given a pair of pictures belonging to the same person, the probability that the system judges it correctly) and the misrecognition rate (given a pair of pictures that do not belong to the same person, the system judges it wrong, probability that they belong to the same person). [0003] The traditional face recognition algorithm can be mainly divided into two steps: first, for any face...

AI Technical Summary

Problems solved by technology

Traditional methods have two main limitations. 1) The eigenvector extraction algorithm for face images is generally designed manually or relies heavily on expert experience, and it is difficult to automatically learn from samples to generate a large number of differentiated eigenvectors for subsequent similarity calculation systems 2) No matter what method is used, it is difficult for a single similarity calculation system to achieve a better balance between the recognition rate and the false recognition rate. Practical application scenarios often require a very low false recognition rate (less than one thousandth or one in 10,000), the recognition rate should be as high as possible, and it is often difficult for a single similarity system to accurately meet the balance between the two indicators

Images