Pedestrian re-identification model optimization method based on fusion loss function

Abstract

The invention discloses a pedestrian re-identification model optimization method based on a fusion loss function, and the method comprises the steps: (1) obtaining a pedestrian re-identification modelthrough the training of a neural network, and calculating the feature spatial similarity of each pedestrian image in the model through a cross entropy loss function; (2) setting a loss threshold value according to the spatial similarity, and constructing a triple object by using the threshold value; and (3) constructing triple loss function iterative optimization by using the triple object to obtain an optimized pedestrian re-identification model. According to the method, the cross entropy loss function and the triple loss function are combined, deep learning and feature recognition are applied to the pedestrian re-recognition problem, and the important significance is achieved for improving public security management.

Description

technical field [0001] The invention relates to a pedestrian re-identification method, in particular to a pedestrian re-identification optimization calculation method based on fusion loss. Background technique [0002] With the rapid development of artificial intelligence technology and the increasing popularity of video surveillance equipment, intelligent surveillance has attracted widespread attention from all walks of life for its accuracy, timeliness and rich functions. At present, the pedestrian re-identification problem has received extensive attention in the field of machine vision. In recent years, many related algorithms have been well implemented on this problem, and excellent results have been obtained in terms of recognition accuracy. With the continuous improvement of the cost performance of security systems in the future and the development of digital high-definition and intelligent technologies, the market application space will continue to grow. [0003] We ...

AI Technical Summary

Problems solved by technology

However, we found in the process of practice that the acquisition of pedestrian pictures is continuous. Once the amount of picture data is too large, such as the Mars data set, a deeper network architecture is needed to extract features, but this often occurs Gradient backpropagation to In the previous layer, repeated multiplication makes the gradient infinitely small, or even disappears, resulting in a decrease in the accuracy of the training set

Secondly, the existing pedestrian re-identification model only trains and extracts a limited number of main features of pedestrians, and the extraction of more detailed feature points is far from enough to meet the requirements of real-world pedestrian image recognition accuracy

Again, in the actual training process, it is found that if only the cross-entropy loss function is used to optimize the existing pedestrian re-identification model, due to the large number of data sets and the large number of similar pedestrian pictures, the gap between the output and the expectation is compared only. Although the calculation speed is relatively fast, it will be lacking in accuracy

And it will also lead to redundant parameters and long training time, and the calculation overhead is very high in problems such as cluster retrieval. When it is hoped to find multiple target persons, the recognition effect is not very good, because each match must be passed with each A network of paired images in a picture library is not suitable for end-to-end recognition of the Internet of Things

Images