Attention point mechanism-based locating loss calculation method and system for target detection system

Abstract

The invention relates to an attention point mechanism-based locating loss calculation method and system for a target detection system. The method comprises the steps of calculating a weight matrix ofa convolutional graph used for generating a target prediction frame by utilizing an attention point mechanism in a forward target detection process of a convolutional neural network; determining a non-overlapping region of the prediction frame and a target frame caused by separate deviations of central point horizontal coordinates, central point longitudinal coordinates, transverse width and longitudinal height of the prediction frame; according to weight values of elements located in the non-overlapping region of the prediction frame and the target frame on the convolutional graph in the weight matrix of the convolutional graph obtained by the attention point mechanism, calculating locating loss; calculating gradients of the locating loss to the prediction frame, the target frame and theweight matrix in back error propagation; and judging whether an iterative process of the convolutional neural network is ended or not, and if yes, ending the process, otherwise, returning to the abovementioned steps. Therefore, the target detection accuracy can be improved; and the manpower, material and time costs can be effectively reduced.

Description

technical field [0001] The invention relates to a method and system for target detection in the direction of computer vision in the field of pattern recognition, in particular to a method and system for calculating positioning loss based on a point of interest mechanism in a target detection system. Background technique [0002] In recent years in the field of target detection Faster-RCNN, SSD and other classic algorithms based on convolutional neural networks, in the final stage of network forward target detection, the system's predicted bounding-box (predicted bounding-box) t of the object position has been known. u After and the target box (ground-truth bounding-box) v in the training data, it is generally used function to compute the object localization loss L loc (t u , v). Therefore, in the reverse error propagation stage, using the chain derivation rule, according to L loc (t u , v) the gradient at each layer of network nodes and links, modify the link weights a...

AI Technical Summary

Problems solved by technology

[0013] The traditional positioning loss calculation method in the target detection system does not make full use of the distribution trend information of the number and importance of the features of the target and background contained in the convolution graph, which leads to the inability of the target detection system to effectively prevent the prediction frame from missing the target object The important part of the relatively concentrated features, and avoiding the prediction frame from mixing with the features of other surrounding objects, making it difficult to ensure the accuracy of target detection

Images