Small target detection method based on SSD network

Abstract

A small target detection method based on SSD is provided by the invention. The feature information of a small target is enriched through multi-scale feature fusion, learning of high-resolution features with a large receptive field, introduction of an attention mechanism and other feature enhancement modes, the feature quality of the small target is improved, and the A nchor is learned through theconfidence of Anchor-object matching strategy joint classification and positioning, and the problem that no interaction exists between classification and positioning and between classification and positioning due to independent supervised network learning of each allocated Anchor is solved, so the detection precision and the positioning precision of the small target are improved; and in addition,the diversity of data set samples can effectively improve the precision and the generalization ability of the model for small target detection, so in the aspect of data sample collection, various influence factors such as different illumination, different postures and shielding degrees are referred.

Description

technical field [0001] With the development of computer technology and the wide application of computer vision principles, the use of computer image processing technology to track targets in real time has become more and more popular. Dynamic real-time tracking and positioning of targets is widely used in intelligent transportation systems, intelligent monitoring systems, and military target detection. It has wide application value in medical and other fields. The invention relates to the field of image processing and target detection. Aiming at the problem of poor detection effect of small targets in the target detection field, a method for improving the detection accuracy of small targets is proposed. Background technique [0002] Target detection technology is divided into traditional methods and deep learning methods. Traditional target detection methods include HOG feature method, Haar-Like feature method, and LBP feature method. The feature layers of these methods are v...

AI Technical Summary

Problems solved by technology

[0005] Nowadays, the detection of small targets is based on the improvement of the target detection method. The Feature Pyramid Network (FPN) uses bilinear interpolation to upsample the high-level feature map and then add it to the shallow feature map at the element level. , which enhances the ability of the network to extract multi-scale features. FPN includes two paths: bottom-up and top-down. The bottom-up path is usually a convolutional network for extracting features, and the bottom-up path will cause feature maps. The spatial resolution is reduced, but the semantic value of the feature map is increased accordingly. In order to skip the connection of the feature map and help the detector to better locate, FPN designs a horizontal connection structure, and the multi-scale prediction structure designed by FPN makes reasonable use of shallow features. and high-level feature information, effectively improving the detection accuracy of small targets, improving the feature fusion method based on FPN, changing FPN from bottom-up upsampling feature fusion to direct fusion, and using multi-scale prediction methods to detect targets. Inspired by the receptive field structure in the human visual system, a spatial convolution structure is proposed to simulate the relationship between the receptive field size and the center in the human visual system, increase the receptive field of the feature map, and enrich the details and edges of the feature map. And other information, although the detection accuracy of the network is improved while ensuring the detection performance, but the network structure does not use the spatial context information of the feature map, and the detection accuracy still has room for improvement

Images