Research On Small Object Detection In Complex Background

Small target detection in complex backgrounds is usually affected by factors such as different light conditions,mutual obstruction of the target,and large changes in the target scale,which will cause the existing target detection model to detect the effect.In this article,small targets can be used in small targets in complex contexts,and the problems of uneven distribution.Study from the two aspects of multi-scale characteristics fusion and effective use of contexts to improve the detection effect of small targets.The main research content in this article is as follows:(1)Briefly analyze the difficult problems existing in the current small target detection.From five aspects of data enhancement,multi-scale characteristics fusion,context information,generating confrontation network and other improved methods,the current mainstream research direction is combed.(2)For the misunderstanding and missed inspection problems existing in small target detection,the anchor-free box detection model CenterNet is based on the center point.The Inception module enables each layer of output features to integrate features after obtaining a wider level of global information,thereby reducing the problem of small targets in the case of mutual blocking;Make small target positioning more accurate.Apply the improved model to the Visdrone2019test.The experimental results show that the model detection performance has been effectively improved and confirmed the feasibility of the module.(3)For the problem of poor detection effects of remote dense targets,further research on multi-scale methods on the basis of the above network proposes a multi-scale characteristic fusion module based on hybrid domain attention.This module can fully change Deep features participate in the level of superficial characteristics,filtering invalid samples in the deep and shallow layers,learning more effective information from the deep feature map,and bringing gain to the detection of small targets;The feature enhancement module,after the output feature layer,to obtain richer semantic information features.The improved algorithm significantly improves the recognition accuracy of the remote dense target,and at the same time reduces the probability of various categories that is recognized as the background,and has good results.