Lightweight improvement of YOLOv6 algorithm for small target detection

Abstract At present, there are still many obstacles for small target detection in the field of computer vision. Although the existing target detection technology can extract the semantic information of small targets by multi-scale fusion and other methods, the detection speed still needs some improvements. In this paper, the target detection algorithm of YOLOv6 is improved by lightweight. On the basis of VGNetG architecture, a lighter Backbone is redesigned, and ECA attention mechanism with extended position information is introduced to make up for some loss in detection accuracy, so as to ensure that the model can obtain lower computation without obvious decrease in detection accuracy. According to the test experiment of DOTA, a public data set, the FPS of VE-YOLOv6 proposed in this paper increases by 196, and the model computation and parameter number decrease by 2.91 and 0.8M, respectively, when the mAP decrease is not obvious with the original YOLOv6 algorithm.