A lightweight small object detection network inspired by the visual area V2

In high-resolution aerial images, due to the low feature density and small pixels of the objects to be detected, it is difficult to capture subtle details during feature extraction, which can easily lead to missed detections and false detections. To address this problem, this paper proposes a small object detection algorithm inspired by the information processing mechanism of visual area V2, in order to improve the network’s ability to extract detailed features such as the edge and direction of small objects. Inspired by the information processing mechanism of complex cells and hypercomplex cells in biological visual area V2, we designed a complex cell module (CCM) that mimics the edge-sensitive properties of complex cells and a hypercomplex cell module (HCM) that simulates the edge and direction-sensitive properties of hypercomplex cells. By simulating the sensitive characteristics of complex cells and hypercomplex cells to edge and direction information, the model’s ability to extract edge and direction features of small objects is enhanced. In addition, inspired by the bottom-up attention mechanism between visual cortexes, this paper designs a spatial enhanced attention module (SEAM) between Neck and Head, which uses shallow features to modulate deep features to retain key shallow information while focusing on small objects. The results show that on the UAV small object dataset VisDrone2019 and the remote sensing small object AITODv2, the network we designed achieved an accuracy index (mAP50) score of 48.9% and 49.1% with 1.3M parameters, successfully achieving a good balance between lightweight network and detection accuracy, achieving the best performance of lightweight models, and effectively reducing the occurrence of missed detections and false detections. The code will be available online at https://github.com/Dzzz614/V2.