Research on Lightweight Small Object Detection Method with Fusion Attention Mechanism

Abstract

In the task of small object detection, due to the relatively small size of the objects, complex situations such as occlusion and overlap may arise in high-density scenes. These factors can significantly increase the issues of missed detections and false detections in object detection tasks. To address the aforementioned issues, this paper proposes a lightweight small object detection model, YOLOv5s-MGC. Firstly, we introduce an improved Mobilenetv3 network to replace the backbone network of YOLOv5s, aiming to improve the network's capability for effective feature extraction. Secondly, we incorporate the GSConv module into the feature fusion network to reduce model parameters and strengthen the network's ability in fusing feature information. Finally, we have integrated Convolutional Block Attention Module (CBAM) at the forefront of the detection heads, achieving the fusion of fine-grained features. We assessed our approach on the openly accessible dataset 'SHWD' and contrasted the results with the baseline YOLOv5s and various SOTA object detection models. Precisely, our method attains an accuracy of 91.1%, exceeding the baseline module by 1.6%. Furthermore, the module size is reduced to 1.97MB and it achieves a 52.4 FPS. In comparison to other prevailing algorithms, YOLOv5s-MGC enables efficient and accurate small object detection.