Improved YOLOv8 for Small Object Detection
Huafeng Xue, Jilin Chen, Ruichun Tang · 2024
Small object detection is a challenging problem. Small objects usually have small size, low contrast and are easily occluded, so it is difficult to detect accurately. This paper offers an improved YOLOv8 model to boost the efficiency of small object detection, which introduces small object detection layer, GAM (Global Attention Mechanism) and SPPFCSPC (Spatial Pyramid Pooling and Fully Connected Spatial Pyramid Convolution). To begin, a small object detection layer is included for feature fusion, considering both the shallow feature map's intricate features and the deep feature map's semantic features. Subsequently, we introduce GAM to make the network more attentive to the minuscule area and enhance the detection precision. Lastly, we optimize SPPF to SPPFCSPC, thus augmenting the model's feature expression capacity. Experimental results demonstrate that our technique accomplish a noteworthy enhancement in small object detection tasks.