Improving the Application of YOLOv8 in Image Object Detection

As deep learning develops, the use of object detection to analyze image content has achieved stage-by-stage success. However, due to the problems of spatial occlusion and slow convergence, the accuracy is poor and models are complex. To address the above problems, a new YOLOv8s network structure that improves the accuracy with almost no increase in training overhead is proposed. First, the hierarchical feature extraction is utilized to enhance the positional information of the feature map by adding Coordinate Attention at backbone and neck networks layer-by-layer. Then, GCNet is introduced at the last layer of neck network to model global context and merge hierarchical features. Finally, SIOU Loss is used to accelerate the convergence speed to further improve the accuracy. The experimental results demonstrate that the enhanced model improves mAP by +4.6% over YOLOv8s and accuracy by +2.7% over YOLOv9s. Additionally, it slightly decreases model complexity and training overhead by 0.2 GFLOPs.