Lightweight network based on improved YOLOv8n for clothing image detection

Abstract In response to the issues of high computational complexity, large model size, and high computational resource requirements in deep learning-based detection models, an improved lightweight clothing image detection algorithm based on YOLOv8n is proposed. Firstly, some GhostConv are employed to replace ordinary Conv in benchmark model to reduce the number of parameters. Additionally, we design a C2f-GD module to enhanced the bottleneck structure in the C2f architecture by using GhostConv and DCNv2 modules, which can improve compression detection performance while reducing model size and parameter count. Finally, the Inner EIoU loss function is substituted for the original CIoU loss function to enhance the regression analysis performance of the model’s bounding boxes. Experimental results on the open-source fashion dataset Deepfashion2 demonstrate that the lightweight clothing image prediction network demonstrates a significant level of reduced size and parameter count, improved detection accuracy. Compared to the YOLOv8n benchmark model, the proposed model achieves a 1.1% increase in precision and a 2.6%increase in mean average precision(mAP),while reducing model size and parameter count by 0.65MB and 0.315×109,which has good practical value.