DSA-YOLO: A Model for Clothing Images Segmentation Based on YOLOv8

With the development of the internet industry, online clothing recommendation and rating systems have grown rapidly. Existing clothing recommendation and analysis systems face issues such as insufficient datasets, low input image resolution, and poor real-time performance. This paper proposes a clothing images segmentation algorithm based on YOLOv8, named DSA-YOLO, to better accomplish the task of clothing images segmentation. The proposed DSA-YOLO algorithm employs HWD downsampling instead of the original downsampling method, reducing the loss of key spatial information. Additionally, DSA-YOLO replaces some C2f components in the YOLOv8 network with the C2f-DSA module, which utilizes Deformable Spatial Attention (DSA) to avoid a quadratic increase in computational load with kernel size. Experimental results demonstrate that DSA-YOLO achieves a 2.6% increase in accuracy on the DeepFashion2 dataset compared to YOLOv8. The model is particularly adept at segmenting clothing images, is conducive to deployment on resource-constrained devices, and satisfies the requirements for real-time detection.