A clothing fabric segmentation method based on improved YOLOv8

To address the inefficiencies, high costs of manual fabric cutting, and the inability to meet the demands of the smart clothing era, as well as the limitations of traditional fabric segmentation models such as low accuracy, limited information, and deployment challenges, this paper proposes an improved YOLOv8 model for fabric instance segmentation. First, a C2f_DeformConv module is designed in the backbone network to mitigate the effects of fabric color gradients and complex backgrounds, thereby enhancing the model's detection accuracy. Second, the BiFPN network is introduced as the main feature extraction network to more effectively capture target boundary information and improve edge segmentation accuracy. Finally, the Shape-IoU loss function is employed to provide more refined evaluation results and enhance the model's generalization ability. Experimental results show that the proposed algorithm improves mAPmask@50 by 5.4 percentage points, model accuracy by 5.8 percentage points, and recall by 4.7 percentage points compared to the baseline YOLOv8n-seg model, significantly enhancing the precision and accuracy of fabric segmentation. The findings indicate that the improved algorithm achieves notable improvements in various performance metrics, making it effectively applicable to fabric segmentation tasks.