FF-YOLO: Fashion Fabric Detection Algorithm Based on YOLOv8

In monitoring fashion trend and managing inventory, promptly detecting and identifying popular fashion fabric material is crucial for manufacturers and sellers. However, most traditional fabric detection are working on raw material(e.g. a block of raw fabric material with a solid background), which is not robust enough to handle complex fashion images(e.g. a human model and complex backgrounds). To address this problem, this paper proposes a lightweight and efficient fashion fabric detection algorithm named FF-YOLO, based on the YOLOv8 architecture. First, we introduce a Simple Attetion Module (SimAM) in the backbone to improve the feature extraction capability of the model. Furthermore, we propose a Lightweight Multi-Level Asymmetry Detector Head (LADH) to replace the head, improving the computational efficiency of the model’s inference process. Last, we replace the original loss function with Wise-IoU to improve the localization ability of the detection box. The experimental results show that FF-YOLO achives an average accuracy of 75.5% and a frames per second (FPS) of 105 for fashion fabric detection. Compared to the original YOLOv8 model, the mAP is improved by 2.9% and the FPS is improved by 6 frames. Meanwhile, the floating point operations per second (FLOPS) computational complexity is reduced by 5%. The results show our method proves to be effective and lightweight in fashion fabric detection tasks.