FabricMamba: A fabric surface defect detection system based on large kernel attention and visual state space

In the textile manufacturing industry, fabric defect detection is essential for ensuring product quality. Traditional approaches based on Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) often encounter scalability issues, particularly due to the high computational complexity of the self-attention mechanism in ViTs. To address these limitations, this study introduces FabricMamba, a real-time defect detection framework built on the You Only Look Once version 8 (YOLOv8) CNN architecture. The model enhances detection precision and efficiency for complex fabric defects in high-resolution images while minimizing computational cost. YOLOv8 was selected as the base model due to its strong balance between accuracy and inference speed, which is critical in fast-paced textile production settings. FabricMamba extends YOLOv8 with several innovations: the Parallel Large Separable Kernel Attention (P-LSKA) mechanism for multi-scale perception, the Visual State Space Module (MVSS) for long-range dependency modeling, the lightweight DySample module for reduced resource usage, and Programmable Gradient Information (PGI) to optimize training without increasing inference complexity. Extensive evaluations were conducted using a proprietary industrial fabric defect dataset and two public benchmarks, TILDA Textile Texture Database and FDDS Object Detection Dataset. FabricMamba achieved a mean Average Precision (mAP) of 90.0 %, 97.7 %, and 39.1 % on the respective datasets, outperforming the YOLOv8 baseline by 1.8 %, 2.3 %, and 2.0 %. Compared to Mamba-YOLO, FabricMamba reduced model size and computational requirements by 36.7 % and 33.1 %, respectively, with recall improving by 2.9 %, 1.4 %, and 4.0 %. These results confirm the model effectiveness and practical potential for industrial fabric inspection tasks.