Efficient fabric defect detection based on lightweight model

Abstract

Abstract In the field of textile inspection, the detection of small-sized targets poses a challenge due to its high sensitivity to size variations. Currently, the commonly used methods in this field often rely on a large number of parameters and computing resources. Based on YOLOv8m, this study integrates the Coordinated Attention (CA) mechanism and the Mixed Local Channel Attention (MLCA) strategy and proposes a lightweight textile defect detection model - YOLOv8-mini. The CA mechanism enhances the discriminative ability of features by explicitly modeling the interdependencies between spatial and channel coordinates, enabling more effective feature representation for varying defect sizes. The MLCA strategy captures both global contextual information and local detailed features through a multi-scale approach, which is crucial for accurate defect localization and classification. On the dataset of the Xuelang Manufacturing AI Challenge, compared with the current state-of-the-art methods, the model proposed in this study has achieved a significant reduction in the number of parameters, decreasing to 18.8M (a reduction of 27.4%), while the mean Average Precision at 50% recall (mAP50) has reached 93.3% (an increase of 4%). The improvement in the efficiency of this model is expected to boost productivity and reduce production costs in the textile manufacturing process.