Lightweight Detection Method for Plaid Pattern Defects with Rank Self-selection Decomposition

To address the challenges posed by background patterns and the stringent real-time requirements in defect detection of plaid fabrics, we propose a lightweight defect detection method based on rank self-selection decomposition. Plaid background extraction algorithm is designed for obtaining the minimum periodic patterns of plaid fabrics. Filtering template parameters are derived through rank self-selection decomposition. Singular value decomposition and sparse representation are employed to eliminate the plaid background while preserving high-frequency defect areas. Subsequently, lightweight defect detection is then achieved through transfer learning, leveraging common defect features and transferring model parameters from grey fabric to plaid fabric. Experimental results show that the proposed method mitigates interference from complex plaid textures while achieving real-time, high-precision defect detection. The detection speed can reach up to 298 frames per second on four types of plaid pattern datasets, and up to 96.79% on mean average precision.