LDDNet: a lightweight steel surface detection network based on decoupled structure and dynamic convolution

Automated visual inspection has the potential to improve the efficiency and accuracy of inspection tasks across various industries. In this work, we focus on steel surface defect detection, a critical task for ensuring product quality and safety. Traditionally, this task relies experienced operators, which is subjective and time-consuming. Given this, we propose LDDNet, a novel lightweight convolution neural network based on two key designs, which effectively addresses challenges commonly encountered in industrial environments, including constrained resources and irregular surface defects. Firstly, we propose a decoupled design comprising channel processor and patch processor which significantly reduces the heavy computational burden. Secondly, we incorporate a deformation feature modeling capability to handle the irregular appearances of various defect types, with minimal additional computational cost. Finally, extensive experiments validate that our method surpasses widely used visual methods, achieving a remarkable mean Average Precision of 71.4% with a lightweight 8.2M parameter backbone for steel surface defect detection.