A Real-Time Defect Detector With Content-Guided Feature Extraction and Detail Enhancement

Surface defect detection in industrial materials faces two main challenges: complex backgrounds and missed small defects. To address these problems, this paper presents CGRNet, an efficient real-time detector. CGRNet introduces a content-guided feature extraction network that uses global context to guide adaptive high-frequency filtering. This suppresses low-frequency noise and enhances edge and local features. The detail enhancement module with a double residual attention mechanism integrates channel-specific spatial attention maps for multi-scale feature interaction. This integration improves the detection of small defects. The SIoU loss function and Lion optimizer speed up convergence and improve localization accuracy. Experimental results show that CGRNet achieves a 93.6% mAP on the PVEL_AD dataset, surpassing existing methods in accuracy. On the NEU-DET dataset, the model runs at 81.9 frames per second, satisfying real-time application requirements.