A closed-loop cGAN-based framework for surface defect detection in industrial LPG rubber hoses

Industrial LPG rubber hoses require reliable surface inspection to prevent safety hazards and production losses; however, real defect data are typically scarce and severely imbalanced, making purely supervised detectors prone to missed rare defects. This study proposes cGAN-Eff-Yolo-Net, a closed-loop framework that integrates class-conditioned defect synthesis with a two-stage screening-and-localization pipeline. Unlike conventional offline generative augmentation, synthetic defect images are generated during training, pseudo bounding-box labels are produced by a temporary detector, and downstream screening and localization losses are fed back to guide the generator toward task-relevant synthesis. Confidence-based pseudo-label filtering, periodic temporary-detector synchronization, and discriminator-derived realism-aware weighting are further employed to reduce the influence of unreliable synthetic supervision. Experiments were conducted on an industrial LPG-hose dataset containing 193 expert-confirmed defective images from five defect categories under severe class imbalance. The proposed framework achieved an Accuracy of 93.98 ± 0.46%, Recall of 94.42 ± 0.47%, F1-score of 96.89 ± 0.25%, and mAP@50 of 89.68 ± 2.32%. Compared with the strongest baseline, YOLOv11, mAP@50 increased by 3.42% points. Per-class analysis further showed AP@50 gains of 10.22 and 6.45% points for the under-represented Note and Bubble categories, respectively. Generation-quality evaluation also showed substantially lower FID and KID than the corresponding no-task-feedback variant. These results indicate that closed-loop task-driven synthesis can improve minority-defect coverage while maintaining competitive inspection reliability and practical deployment efficiency under limited industrial defect data.