An Integrated and Robust Vision System for Internal and External Thread Defect Detection with Adversarial Defense

In industrial automation, threaded mechanical components present significant challenges for inspection due to their complex geometries and the high concealment of micro-defects. This paper proposes an integrated detection system for internal and external thread defects, combining image enhancement, data synthesis, lightweight detection, and adversarial defense. A unified image acquisition platform with fisheye lenses and high-definition industrial cameras enables synchronized imaging of both internal and external threads. By incorporating MLWNet-based dynamic deblurring and DarkIR-based low-light enhancement, image quality is significantly improved (PSNR 30.3 dB, SSIM 0.945). A Residual Diffusion Denoising Model (RDDM) is used to diversify samples, reducing the FID from 69.6 to 24.92. For detection, a lightweight enhanced architecture, SLF-YOLO, achieves a precision of 0.881 and mAP@0.5 of 0.813, outperforming multiple YOLO baselines. A dual defense mechanism—input perturbation suppression and output anomaly analysis—effectively mitigates over 95% of mAP loss under Alpha channel attacks. Experimental results demonstrate that the proposed system delivers robust, secure, and efficient performance, offering a practical pathway toward reliable, interpretable, and resilient industrial vision inspection.