With the advancement of intelligent manufacturing, automated inspection of welding quality has become a vital direction in industrial vision research. To address the limitations of traditional methods in complex working conditions, this paper proposes a deep learning-based detection framework tailored to industrial welding scenarios, incorporating systematic image enhancement and adversarial defense mechanisms to improve model robustness and security. In the image preprocessing stage, the SCUNet denoising model is adopted to significantly improve image clarity. To address uneven illumination, a DarkIR module based on Retinex theory is introduced, effectively enhancing local texture and contrast in low-light images. For target detection, we construct a YOLOv11 model, which achieves efficient detection of small welding defects through multi-scale feature fusion and attention mechanisms. To tackle the vulnerability of deep models to adversarial attacks, this study designs an Alpha channel attack method and proposes defense strategies including channel purification, simulated perturbation enhancement, and channel attention mechanisms, significantly enhancing system robustness against adversarial samples. Experimental results validate the effectiveness and practicality of the proposed method in improving the accuracy and safety of welding quality inspection, providing technical support for the engineering deployment of industrial vision systems.