Arc Light Filtering Method for Welding Images Based on a Conditional Diffusion Model

Industrial robots are widely used in automated welding, where vision-based monitoring is crucial for stable and safe operation. However, the intense welding arc light severely interferes with visual systems, compromising the safety and stability of the process. Traditional filtering methods and Generative Adversarial Networks (GANs) struggle with arc light due to complex morphology, loss of texture detail, and training instability. This paper proposes an arc light filtering method for welding images based on a Conditional Diffusion Model. The method introduces a region mask fusion strategy and employs a novel two-stage restoration process—first, repairing the halo region, and then, reconstructing the saturation region. This approach successfully addresses the challenges of structure reconstruction and texture detail loss during arc light filtering. Quantitative evaluation on a self-constructed welding image dataset demonstrates that the proposed method significantly outperforms existing mainstream methods, including Pix2Pix, UNet++P2P, and DifferP2P, across all key metrics. The method achieves a Peak Signal-to-Noise Ratio (PSNR) of \(39.32 \text { dB}\) , yielding an improvement of \(+1.25 \text { dB}\) compared to DifferP2P ( \(38.07 \text { dB}\) ). The Structural Similarity Index (SSIM) and LPIPS metrics reach 0.976 and 0.0102, respectively, outperforming all baselines.