Prompt learning-driven unified image restoration for industrial silicon nitride ceramic ball surface inspection

Silicon nitride ceramic balls are critical in aircraft engines and precision bearings, where surface quality governs service life and reliability, requiring robust inspection. Industrial imaging, however, often contains coupled low brightness, blur, and noise, degrading defect detection. We propose DIRNet, a prompt-learning-guided all-in-one restoration network for degraded ceramic ball images. Unlike prompt-guided methods using prompts mainly as implicit degradation cues, DIRNet explicitly decouples degradation and structure via hybrid attention with visual prompts in multi-scale residual interactions, while a lightweight Transformer supports efficient high-resolution processing. We construct SINCB-DID for evaluation. DIRNet achieves average PSNR/SSIM of 34.73 dB/0.910 on SINCB-DID, restores composite degradations stably, and improves ceramic ball surface inspection reliability. Tests on public restoration datasets further indicate stable restoration and generalization beyond the industrial dataset.