Deep Reinforcement Learning for Multi-Illumination Optimization in Stainless Steel Surface Inspection

Abstract

The inspection of stainless steel surfaces is among the most challenging tasks in industrial quality control, often requiring on-site studies of complex illumination setups. Variations in surface geometry and roughness introduce irregular specular reflections, negatively affecting image quality and the performance of vision-based inspection systems. This paper proposes a Deep Reinforcement Learning (DRL) framework for stainless steel surface inspection that adaptively selects illumination conditions to improve defect segmentation performance. Images were acquired using two high-resolution cameras under a controlled, multi-level illumination setup. Based on this acquisition process, we constructed a multi-illumination dataset of stainless steel production samples collected from an industrial processing line, in which each sample is captured under several predefined illumination levels. Using this dataset, we train a Proximal Policy Optimization (PPO) agent guided by a segmentation-driven reward. The proposed approach is evaluated across two different segmentation architectures and compared against fixed-illumination baselines and alternative DRL strategies. Experimental results demonstrate that the learned policy selects illumination levels that improve segmentation accuracy, achieving gains of 5.5% in mIoU and 6.2% in hIoU. A further ablation study examining action space granularity, reward formulation, and termination criteria demonstrates an additional improvement of 1.7% in mIoU and 3.1% in hIoU.