Deep Learning Defect Detection Based on Illumination Correction and Image Simulation

In industrial scenarios, defect detection is often challenged by complex lighting conditions that lead to illumination distortion in captured images, and significant variations in surface defect features. This paper proposes a defect detection method that combines Retinex-based illumination component correction and single-image High Dynamic Range (HDR) simulation with deep learning. Initially, the illumination component is corrected based on Retinex theory and single-image HDR simulation is performed to eliminate local overexposure and expand dynamic range. The images and labels are then fed into a deep learning network for training. The model parameters are adjusted according to defect characteristics to achieve convergence, resulting in detection outcomes. Experimental results demonstrate that the proposed method effectively mitigates the impact of illumination distortion on defect detection, achieving an accuracy rate of 74.6%.