91‐1: Automated Methods for Panel Defect Image Generation and Assisting Defect Detection

In panel production, defective products are inevitable. By utilizing deep learning models for automated defect detection and classification, defective items can be swiftly filtered out, aiding in the diagnosis of defect causes and enabling faster resolution of production issues. However, training deep learning models to achieve high accuracy typically relies on a substantial volume of samples. In real production line applications, many types of defects lack sufficient samples, which limits model performance. This paper introduces a method that leverages prompts and random noise to generate defect images, allowing the detection model to select training samples from the generated images. These selected samples are then integrated with real samples to optimize the model, thereby enhancing detection performance for rare defect types. The defect images are created using the inpainting method of the Stable Diffusion (SD) model. Multiple mask generation techniques are applied during the generation process to meet the location requirements of different defect types. The generated images exhibit randomness and diversity, and a pre‐trained detection model filters out redundant or irrelevant images. This refined, unique, and varied dataset helps the detection model achieve higher accuracy. For instance, by adding 500 generated samples to the original 6 real samples, the F1 score increases by approximately 5.4%.