Deep learning for PCB X-ray image generation and restoration

Abstract

This project explores the challenge of limited availability of X-ray PCB detection image datasets and proposes a solution using generation methods to generate X-ray style images as training datasets. The study compares the performance of supervised learning methods such as Generative Adversarial Networks (GANs) and regressive methods such as U-net and Resnet in generating fake Xray images for PCB anomaly detection. The experiments showed that the U-net framework with L1 loss achieved the best results in generating high-quality fake X-ray images. The study also suggests that using SSIM as the final evaluation metric can result in highly consistent evaluation with human judgement. The work provides a novel approach to X-ray data augmentation for PCB anomaly detection and offers insights into the use of regression training for synthesizing high-resolution images. Keywords: X-ray image, PCB, Generation, GAN, U-Net.