Improving StyleGAN-ADA with efficient channel attention and multidimensional enhancement heuristic for solder joint defect detection

Generative Adversarial Networks have excelled in generating high-quality images but face challenges when trained on limited datasets due to the discriminator’s tendency to overfit, which compromises the generator’s performance. To address this issue, we propose StyleGAN2-ADAEH, an improved algorithm based on StyleGAN2-ADA. Our method incorporates an attention mechanism into the generator, enhancing its ability to focus on critical input features and improving image realism and detail. In addition, we introduce two heuristic data augmentation strategies in the discriminator, fostering a balanced and competitive dynamic between the generator and discriminator, thereby preventing mode collapse and overfitting. To validate our approach, we developed a new dataset, PCBA-SJD, comprising 2575 images across four types of solder joint defects, and conducted experiments on the PCBA-SJD, FFHQ256-2k, and AFHQ-CAT/WILD datasets. Our method consistently outperforms the latest techniques, achieving a reduced FID score of 15.3 and a KID value of 3.8 on PCBA-SJD, reflecting an improvement of 18.62%. These results demonstrate that our approach improves image synthesis quality under limited data conditions while enhancing training stability.