基于改进BEGAN的钢材缺陷图像数据增强方法OA

Materials science research is trying to develop deep learning-based computer vision methods,but currently limited experimental data is difficult to support the exploration of such big data-based methods.This paper proposes an improved boundary equilibrium generative adversarial network(BEGAN)data augmentation model to address this issue.Firstly,replacing the normalization method in the generator network with spectral normalization reduces the requirement for training sample size compared to batch normalization;Secondly,adding residual modules to the generator/decoder of the model avoids overfitting and accelerates model training;Finally,a self-attention mechanism is added to enhance the model's ability to extract defect details,resulting in smoother and faster convergence of loss parameters during the training process.This paper conducted ablation experiments and comparative experiments using a public dataset of steel defects.Through two evaluation metrics and classification network accuracy,the experiments demonstrated that the improved model significantly outperforms four mainstream generative models.Compared to the BEGAN model's generative dataset,the image classification performance improved by 5.55%;the FID value decreased by 54.35%;the IS value increased by 18.18%.The performance of generated specimens proved this improved method is sufficient as an image enhancement method to cope with small sample problems.