Data augmentation of GMAW weld defect dataset based on improved deep convolution generative adversarial network

Imbalanced data distribution stands as the primary cause of performance deterioration in the majority of supervised classification algorithms. The current publicly available weld defect datasets are very limited, and the samples of various defects are seriously imbalanced. The paper proposes an improved deep convolution generative adversarial network(DCGAN) to balance the weld defect dataset. To solve the problem of poor diversity in the samples generated by the traditional DCGAN, a C-Res unit is constructed, which integrates the convolutional block attention module(CBAM) into the residual block. The transposed convolution in the DCGAN's generator is replaced with the constructed C-Res unit to enhance the attention to image details and improve the stability and learning efficiency of the model. The Pixelshuffle module is added into the generator as the upsampling module to solve the problem that the C-Res unit can't up-sample like the transposed convolution. CBAM is added into the DCGAN's discriminator to further enhance the discriminator's ability to judge the quality of the generated sample. To validate the effectiveness of the improved DCGAN, comparison experiments are carried out. The weld defect dataset is balanced by DCGAN and improved DCGAN, respectively, and then YOLOv8s-cls is used to classify the weld defect sample based on the original dataset, the dataset balanced by the DCGAN, and the dataset balanced by the improved DCGAN, respectively. Among the nine F1 scores of the nine types of samples, seven of them are higher than those of YOLOv8s-cls trained with the original dataset, and six of them are higher than those of YOLOv8s-cls trained with the dataset balanced by the traditional DCGAN. The experiments reveal that the weld defect dataset balanced with improved DCGAN can enhance the performance of the supervised classification model, and is helpful to realize automation of weld defect detection.