Multi-layer feature fusion-based data augmentation of defect images

The occurrence of defects is a low-probability event in industrial production, and the number of collected samples is often difficult to meet the training requirements of deep learning-based defect detection. So, data augmentation has become one of the research directions to address few-shot or low-shot problems. First, Deep Convolution Generative Adversarial Networks (DCGAN) is introduced to expand the training set with generated defect images that have better clarity, realism, and diversity. Then, the DCGAN+ model is proposed to solve a remaining problem: the poor quality of the generated large-sized defect images. A multi-layer feature fusion mechanism is designed in this model to reduce feature loss during image upscaling. The fusion layer for multiple feature maps in the Discriminator works collaboratively with the Generator of additional layers to generate higher resolution defect images of large size. At the same time, the skip connections in the Discriminator further enhance the utilization of defect features and accelerate the convergence speed of network training. By adding the defect images generated by the DCGAN+ to the training set, the mAP of defect detection was increased by 5.3%.