Perspective and Attribute-Based Self-Supervised Conditional GAN for Few-Shot Bolt Defect Image Generation and Classification

The scarcity of bolt defect samples in transmission lines presents significant challenges for classification and detection tasks. To address this issue, this paper proposes a Perspective and Attribute-based Self-supervised Conditional Generative Adversarial Network, aiming to enhance the generative capability of few-shot bolt defect images. First, a label generator based on Contrastive Language-Image Pre-training is proposed for automatic bolt perspective division, which autonomously assigns perspective labels to a bolt dataset to alleviate the issue of visual discrepancies. Then, the perspective and attribute information as conditional guides in the generator is introduced to increase the controllability of generation, and a Residual Skip-Layer Excitation is proposed to enhance the gradient information flow. Finally, the U-Net Self-supervised Reconstruction Discriminator is proposed, and a unique local cropping technique named Shape-crop is designed specifically for T -shaped bolts, thereby enhancing the bolts' feature extraction capability and texture detail generation ability. We constructed a few-shot bolt defect dataset and conducted experiments on it. The experimental results indicate that the method proposed in this paper can generate high-quality bolt defect images and assist in enhancing the accuracy of bolt defect classification, which validates the effectiveness of the methodology presented herein.