MC-UVCGAN: Cycle-Consistent GAN with Multi-Scale Channel Block for Generating Defective Photovoltaic Module Images

With the rapid growth of photovoltaic (PV) installations, efficient inspection based on unmanned aerial vehicle (UAV) is crucial for maintaining PV power stations. To address the need for constructing datasets suitable for high-altitude PV module defects inspection based on deep learning, we propose the UNet Vision Transformer Cycle-Consistent Generative Adversarial Network with Multi-Scale Channel Block (MC-UVCGAN), a model designed to generate diverse and high-quality defective PV module images. The generator of our proposed model integrates the strengths of UNet and Transformers to enhance the preservation of image details while maintaining cycle consistency. The MC Block, a novel attention-based module, is introduced into the generator's encoding and decoding processes to efficiently capture correlations between defective image channels and spatial features across multiple scales. The proposed model is evaluated across three defective types of generation tasks, demonstrating superior performance compared to previous Generative Adversarial Network (GAN)-based models.