Enhanced Photovoltaic Defect Detection Using Perceptual Loss in DCGAN and VGG16-Integrated Models on Electroluminescence Images

Defects in photovoltaic (PV) modules significantly impact their efficiency and reliability, making the accurate detection essential for quality control. This study presents an enhanced generative adversarial network framework, combining deep convolutional generative adversarial network (DCGAN) with visual geometry group 16 (VGG16) and a perceptual loss function, to generate high-quality synthetic defect images and improve defect classification. The proposed model increases classification accuracy from 84% to 90%, demonstrating superior performance over the standard DCGAN. Key improvements include generating more realistic synthetic images, reducing image quality discrepancies, and addressing class imbalances in defect datasets. The enhanced framework performs particularly well in representing rare and complex defects, improving classification outcomes for challenging patterns while maintaining diversity and realism in synthetic data. Despite some limitations in capturing fine defect details, this approach establishes a benchmark for defect detection and synthetic image generation in PV manufacturing. It also offers potential for real-time applications in industrial production lines, ensuring more efficient quality control.