This study proposes a comprehensive evaluation framework to objectively assess the performance of generative adversarial networks (GANs) for generating photovoltaic infrared defect images under small-sample conditions. Focusing on realism and diversity, the framework integrates quantitative metrics with semantic-level evaluation. A weighting system is constructed by combining the Entropy Weight Method (EWM) and the Analytic Hierarchy Process (AHP), enabling systematic calculation of comprehensive scores. Five GAN models, namely DCGAN, LSGAN, WGAN, WGAN-GP, and R3GAN-LS, were evaluated, and their generated images were used to augment the datasets for two defect categories (“Hot-Spot-Multi” and “Soiling”). Classification experiments were subsequently conducted by combining these augmented datasets with images from the “No_Anomaly” category. Results indicate that LSGAN achieves the highest comprehensive score of 0.537 and a classification accuracy of 89.61% for the Hot-Spot-Multi task, whereas WGAN-GP performs best for Soiling, with a comprehensive score of 0.655 and an accuracy of 93.62%. To further validate the framework’s generalizability, additional defect types (“Diode-Multi” and “Cell-Multi”) were respectively assigned to the established “Hot-Spot-Multi” and “Soiling” weighting schemes based on their visual characteristics. Different GAN models were then used to generate “Diode-Multi” and “Cell-Multi” defect images, which were used to augment the corresponding datasets in the subsequent classification experiments. Results show strong consistency between comprehensive evaluation scores and actual test accuracies, confirming the robustness and reliability of the proposed framework. These findings highlight the practical applicability of the framework for guiding GAN-based data augmentation in photovoltaic defect classification.
扫码关注我们
求助内容:
应助结果提醒方式:
