As power producers increasingly adopted renewable energy sources, solar energy emerged as a key solution due to its sustainability and efficient energy harvesting. Photovoltaic (PV) modules convert solar irradiance into electricity, but defects in these modules significantly impact power generation efficiency. This study aims to classify PV defect images through deep learning, focusing on Convolutional Neural Networks (CNNs). It explores two different transfer learning architectures: Series Networks, including AlexNet, VGG-16, and VGG-19, and Directed Acyclic Graph (DAG) Networks, such as ResNet-18, ResNet-50, Inception-V3, and GoogLeNet. Series Networks, known for their structured, layer-wise feature extraction, performed well in specificity and precision; meanwhile, DAG Networks, with deeper architectures and residual connections, enhanced feature learning and classification accuracy. The study evaluated augmented and non-augmented datasets to determine the most effective approach. Results showed that the augmented dataset consistently outperformed the non-augmented dataset. Among the architectures, ResNet-50, a DAG Network, achieved the highest classification accuracy (98.96%) and F1 score (97.89%) due to its ability to retain critical features across multiple layers. Meanwhile, AlexNet excelled in specificity (99.53%), and VGG-16 demonstrated superior precision (98.65%), highlighting the strengths of Series Networks in structured, less complex classification tasks.