Enhancing Defective Solar Panel Detection with Attention-Guided Statistical Features Using Pre-Trained Neural Networks

For defective solar panel detection, the use of resource-depleting methods such as end-to-end deep learning models does not serve the purpose of sustainable green energy. A recent study shows how this problem could be mitigated by exploiting attention-guided statistical features from an MNIST pre-trained attention map while achieving accurate defect detection of solar panels. However, the performance evaluation on attention mechanisms obtained from different training datasets and neural network models has never been reported. This work compares the defect detection performance of attention-guided statistical features from different pre-trained attention mechanisms. We have confirmed that the characteristics of attention mechanisms vary depending on the training dataset and neural network structure, with a stronger reliance on the training dataset. In addition, we present a method, dubbed Attention-Guided Dual Masking (AGDM), to ensure reliable performance regardless of attention mechanism characteristics. AGDM utilizes two disjoint masks not to miss out defective information by complementing each other. Extensive experimental results on the ELPV dataset show that AGDM generalizes the attention-utilizing defect detection models, leading to better performance and reliability.