Attention-Enhanced Pretrained Networks for Electroluminescence-Based Photovoltaic Cell Defect Classification

Abstract

The rising global deployment of photovoltaic (PV) systems has heightened the necessity for automated defect detection systems that is accurate and computationally efficient. While deep learning offers promising solutions, identifying the optimal configuration of network structures and attention modules remains vital for practical deployment. This paper addresses this gap by proposing the integration of four different attention mechanisms with four different deep learning architectures to identify the optimal architectural configuration that has superior classification performance and is computationally lightweight. The four attention mechanisms evaluated in this study are Frequency Channel Attention (FCA), Gather-Excite (GE), Efficient Multi-Scale Attention (EMA) and Coordinate Attention (CA). The deep learning frameworks tested in this study consists of a custom Convolutional Neural Network (CNN) and three ImageNet-pretrained which are MobileNet-v3 Small, SqueezeNet-1.1 and EfficientNet-B4. Based on the results, it can be concluded that CA integrated with SqueezeNet-1.1 is the best performing model achieving the highest accuracy 94.05%, F1-score 0.9432 and MCC 0.8850 for binary classification while being computationally lightweight with only 0.2679 GFLOPs and 0.8241M parameters.