Attention-Based SegNet: Toward Refined Semantic Segmentation of PV Modules Defects

Abstract

Proper surveillance and maintenance of photovoltaic (PV) systems are crucial to ensure continuous power generation and prevent operational downtimes. However, manual analysis of electroluminescence (EL) images is subjective, time-intensive, and requires significant expertise. To address this issue, a comprehensive deep learning architecture has been developed for the semantic segmentation of 29 different features and defects within EL images of PV panels. The SegNet architecture encoder has been replaced with the VGG16 encoder, which incorporates pre-trained weights to leverage transfer learning during the feature extraction stage. A Convolutional Block Attention Module (CBAM) block has also been introduced to enhance the decoder’s ability to generate fine-grained segmentations. Additionally, the suggested architecture has been evaluated through the application of three different loss functions: weighted categorical cross-entropy loss, categorical cross-entropy, and focal loss. The Attention-Based SegNet architecture proposed with a weighted categorical cross-entropy loss exhibits superior performance in terms of accuracy, F1 score, intersection over union (IoU), precision, recall, mean IoU (mIoU), specificity, Jaccard index, and Dice coefficient. It achieves a Dice coefficient of 0.9408 and an mIoU of 0.9101, outperforming the state-of-the-art SEiPV-Net trained on the same dataset by 8.77% and 4.97%, respectively.