Enhanced Firefly Optimization with Lévy–Opposition Learning for Dilated–Residual U-Net Based PV Defect Segmentation

Abstract Accurate segmentation of photovoltaic (PV) defects is essential for automated inspection, performance monitoring, and long-term reliability of solar modules. This paper presents a novel defect segmentation framework that integrates a Dilated Residual U-Net with a Lévy–Opposition Learning enhanced Firefly Algorithm (FA-LOL) for hyperparameter optimization. The proposed architecture combines dilated convolutions to capture long-range spatial dependencies with residual connections that stabilize gradient flow and enhance deep feature representation. FA-LOL introduces Lévy-flight exploration and opposition-based diversity, enabling efficient navigation of the hyperparameter search space and preventing premature convergence. Experiments were conducted on two publicly available datasets: the PVEL-AD electroluminescence dataset and the Solar Cell Defect Segmentation dataset. The proposed method achieved a Dice score of 0.9353 and an IoU of 0.88 , outperforming vanilla U-Net, Residual U-Net, Dilated U-Net, and multiple optimization baselines including standard Firefly Algorithm, PSO, GWO, and ACO. Ablation studies confirm that both dilation and residual pathways significantly contribute to the observed improvements, while FA-LOL provides measurable optimization gains. These results demonstrate that the proposed FA-LOL optimized Dilated Residual U-Net is a reliable and efficient solution for high-precision PV defect segmentation across diverse imaging modalities.