Multiscale feature enhancement-based dehazing method for photovoltaic panel imagery

Abstract

To address the problems of blurred boundaries, weakened textures, and color degradation of photovoltaic panels in remote sensing images caused by atmospheric scattering from haze and thin clouds, we propose a multi-scale feature enhancement-based dehazing method for photovoltaic panel imagery. The proposed method constructs a Transformer-based encoder–decoder network and introduces structural improvements focused on feature enhancement and model lightweighting. First, a Dynamic Receptive-field Attention Module (DRAM) with spatial adaptability dynamically fuses local and global features to restore target boundaries. Second, a Learnable Adaptive Frequency Node (LAFN) module is designed to preserve luminance distribution and edge texture details in the low-frequency and high-frequency branches respectively. Finally, a Lightweight Context Enhancement Block (LightCEB) is introduced to enhance semantic and detail representation while reducing computational complexity. Experimental results demonstrate that the proposed method outperforms existing approaches in dehazing performance across three datasets, with model parameters and computational cost of only 5.79 M and 0.16 FLOPs, representing reductions of approximately 65% and 73% compared to the baseline, respectively, achieving significant model complexity reduction while maintaining dehazing quality.