MECNN: Mixed-Attention Enhanced Convolutional Neural Network for Remote Sensing Dehazing
Zhenxiao Hua · 2024
Haze in the atmosphere degrades remote sensing image quality, impacting applications like land cover classification and target detection. Traditional dehazing methods relied on heuristic information, limiting image representation. Deep learning has led to more effective dehazing models, but existing methods are inadequate for remote sensing due to haze distribution variations and limited datasets. In order to tackle this issue, a novel end-to-end mixed-attention enhanced convolutional neural network is introduced. The network leverages an attention mechanism to handle varying haze distributions and incorporates wavelet transform for multi-frequency processing. CloudGAN is used to generate training datasets, enhancing model generalization. Comparative analysis on synthetic and real datasets from Landset-8 OLI remote sensing data demonstrates superior dehazing performance while preserving image structure. The network effectively removes haze and retains original image color and texture.