Research on Dehazing Algorithm Based on Spatial Pyramid Attention

To address the issue that existing image dehazing methods using convolutional neural networks (CNNs) as the backbone fail to utilize the deep detail features of hazy images, leading to suboptimal dehazing performance of trained models, this paper innovatively integrates network structure improvement and a two-stage feature fusion strategy based on a single-image dehazing method using Detail-Enhanced Convolution (DEConv) and Content-Guided Attention (CGA). For network architecture optimization, the improvement involves serially fusing the Spatial Pyramid Attention (SPA) module with the Detail Enhancement Attention Block (DEAB) to construct a hierarchical feature extraction system. The two-stage feature fusion establishes a cascade mechanism of cross-level attention mixing and multi-scale attention enhancement. These methods enhance the model's ability to capture multi-scale structural information and accurately perceive haze concentration distributions.