Gated context aggregation dehazing algorithm based on detail-enhanced convolution
Jialing He, Liang Zhou · 2025
Single image dehazing is considered as a challenging and poorly defined issue, aimed at restoring the detail information in hazy image. The gated context aggregation network can effectively solve the grid artifact problem during dehazing by introducing smoothed dilated convolution, but there are still issues with local detail processing and color preservation. Therefore, we propose an improved gated context aggregation dehazing algorithm. Firstly, an improved detail-enhanced convolution module is added before the smoothed dilated convolution, which add a gated weight stitching module to highlight the feature enhancement of multiple features, and residual connections are added to prevent information loss. After each smoothed dilated convolution, a united attention module is added, which is composed of channel attention and spatial attention. Finally, besides the original MSE loss, perceptual loss is also added to the loss function. Results of our experiments on RESIDE dataset show the efficiency of the proposed algorithm.