Multi-Stage Fusion Dehazing Network Based on Multi-Resolution

In recent years, the haze phenomenon has become more frequent and significantly impacts high-level vision tasks. Image dehazing has become an increasingly important technique with significant research value. The traditional dehazing algorithms are image enhancement-based dehazing algorithms and image restoration-based dehazing algorithms. However, due to the random distribution of haze, these methods do not recover high-frequency details well. Therefore, to solve the existing problems, in this paper, we have proposed a multi-stage fusion dehazing network based on multi-resolution. Iteratively stacking Residual-based Atrous Spatial Pyramid Convolution (RASPC) and down-sampling convolution layers according to the hazy input image to produce primary feature maps with various resolutions. Then, the adjacent feature maps are fused, and finally, global aggregation is performed to output haze-free images. We compared our method to many popular state-of-the-art approaches on public RESIDE dataset. The experiments demonstrate that our approach can successfully recreate high-frequency details and high-fidelity colors.