Multi-scale Dehazing Network Based on Error-backward Mechanism

The existing dehazing methods can not estimate large-scale target features accurately due to the loss of spatial context information, leading to the destruction of image structure and remaining of haze. To solve this problem, we propose a novel multi-scale dehazing network based on error-backward mechanism, which is composed of error-backward multi-scale dehazing group (EMDG), gated fusion module (GFM) and optimization module. Error-backward multi-scale dehazing group consists of error-backward block (EB) and haze aware unit (HAU). With comparing the difference between feature maps of the coarse-scale sub-network and those of the fine-scale sub-network, error-backward block produces an error map and then transmits it to the last coarse-scale sub-network. So the structure and context information can be reused effectively. Haze aware unit is the core of all sub-networks, which consists of residual dense blocks (RDB) and haze density adaptive detection block (HDADB). It helps to extract local information and accomplish adaptive dehazing according to haze density. Differently from the existing fusion-based methods stacking features from different scales directly, the proposed gated fusion module learns the optimal weights of feature maps from different sub-networks, which prevents interferences to destroy image structure and details. The output of optimization module will be the final dehazed image. Extensive experiments on synthetic datasets and real datasets validate the superiority of our proposed network, especially for the haze removal at a distant view. © 2023 Science Press. All rights reserved.