GridDehazeNet+: An Enhanced Multi-Scale Network With Intra-Task Knowledge Transfer for Single Image Dehazing

Adverse weather conditions such as haze can deteriorate the performance of autonomous driving and intelligent transport systems. As a potential remedy, we propose an enhanced multi-scale network, dubbed GridDehazeNet+, for single image dehazing. The proposed dehazing method does not rely on Atmosphere Scattering Model (ASM), explanation to why it is necessarily performing dimension reduction offered by this model provided. GridDehazeNet+ consists three modules: pre-processing, backbone, post-processing. trainable pre-processing module generate learned inputs with better diversity more pertinent features compared those derived produced hand-selected methods. backbone implements estimation two major enhancements: 1) novel grid structure that effectively alleviates bottleneck issue via dense connections across different scales; 2) spatial-channel attention block facilitate adaptive fusion consolidating dehazing-relevant features. post-processing helps reduce artifacts in final output. Due domain shift, trained synthetic data may generalize well real data. To address issue, shape distribution match data, use resulting translated finetune our network. We also intra-task knowledge transfer mechanism memorize take advantage assist learning process Experimental results demonstrate outperforms state-of-the-art several datasets, achieves superior real-world hazy images after finetuning.