Efficient Re-parameterization Residual Attention Network For Nonhomogeneous Image Dehazing

TL;DR

This paper introduces ERRA-Net, an efficient re-parameterization residual attention network for nonhomogeneous image dehazing, featuring a novel multi-branch attention block, lightweight design, and specialized loss functions, achieving high speed and competitive performance.

Contribution

The paper presents a novel multi-branch attention block with re-parameterization, a lightweight network architecture, and new loss functions for improved nonhomogeneous image dehazing.

Findings

High processing speed of 166.11 fps on HD images

Outperforms state-of-the-art methods on real-world hazy images

Effective reconstruction of nonhomogeneous hazy images

Abstract

This paper proposes an end-to-end Efficient Re-parameterizationResidual Attention Network(ERRA-Net) to directly restore the nonhomogeneous hazy image. The contribution of this paper mainly has the following three aspects: 1) A novel Multi-branch Attention (MA) block. The spatial attention mechanism better reconstructs high-frequency features, and the channel attention mechanism treats the features of different channels differently. Multi-branch structure dramatically improves the representation ability of the model and can be changed into a single path structure after re-parameterization to speed up the process of inference. Local Residual Connection allows the low-frequency information in the nonhomogeneous area to pass through the block without processing so that the block can focus on detailed features. 2) A lightweight network structure. We use cascaded MA blocks to extract high-frequency features step by step, and the Multi-layer attention fusion tail combines the shallow and deep features of the model to get the residual of the clean image finally. 3)We propose two novel loss functions to help reconstruct the hazy image ColorAttenuation loss and Laplace Pyramid loss. ERRA-Net has an impressive speed, processing 1200x1600 HD quality images with an average runtime of 166.11 fps. Extensive evaluations demonstrate that ERSANet performs favorably against the SOTA approaches on the real-world hazy images.