[Paper Review] GridDehazeNet: Attention-Based Multi-Scale Network for Image Dehazing

Xiaohong Liu, Yongrui Ma+2 more|ArXiv.org|Aug 8, 2019

Image Enhancement Techniques40 references133 citations

TL;DR

GridDehazeNet is an end-to-end CNN for single-image dehazing that uses a trainable pre-processing module, an attention-based multi-scale grid backbone, and a post-processing stage to outperform state-of-the-art methods without relying on the atmosphere scattering model.

ABSTRACT

We propose an end-to-end trainable Convolutional Neural Network (CNN), named GridDehazeNet, for single image dehazing. The GridDehazeNet consists of three modules: pre-processing, backbone, and post-processing. The trainable pre-processing module can generate learned inputs with better diversity and more pertinent features as compared to those derived inputs produced by hand-selected pre-processing methods. The backbone module implements a novel attention-based multi-scale estimation on a grid network, which can effectively alleviate the bottleneck issue often encountered in the conventional multi-scale approach. The post-processing module helps to reduce the artifacts in the final output. Experimental results indicate that the GridDehazeNet outperforms the state-of-the-arts on both synthetic and real-world images. The proposed hazing method does not rely on the atmosphere scattering model, and we provide an explanation as to why it is not necessarily beneficial to take advantage of the dimension reduction offered by the atmosphere scattering model for image dehazing, even if only the dehazing results on synthetic images are concerned.

Motivation & Objective

• Motivate dehazing without strict reliance on the atmosphere scattering model.
• Introduce a trainable pre-processing module to generate diverse learned inputs.
• Develop an attention-based grid backbone to enable efficient multi-scale information exchange.
• Incorporate a post-processing stage to reduce output artifacts.
• Demonstrate state-of-the-art performance on synthetic and real-world hazy images.

Proposed method

• Trainable pre-processing module that converts a hazy image into 16 learned inputs.
• Grid-based backbone with three rows and six columns enabling dense cross-scale information exchange.
• Channel-wise attention to fuse multi-scale features from row/column streams.
• Post-processing module to suppress artifacts in the final dehazed image.
• Loss combines smooth L1 loss with perceptual loss using VGG16 features from Conv1-2, Conv2-2, and Conv3-3 layers with lambda=0.04.

Experimental results

Research questions

• RQ1 Can a fully trainable dehazing network outperform model-based and other data-driven methods on synthetic and real hazy images?
• RQ2 Does a grid-based, attention-guided multi-scale backbone better handle scale bottlenecks than encoder-decoder or standard multi-scale networks?
• RQ3 Is it advantageous to bypass the atmosphere scattering model for dehazing, even on synthetic data?
• RQ4 Do learned pre-processing inputs provide diversity gains over hand-crafted or derived inputs?

Key findings

• Outperforms state-of-the-art methods on SOTS indoor/outdoor and Sun RGB-D datasets in PSNR/SSIM.
• Learned inputs offer diversity gains over derived inputs, boosting performance.
• GridDehazeNet with attention-based fusion yields the best qualitative and quantitative results, including reduced artifacts.
• Bypassing the atmosphere scattering model can yield better optimization surfaces and dehazing performance than indirect model-based estimation.
• Ablation studies show the full model components (grid backbone, attention fusion, post-processing, perceptual loss) each contribute to performance gains.