[Paper Review] Multi-Scale Boosted Dehazing Network with Dense Feature Fusion

Hang Dong, Jinshan Pan+5 more|arXiv (Cornell University)|Apr 28, 2020

Image Enhancement Techniques59 references63 citations

TL;DR

The paper presents MSBDN-DFF, a U-Net based dehazing network that uses SOS boosting and a back-projection inspired dense feature fusion to outperform state-of-the-art methods on multiple hazy-image benchmarks.

ABSTRACT

In this paper, we propose a Multi-Scale Boosted Dehazing Network with Dense Feature Fusion based on the U-Net architecture. The proposed method is designed based on two principles, boosting and error feedback, and we show that they are suitable for the dehazing problem. By incorporating the Strengthen-Operate-Subtract boosting strategy in the decoder of the proposed model, we develop a simple yet effective boosted decoder to progressively restore the haze-free image. To address the issue of preserving spatial information in the U-Net architecture, we design a dense feature fusion module using the back-projection feedback scheme. We show that the dense feature fusion module can simultaneously remedy the missing spatial information from high-resolution features and exploit the non-adjacent features. Extensive evaluations demonstrate that the proposed model performs favorably against the state-of-the-art approaches on the benchmark datasets as well as real-world hazy images.

Motivation & Objective

• Motivate robust dehazing for diverse scenes where priors are unreliable.
• Design a network that combines boosting with error feedback for progressive haze removal.
• Develop a dense feature fusion module to recover spatial details across scales.

Proposed method

• Incorporate Strengthen-Operate-Subtract (SOS) boosting in the decoder to progressively restore haze-free images.
• Introduce a Dense Feature Fusion (DFF) module based on back-projection to fuse multi-scale features and recover spatial information.
• Utilize an encoder–decoder (U-Net) architecture with skip connections and residual groups for feature restoration.
• Train MSBDN-DFF end-to-end with a mean squared error loss.
• Evaluate on RESIDE, HazeRD, and NTIRE2018-dehazing datasets and compare to state-of-the-art methods.

Experimental results

Research questions

• RQ1 Does SOS-based boosting in the decoder improve dehazing quality compared to non-boosted decoders?
• RQ2 Can dense back-projection based feature fusion effectively combine multi-scale information to preserve spatial details in hazy images?
• RQ3 How does MSBDN-DFF perform relative to other end-to-end dehazing networks on standard benchmarks?
• RQ4 What is the impact of the DFF module on preserving high-frequency details and non-adjacent feature utilization?

Key findings

• MSBDN-DFF outperforms several state-of-the-art dehazing methods on benchmark datasets according to qualitative and quantitative evaluations.
• The SOS boosted decoder provides substantial improvements over baselines and alternative boosting strategies.
• The DFF module effectively preserves spatial information and exploits non-adjacent features, yielding higher restoration quality.
• Ablation studies show that both the boosting and the dense feature fusion contribute significantly to performance gains.
• Perceptual evaluation on a KITTI-based haze dataset indicates improved object detection performance on dehazed images.