Multi-Scale Boosted Dehazing Network with Dense Feature Fusion

1. Introduction

• The paper frames dehazing as highly ill-posed: small measurement errors can cause dramatic changes, while simply adding depth or width is inefficient for substantial gains.It argues that these challenges make dehazing-specific network design important.Source: §1 · p. 1
• MSBDN integrates boosting and back-projection into a dehazing network, using a Strengthen-Operate-Subtract decoder to progressively restore haze-free images.The authors describe the boosting design as simple and effective in practice, and state that boosting can help under a certain axiom.Source: §1 · p. 2
• Dense Feature Fusion uses back-projection to fuse features across levels, preserving high-resolution spatial information while exploiting non-adjacent features.The module is intended to address spatial information compressed by U-Net encoder downsampling.Source: §1 · p. 2
• The authors report that MSBDN performs favorably against state-of-the-art dehazing methods in extensive evaluations.The introduction does not provide quantitative comparisons.Source: §1 · p. 2

Image dehazing

• Prior-based methods constrain transmission, atmospheric light, and scene radiance, but can fail when their scene or atmospheric assumptions do not hold.For example, the dark channel prior performs poorly on images without zero-intensity pixels.Source: §2 · p. 2
• Deep methods either estimate transmission maps and atmospheric light, risking color distortion from airlight-albedo ambiguity, or directly restore radiance with generic architectures [43] [9] [41] [40].The latter are described as inefficient for dehazing, while estimation-based methods depend on accurate transmission and atmospheric-light estimates [5].Source: §2 · p. 2

Boosting algorithms for image restoration

• Prior boosting methods progressively refine denoising results, and deep CNN boosting has been used for classification [21] and denoising; this work applies boosting to dehazing.Source: §2 · p. 2

Multi-scale feature fusion

• Existing fusion approaches reuse same-scale features, concatenate projected multi-level features [25] [37], or connect adjacent levels in grid architectures [40], but do not explicitly exploit non-adjacent levels.Unlike the two-level feedback method [22], the proposed DFF module fuses features from multiple scales.Source: §2 · p. 2

3.1. Multi-Scale Boosted Dehazing Network

• The network builds on U-Net and places SOS boosting in a multi-scale decoder that strengthens upsampled features with encoder features before learned refinement.Each refinement unit is implemented as a residual group, and a final convolution reconstructs the haze-free image.Source: §3.1 · pp. 2–3
• Under Axiom 1—that less-hazy same-scene images remain better in PoH after dehazing—the SOS iterations reduce the estimated image's PoH.PoH is defined as (1 − T)A/J and is proportional to 1 − T for hazy images of the same scene.Source: §3.1 · p. 3

3.2. Dense Feature Fusion Module

• DFF modules in the encoder and decoder progressively fuse preceding levels through learned down/up-sampling and back-projected error feedback, addressing lost spatial detail and exploiting non-adjacent features.The decoder DFF enhances each boosted feature using preceding enhanced features, and the feedback mechanism is described as recovering high-frequency information from high-resolution features.Source: §3.2 · pp. 4–5, Figure 3 · p. 5

3.3. Implementations

• Implementation uses four strided convolutional and deconvolutional layers, LReLU with slope 0.2, and jointly trains MSBDN and DFF with MSE for 100 ADAM epochs.Training uses batch size 16, β1 = 0.9, β2 = 0.999, and an initial learning rate of 10^-4 decayed by 0.75 every 10 epochs.Source: §3.3 · p. 5

4.1. Datasets

• The evaluation uses RESIDE training data for SOTS, NYUv2 training for HazeRD, and each NTIRE challenge dataset's own training set and corresponding test set.RESIDE training uses 9,000 outdoor and 7,000 indoor pairs; several comparison methods are retrained on the same training data for fair evaluation.Source: §4.1 · pp. 5–6, §4.2 · p. 6

4.2. Performance Evaluation

• On SOTS, the proposed method outperforms the compared algorithms in PSNR and SSIM; on HazeRD it achieves better CIEDE2000 and SSIM, and on NTIRE18 it is comparable to MsPPN.The NTIRE18 comparison is against MsPPN, which uses extra pre- and post-processing for ultra-high-resolution datasets.Source: Table 1 · p. 6, §4.2 · p. 6
• The proposed method produces clearer SOTS results and a sharper, brighter real-image dehazing result than the compared approaches.The SOTS examples show color distortions or residual artifacts in other methods, while the real-image example is clearer with less color distortion.Source: §4.2 · p. 6, Figure 5 · p. 7, Figure 6 · p. 7
• On the synthesized KITTI Haze dataset, dehazed images from the proposed method yield the highest YOLOv3 detection accuracy among the evaluated methods.The evaluation synthesizes haze from KITTI detection images using estimated depth maps and RESIDE protocols; Figure 4 also shows improved recognition of cars and pedestrians.Source: §4.2 · p. 6, Table 2 · p. 7, Figure 4 · p. 7

4.3. Ablation Study and Analysis

• Larger feature-level counts and more ResBlocks improve performance, and SOS boosting with DFF yields gains across configurations; the default is L = 5 and B = 18.The study varies feature levels L and ResBlocks B in GRes on SOTS.Source: Table 3 · p. 7, §4.3 · p. 7
• The SOS boosted decoder outperforms the other tested boosting strategies by at least 0.36 dB without adding extra layers.Boosting modules improve on the PFFNet baseline, while the simple Diffusion Decoder performs worst.Source: Table 4 · p. 8, §4.3 · p. 8
• MSBDN-DFF outperforms MSBDN-M by 0.55 dB without extra layers, while PFFNet-DFF improves on PFFNet by 3.73 dB.MSBDN-S is 0.17 dB higher in average PSNR than MSBDN, showing a benefit from fusing non-adjacent preceding features.Source: §4.3 · p. 8

• Extensive evaluations show that MSBDN performs favorably against state-of-the-art methods on image dehazing datasets.Source: §5 · p. 8
• The encoder-decoder network uses an SOS-based boosted decoder to restore hazy images.Source: §5 · p. 8
• Its back-projection-based DFF module preserves spatial information and exploits features from non-adjacent levels.Source: §5 · p. 8
• Ablation studies demonstrate that the proposed modules are effective for image dehazing.Source: §5 · p. 8