DRACO-DehazeNet: An Efficient Image Dehazing Network Combining Detail Recovery and a Novel Contrastive Learning Paradigm

TL;DR

DRACO-DehazeNet tackles the data- and compute-hungry nature of contemporary image dehazing by fusing dense dilated inverted residual blocks with an attention-based detail recovery network and a quadruplet contrastive learning framework. The architecture introduces DDIRB and ATTDRN to efficiently extract and refine hazy features, while the quadruplet loss guides intermediate dehazed outputs toward ground-truth quality using both published and intermediate representations. Empirical results on RESIDE, O-HAZE, NH-HAZE, and DENSE-HAZE demonstrate superior or competitive PSNR/SSIM with significantly reduced FLOPs compared to related approaches like DPE-Net, highlighting strong performance under limited data and non-uniform haze. The work underscores the practical potential for real-time dehazing on resource-constrained platforms, such as mobile devices and autonomous systems, by delivering high-quality restoration with low computational overhead.

Abstract

Image dehazing is crucial for clarifying images obscured by haze or fog, but current learning-based approaches is dependent on large volumes of training data and hence consumed significant computational power. Additionally, their performance is often inadequate under non-uniform or heavy haze. To address these challenges, we developed the Detail Recovery And Contrastive DehazeNet, which facilitates efficient and effective dehazing via a dense dilated inverted residual block and an attention-based detail recovery network that tailors enhancements to specific dehazed scene contexts. A major innovation is its ability to train effectively with limited data, achieved through a novel quadruplet loss-based contrastive dehazing paradigm. This approach distinctly separates hazy and clear image features while also distinguish lower-quality and higher-quality dehazed images obtained from each sub-modules of our network, thereby refining the dehazing process to a larger extent. Extensive tests on a variety of benchmarked haze datasets demonstrated the superiority of our approach. The code repository for this work is available at https://github.com/GreedYLearner1146/DRACO-DehazeNet.

Figures (11)

• Figure 1: Comparison of the ordinary ResNet block (top) with the inverted ResNet block (bottom). The image is adapted from Ding et al. ding2022slimyolov4.
• Figure 2: Illustration of the gridding problem that arose when using the same vs different dilation rates. The purpose of the dilated convolution is to output a big receptive field encompassing the entire image for a given center pixel, as shown on the left of the figure. If we utilize the same dilation rate (2) throughout, we can also see a uniform grid of pixels which does not contribute to the output center pixel, illustrating the unwanted gridding problem. Conversely, if we were to use differing dilation rate (1,2,4) as in the right of the figure, pixels that are closer to the centre pixel to contribute even more than that of those further away, hence reducing the checkerboard receptive field issue. Thefigure is adapted from the github page "Dilation Rate Gridding Problem and How to Solve It With the Fibonacci Sequence" (https://github.com/Jonas1312/dilation-rate-as-fibonacci-sequence)
• Figure 3: Illustration of the overall architecture of our DRACO-DehazeNet. Once again, C means the number of channels, K means the kernel size, and D means the dilation rate. All strides used is of value 1. The component of the feature extraction module are also illustrated for completeness.
• Figure 4: Illustration of the Densely Connected Dilated Residual Block (DDIRB), along with the ATTention-based Detail Recovery Network (ATTDRN) component of our DRACO-DehazeNet. The Squeeze and Excite (SE) block component is also illustrated. In each individual block of the components, C represents the number of channels, K represents the kernel size, and D represents the dilation rate. All strides used is of value 1.
• Figure 5: Illustration of the triplet network-based contrastive paradigm (as laid out by the AECR-Net) on the anchor (dehazed), positive (GT) and negative (hazy) image sample.