Multi-Scale Boosted Dehazing Network with Dense Feature Fusion (2004.13388v1)

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.

• The paper introduces a boosted decoder architecture using a Strengthen-Operate-Subtract strategy to progressively restore haze-free images.
• It presents a dense feature fusion module guided by back-projection techniques to recover lost spatial details in hazy images.
• The model achieves superior PSNR and SSIM scores across standard datasets, demonstrating robust, state-of-the-art performance.

An Overview of Multi-Scale Boosted Dehazing Network with Dense Feature Fusion

The paper under review presents a sophisticated approach to the problem of image dehazing, proposing a Multi-Scale Boosted Dehazing Network (MSBDN) enhanced with Dense Feature Fusion (DFF). As dehazing is a particularly challenging ill-posed problem, effective techniques are required to establish constraints that aid in restoring clear images from hazy ones. This work addresses core challenges by leveraging principles of boosting and error feedback within a U-Net architecture framework.

Key Contributions

1. Boosted Decoder Architecture: The authors extend the U-Net architecture with a boosted decoder that employs a Strengthen-Operate-Subtract (SOS) boosting strategy. This method facilitates progressive restoration of haze-free images by recursively refining intermediate outputs.
2. Dense Feature Fusion Module: To overcome limitations in preserving spatial information and exploiting non-adjacent features, a dense feature fusion module, guided by back-projection techniques, is introduced. This fusion mechanism exploits multi-scale features to effectively recover details lost during the compression and degradation processes in hazy images.
3. Comprehensive Evaluation: Extensive experimentation demonstrates that the proposed model surpasses existing state-of-the-art methods across standard datasets, highlighting the practical efficacy and generalizability of the approach in diverse hazy conditions.

Numerical Results and Analysis

The model's performance is rigorously evaluated across several benchmark datasets, including SOTS, HazeRD, I-HAZE, and O-HAZE. Notably, the MSBDN-DFF model achieves superior results, with the highest PSNR and SSIM scores, indicating highly accurate restoration of haze-free images compared to other methods. These results underscore the effectiveness of the proposed dense feature fusion and boosting strategies.

Practical and Theoretical Implications

From a practical perspective, the successful implementation of dense feature fusion and boosting in image restoration tasks opens avenues for improved pre-processing steps in high-level computer vision tasks such as object detection and scene understanding. Theoretically, the integration of error feedback mechanisms such as back-projection into network designs could inform future model architectures geared towards other ill-posed vision problems.

Future Developments

While the authors' work sets a precedent in image dehazing, future research could explore the application of these principles to other domains such as video dehazing and real-time processing. Additionally, expanding the techniques to accommodate synthetic data augmentation or integration into larger vision systems may provide further enhancements.

In summary, this paper enriches the domain of image dehazing with a novel method that combines boosting strategies with cutting-edge feature fusion techniques. The results clearly demonstrate the potential of the proposed methodologies to make significant contributions to related areas in computational imaging and computer vision.