GridDehazeNet: Attention-Based Multi-Scale Network for Image Dehazing (1908.03245v1)

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.

• The paper presents an end-to-end trainable CNN that uses attention-based multi-scale estimation within a grid network to achieve effective image dehazing.
• It bypasses the traditional atmosphere scattering model, avoiding its limitations and reducing artifacts in the dehazed outputs.
• Experiments demonstrate that GridDehazeNet consistently outperforms state-of-the-art methods by significantly boosting PSNR and SSIM on both synthetic and real-world images.

An Overview of GridDehazeNet: Attention-Based Multi-Scale Network for Image Dehazing

The paper "GridDehazeNet: Attention-Based Multi-Scale Network for Image Dehazing" introduces an end-to-end trainable Convolutional Neural Network (CNN) specifically designed for single image dehazing. The goal of image dehazing is to recover clear versions of hazy images, a task crucial for intelligent surveillance systems.

Key Contributions

GridDehazeNet Architecture: The proposed GridDehazeNet consists of three main modules: pre-processing, backbone, and post-processing. The pre-processing module generates multiple learned inputs through a trainable mechanism, providing more relevant features than traditional hand-crafted methods. The backbone employs a novel attention-based multi-scale estimation on a grid network to alleviate bottleneck issues in conventional multi-scale approaches. The post-processing module effectively reduces artifacts, enhancing the final output quality.

Independence from Atmosphere Scattering Model: Unlike many existing methods that rely on the atmosphere scattering model for haze removal, GridDehazeNet operates independently of this model. The paper argues that bypassing this model avoids potential pitfalls in performance, even on synthetic images.

Numerical Results

The experimental results are noteworthy, as GridDehazeNet consistently surpasses state-of-the-art methods on both synthetic and real-world datasets. The network achieves significant improvements in PSNR and SSIM on the SOTS dataset, confirming its superior performance in comparison to predecessors like DehazeNet and MSCNN.

Implications and Future Directions

The implications of this research extend beyond image dehazing. The architecture and components of GridDehazeNet are generic enough to potentially benefit other image restoration tasks. The findings challenge the conventional wisdom of employing physical models like the atmosphere scattering model, suggesting a reevaluation of their roles in algorithm design.

The introduction of an attention-based grid network and trainable modules highlights future prospects in enhancing CNN architectures for various computer vision applications. The focus on reducing reliance on physical models could steer new methodologies in data-driven approaches across broader domains.

In summary, GridDehazeNet presents a solid advancement in image dehazing technology, with methodological innovations that may influence future research in image processing and restoration. The exploration into bypassing traditional models offers intriguing avenues for future inquiry and development within the field of AI.