DehazeNet: An End-to-End System for Single Image Haze Removal (1601.07661v2)

Abstract: Single image haze removal is a challenging ill-posed problem. Existing methods use various constraints/priors to get plausible dehazing solutions. The key to achieve haze removal is to estimate a medium transmission map for an input hazy image. In this paper, we propose a trainable end-to-end system called DehazeNet, for medium transmission estimation. DehazeNet takes a hazy image as input, and outputs its medium transmission map that is subsequently used to recover a haze-free image via atmospheric scattering model. DehazeNet adopts Convolutional Neural Networks (CNN) based deep architecture, whose layers are specially designed to embody the established assumptions/priors in image dehazing. Specifically, layers of Maxout units are used for feature extraction, which can generate almost all haze-relevant features. We also propose a novel nonlinear activation function in DehazeNet, called Bilateral Rectified Linear Unit (BReLU), which is able to improve the quality of recovered haze-free image. We establish connections between components of the proposed DehazeNet and those used in existing methods. Experiments on benchmark images show that DehazeNet achieves superior performance over existing methods, yet keeps efficient and easy to use.

• The paper introduces an end-to-end CNN that estimates medium transmission maps to effectively restore haze-free images.
• It employs innovative layers such as Maxout, multi-scale convolutions, and BReLU to overcome limitations of handcrafted priors.
• Performance evaluations on synthetic and real-world datasets demonstrate superior results in MSE, SSIM, and PSNR compared to traditional techniques.

DehazeNet: An End-to-End System for Single Image Haze Removal

The paper "DehazeNet: An End-to-End System for Single Image Haze Removal" by Bolun Cai et al. introduces a convolutional neural network (CNN) designed to tackle the challenging problem of haze removal from single images. The key innovation of this work lies in its end-to-end learning approach, which deviates from traditional methods relying heavily on handcrafted features and prior assumptions.

The Problem and Traditional Approaches

Haze removal is a well-known problem in the field of image processing and computer vision. It is characterized by the degradation of image quality due to particles such as dust and smoke, which scatter light and reduce visibility. Traditional methods for dehazing generally rely on various priors and constraints. Notable among these are Dark Channel Prior (DCP), Contrast Maximization, and Color Attenuation. However, these methods often falter in scenarios where their underlying assumptions do not hold, such as in the presence of white objects or bright skies.

DehazeNet Architecture

DehazeNet, as proposed in the paper, is a CNN-based framework specifically designed to estimate the medium transmission map of a hazy image. This transmission map is subsequently used to recover a haze-free image via the atmospheric scattering model. The network's architecture includes several unique components:

1. Feature Extraction Layer: Utilizing Maxout units, this layer extracts haze-relevant features from the hazy input images. Maxout units serve as a dimensionality reduction mechanism, capturing essential details while discarding noise.
2. Multi-scale Mapping Layer: This layer employs convolutions with varying filter sizes to capture features at different scales, enhancing the network's robustness to changes in image content and scale.
3. Local Extremum Layer: Implemented using max-pooling operations, this layer is designed to ensure spatial invariance and robustness, particularly effective in handling local variations.
4. Non-linear Regression Layer: The Bilateral Rectified Linear Unit (BReLU) is introduced here, extending the traditional ReLU by imposing bilateral bounds, thus improving convergence and ensuring that the output values remain within plausible limits.

Performance Evaluation

DehazeNet was evaluated on both synthetic and real-world datasets. The network outperformed traditional methods and state-of-the-art techniques on several metrics, including Mean Squared Error (MSE), Structural Similarity Index (SSIM), and Peak Signal-to-Noise Ratio (PSNR). The results demonstrate that DehazeNet significantly reduces the haze effect while preserving image details, even in complex scenes with sky regions or white objects.

Implications and Future Work

The results obtained using DehazeNet suggest notable practical and theoretical implications. Practically, this approach provides a more efficient and accurate alternative to previous methods, particularly in applications requiring real-time processing, given its reduced computational complexity. Theoretically, it opens avenues for further exploration in the domain of image restoration using deep learning models.

Future research could focus on incorporating adaptive mechanisms that dynamically adjust the network parameters based on the type and severity of haze. Additionally, integrating the atmospheric light estimation directly into the network architecture could lead to an even more streamlined and robust dehazing process. Lastly, extending the model to handle other atmospheric conditions, such as fog or underwater haze, could greatly expand its applicability.

Conclusion

In summary, DehazeNet represents a significant leap in the field of single image haze removal. By leveraging convolutional neural networks, it provides an end-to-end learning solution that effectively mitigates the limitations of traditional heuristic-based methods. The proposed architecture and experimental results underscore its potential as a robust tool for dehazing tasks, marking a noteworthy contribution to the field of computer vision.