DehazeNet: An End-to-End System for Single Image Haze Removal
read more
Citations
AOD-Net: All-in-One Dehazing Network
Benchmarking Single-Image Dehazing and Beyond
Densely Connected Pyramid Dehazing Network
Deep Joint Rain Detection and Removal from a Single Image
Pre-Trained Image Processing Transformer
References
ImageNet Classification with Deep Convolutional Neural Networks
Gradient-based learning applied to document recognition
Image quality assessment: from error visibility to structural similarity
Going deeper with convolutions
ImageNet Large Scale Visual Recognition Challenge
Related Papers (5)
Frequently Asked Questions (12)
Q2. What have the authors stated for future works in "Dehazenet: an end-to-end system for single image haze removal" ?
The authors leave this problem for future research.
Q3. What is the noise model for NRE?
As a basic noise model, additive white Gaussian (AWG) noise with standard deviation σ ∈ {10, 15, 20, 25} is used for noise robustness evaluation (NRE).
Q4. How can the authors learn haze-relevant features in a deep neural network?
the authors think11atmospheric scattering model can also be learned in a deeper neural network, in which an end-to-end mapping between haze and haze-free images can be optimized directly without the medium transmission estimation.
Q5. How many patches are generated from RF?
According to the testing measure of RF, 2000 image patches are randomly sampled from haze-free images with 10 random transmission t ∈ (0, 1) to generate 20,000 hazy patches for testing.
Q6. Why is the local extremum in DehazeNet?
In addition, the local extremum is in accordance with the assumption that the medium transmission is locally constant, and it is commonly to overcome the noise of transmission estimation.
Q7. What is the important step to recover a haze-free image?
The atmospheric scattering model in Section II-A suggests that estimation of the medium transmission map is the most important step to recover a haze-free image.
Q8. What is the way to evaluate the dehazing performance of different atmosphere airlight?
An airlight robustness evaluation (ARE) is proposed to analyze the dehazing methods for different atmosphere airlight α. Although DehazeNet is trained from the samples generated by setting α = 1, it also achieves the greater robustness on the other values of atmosphere airlight.
Q9. What is the state-of-the-art score for DehazeNet?
DehazeNetachieves the best state-of-the-art score, which is 1.19e-2; the MSE between their method and the next state-of-art result (RF [17]) in the literature is 0.07e-2.
Q10. What is the haze-relevant feature extracted from the image?
To address the ill-posed nature of image dehazing problem, existing methods propose various assumptions and based on these assumptions, they are able to extract haze-relevant features (e.g., dark channel, hue disparity, and color attenuation) densely over the image domain.
Q11. What is the reason why the CAP and RF avoid color distortion in the sky?
Based on the learning framework, CAP and RF avoid color distortion in the sky, but non-sky regions are enhanced poorly because of the non-content regression model (for example, the rock-soil of the first image and the green flatlands in the third image).
Q12. How does DehazeNet achieve the dehazing effects?
With this lightweight architecture, DehazeNet achieves dramatically high efficiency and outstanding dehazing effects than the state-of-the-art methods.