Back

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

Bolun CaiXiangmin XuKui JiaChunmei QingDacheng Tao

article2016IEEE TIP3,111 citations

Proposes DehazeNet, an end-to-end convolutional neural network architecture that integrates domain-specific dehazing priors with Maxout units and a Bilateral Rectified Linear Unit activation function to accurately estimate medium transmission maps for single-image restoration.

Listen

DehazeNet is a convolutional neural network designed to estimate the medium transmission map directly from a single hazy input image, after which the haze-free output is recovered through the standard atmospheric scattering model. The work addresses the long-standing difficulty of single-image dehazing, an ill-posed inverse problem in which haze thickness depends on unknown and spatially varying scene depth; conventional methods rely on hand-crafted priors that often produce oversaturated skies, fail on dense haze, or require heavy post-processing.

The authors built a four-stage CNN whose layers explicitly mirror established dehazing assumptions: a Maxout feature-extraction stage that can reproduce dark-channel, contrast, color-attenuation and hue-disparity cues; a multi-scale convolutional mapping; a local-extremum pooling layer that enforces spatial constancy of transmission; and a final regression layer using a newly proposed Bilateral Rectified Linear Unit (BReLU) that constrains outputs to the physically valid interval [0,1]. The network was trained end-to-end on 100,000 synthetic 16×16 patches generated from Internet-collected haze-free images by applying the scattering model with random transmission values; it contains only 8,240 parameters and runs at roughly 1.5 seconds per 640×480 image on a standard CPU.

On held-out synthetic patches DehazeNet reduced mean-squared transmission error to 1.19×10^{-2}, outperforming the next-best learning method (Random Forests) by 0.07×10^{-2} and classical priors by larger margins. Across complete synthetic images it recorded the lowest MSE, highest SSIM, PSNR and weighted SNR under varied haze densities, atmospheric colors, image scales and additive noise. On five challenging real-world photographs containing large sky or white regions, the method avoided the oversaturation and color distortion typical of prior-based approaches while still removing haze from foreground objects.

These gains matter because accurate, automatic dehazing improves visibility for consumer photography, surveillance, and downstream vision tasks without requiring multiple exposures, depth sensors or manual tuning. The architecture is already fast enough for practical use and generalizes across haze conditions better than earlier regression models.

Further work should integrate atmospheric-light estimation inside the same network and explore direct learning of the full scattering model so that transmission need not be recovered as an explicit intermediate. The main limitations are reliance on synthetic training data and the assumption that transmission is locally constant; both introduce possible domain-shift risk on real scenes whose statistics differ markedly from the training distribution.

South China University of TechnologyUniversity of MacauUniversity of Technology Sydneyconvolutional neural networksfeature extractionsingle imagessingle image haze removalhaze transmissionatmospheric light
arXiv: 1601.07661caibolun/DehazeNet
  • Paper: Deep Residual Learning for Image Recognition, Kaiming He et al. (2016). ResNet extends DehazeNet's shallow CNN philosophy by introducing deep residual learning to resolve degradation problems across hundreds of layers.
Cover for DehazeNet: An End-to-End System for Single Image Haze Removal

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.

Topics

convolutional neural networksfeature extractionsingle images

Keywords

single image haze removalhaze transmissionatmospheric lightdark channel featurehaze-relevant featuresvisibility restorationhue disparitycontrast restoration

Table of Contents

  • I. INTRODUCTION
  • II. RELATED WORKS
  • A. Atmospheric Scattering Model
  • B. Haze-relevant features
  • III. THE PROPOSED DEHAZENET
  • A. Layer Designs of DehazeNet
  • B. Connections with Traditional Dehazing Methods
  • C. Training of DehazeNet
  • IV. EXPERIMENTS
  • A. Model and performance
  • B. Filter number and size
  • C. Quantitative results on synthetic patches
  • D. Quantitative results on synthetic images
  • E. Qualitative results on real-world images
  • V. CONCLUSION

Knowls

  1. Knowl 1 — DehazeNet Architecture for Medium Transmission Estimation

    model/method

    DehazeNet is a feed-forward convolutional neural network designed to estimate the medium transmission map t(x)∈[0,1]t(x) \in [0, 1] from an input RGB hazy image patch I(x)∈R3×16×16I(x) \in \mathbb{R}^{3 \times 16 \times 16}. The architecture consists of four sequential operations:

    1. Feature Extraction (F1F_1): A convolutional layer with k×n1=4×16=64k \times n_1 = 4 \times 16 = 64 filters W1i,j∈R3×5×5W_1^{i,j} \in \mathbb{R}^{3 \times 5 \times 5} (stride 1, padding 0) and biases B1i,j∈RB_1^{i,j} \in \mathbb{R}, followed by a Maxout unit taking the pixel-wise maximum across k=4k = 4 feature maps to produce n1=16n_1 = 16 feature maps of spatial size 12×1212 \times 12: F1i(x)=max⁡j∈{1,…,k}(W1i,j∗I(x)+B1i,j),i∈{1,…,n1}F_1^i(x) = \max_{j \in \{1, \dots, k\}} \left( W_1^{i,j} * I(x) + B_1^{i,j} \right), \quad i \in \{1, \dots, n_1\}

    2. Multi-Scale Mapping (F2F_2): A parallel multi-scale convolutional layer that applies n2=48n_2 = 48 filters divided into three equal groups of 16 filters with kernel sizes 3×33 \times 3 (padding 1), 5×55 \times 5 (padding 2), and 7×77 \times 7 (padding 3), operating on F1F_1 to output n2=48n_2 = 48 feature maps of size 12×1212 \times 12: F2i=W2⌈i/3⌉,(i mod 3)∗F1+B2⌈i/3⌉,(i mod 3),i∈{1,…,n2}F_2^i = W_2^{\lceil i/3 \rceil, (i \bmod 3)} * F_1 + B_2^{\lceil i/3 \rceil, (i \bmod 3)}, \quad i \in \{1, \dots, n_2\}

    3. Local Extremum (F3F_3): A dense spatial max-pooling operation over a sliding neighborhood Ω(x)\Omega(x) of size 7×77 \times 7 (stride 1, padding 0), producing n3=48n_3 = 48 feature maps of spatial size 6×66 \times 6: F3i(x)=max⁡y∈Ω(x)F2i(y),i∈{1,…,n3}F_3^i(x) = \max_{y \in \Omega(x)} F_2^i(y), \quad i \in \{1, \dots, n_3\}

    4. Non-linear Regression (F4F_4): A convolutional layer with a single filter W4∈R48×6×6W_4 \in \mathbb{R}^{48 \times 6 \times 6} and bias B4∈RB_4 \in \mathbb{R} (padding 0), followed by a Bilateral Rectified Linear Unit (BReLU) activation with bounds tmin⁡=0t_{\min} = 0 and tmax⁡=1t_{\max} = 1, yielding a single scalar transmission value for the patch: F4=min⁡(tmax⁡,max⁡(tmin⁡,W4∗F3+B4))F_4 = \min\left(t_{\max}, \max\left(t_{\min}, W_4 * F_3 + B_4\right)\right)

    The complete default model contains 8,240 trainable parameters.

  2. Knowl 2 — Bilateral Rectified Linear Unit (BReLU)

    definition

    The Bilateral Rectified Linear Unit (BReLU) is a bounded piecewise-linear activation function designed for image restoration and regression tasks where the target outputs lie within a known bounded range [tmin⁡,tmax⁡][t_{\min}, t_{\max}]. For an input scalar or tensor yy, BReLU is defined as:

    f(y)=min⁡(tmax⁡,max⁡(tmin⁡,y))f(y) = \min\left(t_{\max}, \max(t_{\min}, y)\right)

    For medium transmission estimation in single image dehazing, the marginal values are fixed to tmin⁡=0t_{\min} = 0 and tmax⁡=1t_{\max} = 1. The derivative of BReLU with respect to its input yy is:

    ∂f(y)∂y={1,if tmin⁡≤f(y)<tmax⁡0,otherwise\frac{\partial f(y)}{\partial y} = \begin{cases} 1, & \text{if } t_{\min} \le f(y) < t_{\max} \\ 0, & \text{otherwise} \end{cases}

    BReLU provides two properties:

    • Bilateral restraint: Constrains predictions to the physical interval [tmin⁡,tmax⁡][t_{\min}, t_{\max}], preventing response overflow and invalid negative or super-unitary transmission estimates.
    • Local linearity: Maintains a unit gradient within the operating range [tmin⁡,tmax⁡)[t_{\min}, t_{\max}), mitigating vanishing gradients during backpropagation compared to smooth sigmoidal activations.

  3. Knowl 3 — Single Image Dehazing Pipeline with DehazeNet

    algorithm

    The single image dehazing framework takes an observed hazy RGB image I(x)I(x), estimates its medium transmission map using DehazeNet, refines the map, estimates global atmospheric light, and recovers the haze-free radiance J(x)J(x).

    Input: Hazy RGB image I(x)∈[0,1]3×H×WI(x) \in [0, 1]^{3 \times H \times W}, trained DehazeNet model F\mathcal{F}, Guided Image Filter radius rr and regularization ϵ\epsilon, threshold percentile p=0.001p = 0.001
    Output: Restored haze-free scene radiance J(x)∈[0,1]3×H×WJ(x) \in [0, 1]^{3 \times H \times W}
    1. Extract raw transmission map t~(x)\tilde{t}(x) by applying DehazeNet densely over I(x)I(x)
    2. Compute refined transmission map t(x)t(x) by applying Guided Image Filtering to t~(x)\tilde{t}(x) using I(x)I(x) as the guidance image:
       t(x)=GuidedFilter(t~(x),I(x),r,ϵ)t(x) = \text{GuidedFilter}(\tilde{t}(x), I(x), r, \epsilon)
    3. Identify the lower threshold t0t_0 corresponding to the lowest 0.1%0.1\% percentile of pixel values in t(x)t(x)
    4. Identify candidate pixels with dense haze: Ωt0={x∣t(x)≤t0}\Omega_{t_0} = \{x \mid t(x) \le t_0\}
    5. Estimate global atmospheric light α∈R3\alpha \in \mathbb{R}^3 as the brightest pixel in I(x)I(x) within the candidate region:
       α=max⁡y∈Ωt0I(y)\alpha = \max_{y \in \Omega_{t_0}} I(y)
    6. Recover scene radiance J(x)J(x) pixel-wise via the inverted atmospheric scattering model:
       J(x)=I(x)−α(1−t(x))t(x)J(x) = \frac{I(x) - \alpha (1 - t(x))}{t(x)}
    7. return J(x)J(x)

    The entire pipeline processes a 640×480640 \times 480 pixel image in approximately 1.5 seconds on a 3.4 GHz CPU.

  4. Knowl 4 — Synthetic Patch Generation for Transmission Map Training

    model/method

    Because large collections of natural paired hazy and haze-free images are unavailable, training data for DehazeNet is generated synthetically using the physical atmospheric scattering model:

    IP(x)=JP(x)t+α(1−t)I^P(x) = J^P(x)t + \alpha(1 - t)

    where JP(x)J^P(x) is a clean haze-free image patch, t∈(0,1)t \in (0, 1) is the medium transmission, α\alpha is the global atmospheric light, and IP(x)I^P(x) is the synthesized hazy patch. The synthesis rests on two assumptions:

    1. Content independence: Image content is independent of scene depth and medium transmission.
    2. Local constancy: Pixels within a small patch (16×1616 \times 16) have approximately constant depth, justifying a single uniform scalar transmission tt per patch.

    To minimize parameter uncertainty during training, the atmospheric light is fixed to α=1\alpha = 1. The training dataset is constructed by randomly sampling 10,000 haze-free patches of size 16×1616 \times 16 from natural and urban landscape photographs, and then sampling 10 random transmissions t∼U(0,1)t \sim \mathcal{U}(0, 1) per patch to yield 100,000 synthesized pairs (IiP,ti)(I^P_i, t_i).

    The network parameters Θ={W1,W2,W4,B1,B2,B4}\Theta = \{W_1, W_2, W_4, B_1, B_2, B_4\} are trained by minimizing the Mean Squared Error (MSE) loss:

    L(Θ)=1N∑i=1N∥F(IiP;Θ)−ti∥2L(\Theta) = \frac{1}{N} \sum_{i=1}^N \left\| \mathcal{F}(I^P_i; \Theta) - t_i \right\|^2

    Training uses Stochastic Gradient Descent (SGD) with a batch size of 128 for 500,000 iterations. Weights are initialized from N(0,0.0012)\mathcal{N}(0, 0.001^2) and biases to 0. The learning rate begins at 0.005 and is halved every 100,000 iterations down to 3.125×10−43.125 \times 10^{-4}.

  5. Knowl 5 — Equivalence of DehazeNet Layers to Classical Dehazing Priors

    theoretical result

    The individual layers of DehazeNet functionally generalize established handcrafted image dehazing priors:

    • Dark Channel Prior: When the first-layer convolutional weight W1W_1 acts as an opposite filter (−1-1 at the center of one color channel) with unit bias B1=1B_1 = 1, the Maxout unit across channels computes max⁡(1−Ic(x))=1−min⁡cIc(x)\max(1 - I^c(x)) = 1 - \min_{c} I^c(x), which matches the channel minimum in Dark Channel feature extraction.
    • Maximum Contrast Prior: When W1W_1 learns a round local contrast filter, Maxout over affine projections approximates the local contrast variance operator.
    • Color Attenuation and Hue Disparity Priors: Combining all-pass filters and opposite filters in W1W_1 enables atomic operations for RGB-to-HSV color space transformations, specifically the difference between brightness and saturation A(x)=Iv(x)−Is(x)A(x) = I^v(x) - I^s(x) and semi-inverse hue shifts.
    • Convex Function Approximation: A Maxout unit taking the maximum over k=4k = 4 affine transformations acts as a piece-wise linear approximation to arbitrary convex feature extraction functions.
    • Local Scene Constancy: The 7×77 \times 7 local max-pooling in the third layer enforces the assumption that medium transmission is locally constant, removing local transmission estimation noise over white objects.
    • Boundary Constraints: The BReLU activation in the fourth layer enforces lower and upper transmission bounds (tmin⁡=0,tmax⁡=1t_{\min} = 0, t_{\max} = 1), corresponding to analytical boundary constraints.

  6. Knowl 6 — Transmission Estimation Accuracy on Synthetic Patches

    data/table

    The transmission estimation accuracy of DehazeNet was evaluated against five prior methods on a benchmark of 20,000 synthetic hazy image patches generated by sampling 2,000 haze-free patches with 10 random transmission values t∈(0,1)t \in (0, 1). Accuracy is measured by the Mean Squared Error (MSE) between the predicted transmission and the ground truth transmission.

    Method DCP BPNN CAP RF DehazeNet
    MSE (×10−2\times 10^{-2}) 3.18 4.37 3.32 1.26 1.19

    DehazeNet achieves the lowest MSE (1.19×10−21.19 \times 10^{-2}), outperforming Random Forests (RF, 1.26×10−21.26 \times 10^{-2}), Color Attenuation Prior (CAP, 3.32×10−23.32 \times 10^{-2}), Dark Channel Prior (DCP, 3.18×10−23.18 \times 10^{-2}), and Back Propagation Neural Network (BPNN, 4.37×10−24.37 \times 10^{-2}). Unlike RF, which sorts patch feature values and discards spatial structure, DehazeNet preserves spatial content, preventing transmission estimation errors in sky and white-colored regions.

  7. Knowl 7 — Quantitative Evaluation on Synthesized Stereo Benchmark Images

    data/table

    DehazeNet and six state-of-the-art dehazing methods were evaluated on 12 synthesized hazy images created from the Middlebury Stereo Datasets (2001--2006) with ground truth depth maps d(x)d(x), atmospheric scattering coefficient β=1\beta = 1, and atmospheric light α=1\alpha = 1. Restored images are evaluated using Mean Squared Error (MSE), Structural Similarity (SSIM), Peak Signal-to-Noise Ratio (PSNR in dB), and Weighted Peak Signal-to-Noise Ratio (WSNR in dB) under one-pass evaluation (OPE).

    Metric ATM BCCR FVR DCP CAP RF DehazeNet
    MSE 0.0689 0.0243 0.0155 0.0172 0.0075 (0.0068) 0.0070 0.0062
    SSIM 0.9890 0.9963 0.9973 0.9981 0.9991 (0.9990) 0.9989 0.9993
    PSNR 60.8612 65.2794 66.5450 66.7392 70.0029 (70.6581) 70.0099 70.9767
    WSNR 7.8492 12.6230 13.7236 13.8508 16.9873 (17.7839) 17.1180 18.0996

    Values outside parentheses represent the authors' original code for CAP, and values inside represent a reimplementation. DehazeNet achieves the best scores across all four metrics, yielding the lowest MSE (0.0062) and highest SSIM (0.9993), PSNR (70.9767 dB), and WSNR (18.0996 dB).

  8. Knowl 8 — Robustness to Scattering Coefficient, Airlight, Scale, and Noise

    data/table

    The robustness of DehazeNet was evaluated across four environmental and sensor perturbations on synthetic images from the Middlebury stereo dataset, measured by Mean Squared Error (MSE):

    1. Coefficient Robustness Evaluation (CRE): Evaluates varying haze densities β∈{0.75,1.00,1.25,1.50}\beta \in \{0.75, 1.00, 1.25, 1.50\}.
    2. Airlight Robustness Evaluation (ARE): Evaluates non-white atmospheric light α∈{[1.0,1.0,1.0],[0.9,1.0,1.0],[1.0,0.9,1.0],[1.0,1.0,0.9]}\alpha \in \{[1.0, 1.0, 1.0], [0.9, 1.0, 1.0], [1.0, 0.9, 1.0], [1.0, 1.0, 0.9]\}.
    3. Scale Robustness Evaluation (SRE): Evaluates image scale variations with scale factors s∈{0.40,0.60,0.80,1.00}s \in \{0.40, 0.60, 0.80, 1.00\}.
    4. Noise Robustness Evaluation (NRE): Evaluates additive white Gaussian noise with standard deviation σ∈{10,15,20,25,30}\sigma \in \{10, 15, 20, 25, 30\}.
    Evaluation ATM BCCR FVR DCP CAP RF DehazeNet
    CRE Average 0.0653 0.0254 0.0187 0.0177 0.0105 (0.0095) 0.0094 0.0084
    ARE Average 0.0727 0.0255 0.0159 0.0192 0.0075 (0.0068) 0.0074 0.0067
    SRE Average 0.0581 0.0235 0.0155 0.0144 0.0098 (0.0072) 0.0077 0.0062
    NRE Average – 0.0255 0.0150 0.0100 0.0088 (0.0087) 0.0137 0.0055

    DehazeNet achieves the lowest average MSE in all four categories. Multi-scale convolutional mapping in the second layer provides robustness across image scales ss, while the combination of Maxout in the first layer and local max-pooling in the third layer provides resilience to Gaussian noise.

  9. Knowl 9 — Ablation of Maxout Dimension Reduction and BReLU Activation

    empirical result

    Controlled experiments isolating individual architectural choices in DehazeNet demonstrate the advantages of Maxout units in the feature extraction layer (F1F_1) and BReLU activation in the regression layer (F4F_4):

    • Feature Extraction (F1F_1) Mapping: Replacing the Maxout unit with a linear unit (16×1×116 \times 1 \times 1 convolution) or a ReLU-activated sparse linear unit slows network convergence. Maxout achieves a convergent Mean Squared Error (MSE) of 0.01090.0109, improving over ReLU (0.01400.0140) and the linear unit (0.01400.0140) by approximately 0.30×10−20.30 \times 10^{-2}.
    • Non-Linear Regression (F4F_4) Activation: Replacing BReLU (F4=min⁡(1,max⁡(0,W4∗F3+B4))F_4 = \min(1, \max(0, W_4 * F_3 + B_4))) with standard ReLU (F4=max⁡(0,W4∗F3+B4)F_4 = \max(0, W_4 * F_3 + B_4)) or Sigmoid (F4=(1+exp⁡(−W4∗F3−B4))−1F_4 = (1 + \exp(-W_4 * F_3 - B_4))^{-1}) reduces transmission accuracy on the test set:
      • BReLU: Test MSE = 1.19×10−21.19 \times 10^{-2}. Predicted transmissions align along the ideal diagonal across the entire range [0,1][0, 1].
      • ReLU: Test MSE = 1.28×10−21.28 \times 10^{-2}. Predictions are systematically biased above true values and frequently overflow past the physical limit tmax⁡=1t_{\max} = 1.
      • Sigmoid: Test MSE = 1.46×10−21.46 \times 10^{-2}. Predictions suffer from curvature distortion near boundaries, clustering heavily toward 0 and 1.

  10. Knowl 10 — Ablation on Filter Number and Kernel Sizes in DehazeNet

    data/table

    DehazeNet performance was evaluated across varying network widths (number of output channels n1,n2n_1, n_2) and kernel sizes (f1,f2,f3,f4f_1, f_2, f_3, f_4 across the four layers). Models were trained under identical conditions and evaluated using Train and Test Mean Squared Error (MSE ×10−2\times 10^{-2}).

    Filter Dimension Architecture Train MSE Test MSE # Param
    Channel Number 4-(16×\times3) 1.090 1.190 8,240
    (n1n_1–n2n_2) 8-(32×\times3) 0.972 1.138 27,104
    16-(64×\times3) 0.902 1.112 96,704
    F2F_2 Kernel Size 5-3-7-6 1.184 1.219 4,656
    (f1f_1–f2f_2–f3f_3–f4f_4) 5-5-7-6 1.133 1.225 7,728
    5-7-7-6 1.021 1.184 12,336
    5-M-7-6 1.090 1.190 8,240
    F4F_4 Kernel Size 5-M-6-7 1.077 1.192 8,864
    (f1f_1–f2f_2–f3f_3–f4f_4) 5-M-7-6 1.090 1.190 8,240
    5-M-8-5 1.103 1.201 7,712

    Increasing the channel width improves test MSE from 1.190×10−21.190 \times 10^{-2} to 1.112×10−21.112 \times 10^{-2} at the expense of a 10-fold increase in parameter count. In the second layer, multi-scale mapping (denoted M, using parallel 3×33 \times 3, 5×55 \times 5, and 7×77 \times 7 kernels) achieves a test MSE (1.190×10−21.190 \times 10^{-2}) comparable to a large single-scale 7×77 \times 7 filter (1.184×10−21.184 \times 10^{-2}) while using fewer parameters (8,240 vs. 12,336). Setting f3=7f_3 = 7 and f4=6f_4 = 6 provides the best trade-off between representation capacity and overfitting.

  11. Knowl 11 — Limitations of the DehazeNet Framework

    limitation

    The DehazeNet framework has three principal limitations:

    1. Heuristic Atmospheric Light Estimation: The global atmospheric light α\alpha is treated as a spatial constant and estimated via a heuristic percentile search on the transmission map rather than being learned jointly with the transmission map inside the neural network.
    2. Requirement for Post-Processing Transmission Refinement: Due to the dense local max-pooling in the third layer, raw transmission maps exhibit blocking artifacts and require secondary refinement via Guided Image Filtering before haze-free scene recovery.
    3. Decoupled Two-Stage Restoration: Restoration relies on the classical analytical inversion of the atmospheric scattering model after transmission estimation, rather than directly mapping hazy images to haze-free images end-to-end within a unified deep network.

Coverage note — No substantial contributed material was omitted; the knowls cover the full model architecture, BReLU formulation, dehazing pipeline, synthetic training strategy, theoretical connections to prior dehazing methods, ablation studies, benchmark results, robustness tests, and limitations.

References

  1. 1.T. K. Kim, J. K. Paik, and B. S. Kang, “Contrast enhancement system using spatially adaptive histogram equalization with temporal filtering,” IEEE Transactions on Consumer Electronics, vol. 44, no. 1, pp. 82–87, 1998.
  2. 2.J. A. Stark, “Adaptive image contrast enhancement using generalizations of histogram equalization,” IEEE Transactions on Image Processing, vol. 9, no. 5, pp. 889–896, 2000.
  3. 3.R. Eschbach and B. W. Kolpatzik, “Image-dependent color saturation correction in a natural scene pictorial image,” Sep. 12 1995, uS Patent 5,450,217.
  4. 4.Y. Y. Schechner, S. G. Narasimhan, and S. K. Nayar, “Instant dehazing of images using polarization,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), vol. 1, 2001, pp. I–325.
  5. 5.S. G. Narasimhan and S. K. Nayar, “Contrast restoration of weather degraded images,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 25, no. 6, pp. 713–724, 2003.
  6. 6.J. Kopf, B. Neubert, B. Chen, M. Cohen, D. Cohen-Or, O. Deussen, M. Uyttendaele, and D. Lischinski, “Deep photo: Model-based photo-graph enhancement and viewing,” in ACM Transactions on Graphics (TOG), vol. 27, no. 5, 2008, p. 116.
  7. 7.R. T. Tan, “Visibility in bad weather from a single image,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2008, pp. 1–8.
  8. 8.R. Fattal, “Single image dehazing,” in ACM Transactions on Graphics (TOG), vol. 27, no. 3, 2008, p. 72.
  9. 9.K. He, J. Sun, and X. Tang, “Single image haze removal using dark channel prior,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 33, no. 12, pp. 2341–2353, 2011.
  10. 10.K. Nishino, L. Kratz, and S. Lombardi, “Bayesian defogging,” International journal of computer vision, vol. 98, no. 3, pp. 263–278, 2012.
  11. 11.G. Meng, Y. Wang, J. Duan, S. Xiang, and C. Pan, “Efficient image dehazing with boundary constraint and contextual regularization,” in IEEE International Conference on Computer Vision (ICCV), 2013, pp. 617–624.
  12. 12.K. B. Gibson, D. T. Vo, and T. Q. Nguyen, “An investigation of dehazing effects on image and video coding,” IEEE Transactions on Image Processing, vol. 21, no. 2, pp. 662–673, 2012.
  13. 13.J.-P. Tarel and N. Hautiere, “‘Fast visibility restoration from a single color or gray level image,” in IEEE International Conference on Computer Vision, 2009, pp. 2201–2208.
  14. 14.J. Yu, C. Xiao, and D. Li, “Physics-based fast single image fog re-moval,” in IEEE International Conference on Signal Processing (ICSP), 2010, pp. 1048–1052.
  15. 15.K. He, J. Sun, and X. Tang, “Guided image filtering,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 35, no. 6, pp. 1397–1409, 2013.
  16. 16.A. Levin, D. Lischinski, and Y. Weiss, “‘A closed-form solution to natural image matting,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 30, no. 2, pp. 228–242, 2008.
  17. 17.K. Tang, J. Yang, and J. Wang, “Investigating haze-relevant features in a learning framework for image dehazing,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014, pp. 2995–3002.
  18. 18.Q. Zhu, J. Mai, and L. Shao, “‘A fast single image haze removal algorithm using color attenuation prior,” IEEE Transactions on Image Processing, vol. 24, no. 11, pp. 3522–3533, 2015.
  19. 19.A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” in Advances in neural information processing systems, 2012, pp. 1097–1105.
  20. 20.C. Ding and D. Tao, “‘Robust face recognition via multimodal deep face representation,” IEEE Transactions on Multimedia, vol. 17, no. 11, pp. 2049–2058, 2015.
  21. 21.R. Girshick, J. Donahue, T. Darrell, and J. Malik, “‘Rich feature hierarchies for accurate object detection and semantic segmentation,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014, pp. 580–587.
  22. 22.C. J. Schuler, H. C. Burger, S. Harmeling, and B. Scholkopf, “‘A machine learning approach for non-blind image deconvolution,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2013, pp. 1067–1074.
  23. 23.C. Dong, C. C. Loy, K. He, and X. Tang, “‘Image super-resolution using deep convolutional networks,” IEEE Transactions on Pattern Analysis and Machine Intelligence, no. 99, pp. 1–1, 2015.
  24. 24.D. Eigen, D. Krishnan, and R. Fergus, ‘Restoring an image taken through a window covered with dirt or rain,” in IEEE International Conference on Computer Vision (ICCV), 2013, pp. 633–640.
  25. 25.Z. Wu, D. Lin, and X. Tang, “Adjustable bounded rectifiers: Towards deep binary representations,” arXiv preprint arXiv:1511.06201, 2015.
  26. 26.E. J. McCartney, “‘Optics of the atmosphere: scattering by molecules and particles,” New York, John Wiley and Sons, Inc., vol. 1, 1976.
  27. 27.S. K. Nayar and S. G. Narasimhan, ‘Vision in bad weather,” in IEEE International Conference on Computer Vision, vol. 2, 1999, pp. 820–827.
  28. 28.S. G. Narasimhan and S. K. Nayar, “Contrast restoration of weather degraded images,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 25, no. 6, pp. 713–724, 2003.
  29. 29.C. O. Ancuti, C. Ancuti, C. Hermans, and P. Bekaert, ‘‘A fast semi-inverse approach to detect and remove the haze from a single image,” in Computer Vision–ACCV 2010, 2011, pp. 501–514.
  30. 30.I. Goodfellow, D. Warde-farley, M. Mirza, A. Courville, and Y. Bengio, “Maxout networks,” in Proceedings of the 30th International Conference on Machine Learning (ICML-13), 2013, pp. 1319–1327.
  31. 31.C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich, “‘Going deeper with convolutions,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015, pp. 1–9.
  32. 32.O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein et al., “‘Imagenet large scale visual recognition challenge,” International Journal of Computer Vision, pp. 1–42, 2014.
  33. 33.I. Lampl, D. Ferster, T. Poggio, and M. Riesenhuber, “‘Intracellular measurements of spatial integration and the max operation in complex cells of the cat primary visual cortex,” Journal of neurophysiology, vol. 92, no. 5, pp. 2704–2713, 2004.
  34. 34.Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “‘Gradient-based learning applied to document recognition,” Proceedings of the IEEE, vol. 86, no. 11, pp. 2278–2324, 1998.
  35. 35.G. E. Hinton and R. R. Salakhutdinov, “‘Reducing the dimensionality of data with neural networks,” Science, vol. 313, no. 5786, pp. 504–507, 2006.
  36. 36.V. Nair and G. E. Hinton, ‘Rectified linear units improve restricted boltzmann machines,” in Proceedings of the 27th International Conference on Machine Learning (ICML-10), 2010, pp. 807–814.
  37. 37.Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell, “‘Caffe: Convolutional architecture for fast feature embedding,” in Proceedings of the ACM International Conference on Multimedia. ACM, 2014, pp. 675–678.
  38. 38.J.-P. Tarel and N. Hautiere, “‘Fast visibility restoration from a single color or gray level image,” in IEEE International Conference on Computer Vision, 2009, pp. 2201–2208.
  39. 39.M. Sulami, I. Glatzer, R. Fattal, and M. Werman, “Automatic recovery of the atmospheric light in hazy images,” in IEEE International Conference on Computational Photography (ICCP), 2014, pp. 1–11.
  40. 40.J. Mai, Q. Zhu, D. Wu, Y. Xie, and L. Wang, “‘Back propagation neural network dehazing,” IEEE International Conference on Robotics and Biomimetics (ROBIO), pp. 1433–1438, 2014.
  41. 41.C. Dong, C. C. Loy, K. He, and X. Tang, “‘Learning a deep convolutional network for image super-resolution,” in Computer Vision–ECCV 2014. Springer, 2014, pp. 184–199.
  42. 42.T.-H. Chan, K. Jia, S. Gao, J. Lu, Z. Zeng, and Y. Ma, “‘Pcanet: A simple deep learning baseline for image classification?” IEEE Transactions on Image Processing, vol. 24, no. 12, pp. 5017–5032, 2015.
  43. 43.B. Scholkopft and K.-R. Mullert, ‘Fisher discriminant analysis with kernels,” Neural networks for signal processing IX, vol. 1, p. 1, 1999.
  44. 44.D. Scharstein and R. Szeliski, “‘A taxonomy and evaluation of dense two-frame stereo correspondence algorithms,” International journal of computer vision, vol. 47, no. 1-3, pp. 7–42, 2002.
  45. 45.——, “‘High-accuracy stereo depth maps using structured light,” in IEEE Computer Society Conference on Computer Vision and Pattern Recognition, vol. 1. IEEE, 2003, pp. I–195.
  46. 46.D. Scharstein and C. Pal, “‘Learning conditional random fields for stereo,” in IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 2007, pp. 1–8.
  47. 47.Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,” IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600–612, 2004.
  48. 48.J. L. Mannos and D. J. Sakrison, “‘The effects of a visual fidelity criterion of the encoding of images,” IEEE Transactions on Information Theory, vol. 20, no. 4, pp. 525–536, 1974.

Citation

MLA
Cai, B., et al. “DehazeNet: An End-to-End System for Single Image Haze Removal”. IEEE Transactions on Image Processing, vol. 25, no. 11, 2016, pp. 5187–98, https://doi.org/10.1109/TIP.2016.2598681.
APA
Cai, B., Xu, X., Jia, K., Qing, C., & Tao, D. (2016). DehazeNet: An End-to-End System for Single Image Haze Removal. IEEE Transactions on Image Processing, 25(11), 5187–5198. https://doi.org/10.1109/TIP.2016.2598681
Chicago
Cai, B., X. Xu, K. Jia, C. Qing, and D. Tao. 2016. “DehazeNet: An End-to-End System for Single Image Haze Removal”. IEEE Transactions on Image Processing 25 (11): 5187–98. https://doi.org/10.1109/TIP.2016.2598681.
Harvard
Cai, B. et al. (2016) “DehazeNet: An End-to-End System for Single Image Haze Removal”, IEEE Transactions on Image Processing, 25(11), pp. 5187–5198. Available at: https://doi.org/10.1109/TIP.2016.2598681.
Vancouver
1. Cai B, Xu X, Jia K, Qing C, Tao D (2016) DehazeNet: An End-to-End System for Single Image Haze Removal. IEEE Transactions on Image Processing 25:5187–5198

BibTeX

@article{Cai_2016, title={DehazeNet: An End-to-End System for Single Image Haze Removal}, volume={25}, ISSN={1941-0042}, url={http://dx.doi.org/10.1109/TIP.2016.2598681}, DOI={10.1109/tip.2016.2598681}, number={11}, journal={IEEE Transactions on Image Processing}, publisher={Institute of Electrical and Electronics Engineers (IEEE)}, author={Cai, Bolun and Xu, Xiangmin and Jia, Kui and Qing, Chunmei and Tao, Dacheng}, year={2016}, month=Nov, pages={5187–5198} }
Metadata:Crossref

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF