Back

LLNet: A deep autoencoder approach to natural low-light image enhancement

Kin Gwn LoreAdedotun AkintayoSoumik Sarkar

article2015Pattern Recognition1,863 citations

Introduces LLNet, a deep autoencoder architecture that simultaneously brightens low-light images and removes noise without over-saturating highlights, demonstrating that networks trained on synthetic degradations can effectively restore natural dark scenes.

Listen

Clear visual information is critical for automated and human decision-making across defense surveillance, security monitoring, and commercial systems. However, cost constraints often require deploying low-cost camera sensors that produce dark, heavily degraded images in low-light conditions. Traditional image processing methods struggle in these environments because brightening an image typically amplifies visual noise, washes out brighter regions, or requires tedious manual parameter tuning.

The article sets out to develop and evaluate a deep learning framework capable of simultaneously brightening low-light images and removing visual noise without over-amplifying already bright areas. Specifically, the authors evaluated a deep autoencoder model—a specialized neural network architecture—to determine whether it could be trained entirely on synthetically corrupted imagery and successfully applied to enhance real-world, natural low-light photographs.

To conduct this evaluation, the authors trained their deep learning models on 422,500 synthetic image patches generated by applying randomized nonlinear darkening and varying levels of Gaussian noise to standard datasets. They designed two primary architectures: a simultaneous enhancement network called LLNet and a two-stage sequential network called S-LLNet. These models were evaluated against standard industry baselines, including histogram equalization variants, gamma adjustment, and state-of-the-art denoising filters, across both synthetically degraded images and natural low-light photos taken with a standard mobile phone camera.

The findings show that deep learning provides substantial advantages over traditional methods. First, for dark and noisy images, LLNet and S-LLNet consistently outperformed all baseline methods in quantitative quality metrics, producing substantially higher peak signal-to-noise ratios and structural similarity scores. Second, when tested on natural low-light photographs, the models adaptively illuminated dark areas while suppressing noise, successfully avoiding the severe overexposure and "blooming" artifacts produced by histogram equalization. Third, the two-stage model showed superior performance at higher noise levels, as separate network modules could specialize in contrast enhancement and noise removal. Finally, the authors identified a direct trade-off between noise removal and sharpness governed by patch size, noting that selecting an optimal patch size balances structural clarity and noise suppression.

These results demonstrate that organizations can use deep learning to extract high-quality visual data from low-cost imaging hardware operating in poor lighting conditions. Because the model learns how to brighten and denoise images automatically across a wide range of corruption levels, operational teams can deploy it without the costly and time-consuming manual calibration required by traditional filters. Furthermore, processing an image took approximately 0.42 seconds on standard graphics processing hardware, confirming the viability of this approach for near-real-time field applications.

Decision-makers should consider piloting deep autoencoder architectures in monitoring and surveillance pipelines where lighting cannot be controlled. When deploying this technology, engineering teams should evaluate the trade-off between the single-stage model for faster processing and the two-stage model for environments with severe noise. Moving forward, the framework should be expanded and tested on additional sensor degradation types, including Poisson noise, optical blurring, quantization artifacts, and atmospheric obstructions such as fog or dust.

Confidence in these findings is high for standard low-light scenarios with Gaussian-like sensor noise. However, decision-makers should exercise caution when deploying the system in environments with complex optical blur or non-Gaussian sensor distortions, as the current model was trained specifically on synthetic gamma darkening and Gaussian noise.

Iowa State Universityautoencodersdegraded imagesdenoising autoencodersnatural low-light image enhancementdeep autoencodercontrast enhancement
arXiv: 1511.03995kglore/llnet_color
  • Paper: Learning to See in the Dark, Chen Chen et al. (2018). Building on deep learning for low-light recovery, this work advances the problem to extreme darkness by training an end-to-end network on short-exposure raw sensor data.
  • Paper: Deep Retinex Decomposition for Low-Light Enhancement, Chen Wei et al. (2018). This later approach extends learned low-light enhancement with Retinex-based illumination decomposition and real paired training data to address brightness and noise together.
  • Paper: Zero-Reference Deep Curve Estimation for Low-Light Image Enhancement, Chunle Guo et al. (2020). Zero-DCE continues low-light enhancement by replacing LLNet’s synthetic paired training with image-specific curves learned through non-reference quality constraints.
Cover for LLNet: A deep autoencoder approach to natural low-light image enhancement

Abstract

In surveillance, monitoring and tactical reconnaissance, gathering the right visual information from a dynamic environment and accurately processing such data are essential ingredients to making informed decisions which determines the success of an operation. Camera sensors are often cost-limited in ability to clearly capture objects without defects from images or videos taken in a poorly-lit environment. The goal in many applications is to enhance the brightness, contrast and reduce noise content of such images in an on-board real-time manner. We propose a deep autoencoder-based approach to identify signal features from low-light images handcrafting and adaptively brighten images without over-amplifying the lighter parts in images (i.e., without saturation of image pixels) in high dynamic range. We show that a variant of the recently proposed stacked-sparse denoising autoencoder can learn to adaptively enhance and denoise from synthetically darkened and noisy training examples. The network can then be successfully applied to naturally low-light environment and/or hardware degraded images. Results show significant credibility of deep learning based approaches both visually and by quantitative comparison with various popular enhancing, state-of-the-art denoising and hybrid enhancing-denoising techniques.

Topics

autoencodersdegraded imagesdenoising autoencodersdynamic environmentshigh resolution imagesensorsvisual informationsdenoising auto encoders

Keywords

natural low-light image enhancementdeep autoencodercontrast enhancementimage denoising tasksconvolutional neural networks

Table of Contents

  • 1 Introduction and motivation
  • 2 Related work
  • 3 The Low-light Net (LLNet)
  • 4 Evaluation metrics and compared methods
  • 4.1 Performance metric
  • 4.2 Compared methods
  • 5 Results and discussion
  • 6 Conclusions and future works
  • References

Knowls

  1. Knowl 1 — LLNet and S-LLNet Architectures for Low-Light Image Enhancement

    model/method

    The Low-Light Net (LLNet) framework employs stacked sparse denoising autoencoders (SSDA) to perform adaptive contrast enhancement and denoising on low-light images. Two distinct architectural paradigms are formulated:

    1. LLNet (Simultaneous Module): A single deep SSDA trained jointly on image patches that have been both synthetically darkened and corrupted with Gaussian noise, performing simultaneous contrast enhancement and denoising.

    2. Staged LLNet (S-LLNet): A sequential pipeline of two identical SSDA modules connected in series. Stage 1 is trained exclusively on darkened, noiseless patches to perform localized contrast enhancement. Stage 2 is trained exclusively on noisy, well-lit patches to remove noise artifacts introduced by illumination and enhancement.

    Network Layer Dimensions: Each autoencoder module operates on image patches of size 17×1717 \times 17 pixels flattened into N=289N = 289 input units. The encoder comprises three hidden layers with unit counts of 867867, 578578, and a bottleneck layer of 289289 units. The decoder symmetrically mirrors the encoder with layers of 578578, 867867, and an output layer of 289289 units. Sigmoid activation functions σ(s)=(1+exp⁡(−s))−1\sigma(s) = (1 + \exp(-s))^{-1} are used across all encoder and decoder units.

    Optimization Protocol: Each denoising autoencoder (DA) layer is pre-trained in a greedy, layer-wise unsupervised manner for 3030 epochs with learning rates of 0.10.1 for the first two DA layers and 0.010.01 for the bottleneck layer. The full stacked autoencoder is subsequently fine-tuned end-to-end via error backpropagation with a learning rate of 0.10.1 for the first 200200 epochs and 0.010.01 thereafter, using early stopping when validation error improvement drops below 0.5%0.5\%.

  2. Knowl 2 — Synthetic Training Data Generation for Low-Light Simulation

    model/method

    To train deep autoencoders without requiring large collections of paired natural low-light and well-lit photographs, clean natural images are synthetically corrupted to simulate poorly illuminated environments and camera sensor noise.

    From 169169 standard test images normalized to pixel intensities in [0,1][0, 1], a total of 422,500422,500 patches of size 17×1717 \times 17 pixels are randomly cropped (2,5002,500 patches per source image). Each patch IoriginalI_{\text{original}} is transformed into a corrupted training patch Itrain=n(g(Ioriginal))I_{\text{train}} = n(g(I_{\text{original}})) through two sequential operations:

    1. Nonlinear Gamma Darkening: g(Ioriginal)=A×Ioriginalγg(I_{\text{original}}) = A \times I_{\text{original}}^\gamma where AA is a normalization constant determined by the maximum pixel intensity, and γ\gamma is drawn uniformly at random for each patch from γ∼Uniform(2,5)\gamma \sim \text{Uniform}(2, 5). Setting γ>1\gamma > 1 nonlinearly compresses pixel intensities toward darker values, mimicking low-light capture.

    2. Additive Gaussian Noise Corruption: Itrain=g(Ioriginal)+ϵ,ϵ∼N(0,σ2)I_{\text{train}} = g(I_{\text{original}}) + \epsilon, \quad \epsilon \sim \mathcal{N}(0, \sigma^2) where the standard deviation σ\sigma is randomized per patch according to: σ=(B(25255)2)1/2,B∼Uniform(0,1)\sigma = \left( B \left( \frac{25}{255} \right)^2 \right)^{1/2}, \quad B \sim \text{Uniform}(0, 1)

    The dataset is divided equally into 211,250211,250 training patches and 211,250211,250 validation patches.

  3. Knowl 3 — Sparsity-Regularized Denoising Autoencoder Layer Pre-Training Objective

    equation

    During the greedy layer-wise pre-training of each denoising autoencoder layer in LLNet, the parameter set θ={W,b,W′,b′}\theta = \{W, b, W', b'\} (with encoder weights W∈RK×NW \in \mathbb{R}^{K \times N}, encoder biases b∈RKb \in \mathbb{R}^K, decoder weights W′∈RN×KW' \in \mathbb{R}^{N \times K}, and decoder biases b′∈RNb' \in \mathbb{R}^N) is optimized by minimizing the reconstruction loss with Kullback-Leibler (KL) sparsity regularization and Frobenius norm weight decay:

    LDA(D;θ)=1Nsamples∑i=1Nsamples12∥yi−y^(xi)∥22+β∑j=1KKL(ρ∥ρ^j)+λ2(∥W∥F2+∥W′∥F2)L_{\text{DA}}(\mathcal{D}; \theta) = \frac{1}{N_{\text{samples}}} \sum_{i=1}^{N_{\text{samples}}} \frac{1}{2} \|y_i - \hat{y}(x_i)\|_2^2 + \beta \sum_{j=1}^K \text{KL}(\rho \parallel \hat{\rho}_j) + \frac{\lambda}{2} \left( \|W\|_F^2 + \|W'\|_F^2 \right)

    where xi∈RNx_i \in \mathbb{R}^N is the corrupted input patch, yi∈RNy_i \in \mathbb{R}^N is the clean uncorrupted target patch, and y^(xi)=σ′(W′σ(Wxi+b)+b′)\hat{y}(x_i) = \sigma'(W' \sigma(W x_i + b) + b') is the single-layer reconstruction using sigmoid activations σ(s)=σ′(s)=(1+exp⁡(−s))−1\sigma(s) = \sigma'(s) = (1 + \exp(-s))^{-1}.

    The sparsity penalty KL(ρ∥ρ^j)\text{KL}(\rho \parallel \hat{\rho}_j) enforces the average activation of hidden unit jj to approximate a low target activation ρ∈(0,1)\rho \in (0, 1):

    KL(ρ∥ρ^j)=ρlog⁡ρρ^j+(1−ρ)log⁡1−ρ1−ρ^j\text{KL}(\rho \parallel \hat{\rho}_j) = \rho \log \frac{\rho}{\hat{\rho}_j} + (1 - \rho) \log \frac{1 - \rho}{1 - \hat{\rho}_j}

    where ρ^j=1Nsamples∑i=1Nsampleshj(xi)\hat{\rho}_j = \frac{1}{N_{\text{samples}}} \sum_{i=1}^{N_{\text{samples}}} h_j(x_i) is the empirical mean activation of unit jj across the dataset. The hyperparameters β\beta, λ\lambda, and ρ\rho are positive scalars selected via cross-validation.

  4. Knowl 4 — Stacked Sparse Denoising Autoencoder Global Fine-Tuning Objective

    equation

    Following greedy layer-wise pre-training, all LL encoding layers and LL decoding layers of the stacked autoencoder (2L2L total parameter layers) are initialized and jointly fine-tuned end-to-end using backpropagation to minimize the mean squared reconstruction error regularized by weight decay:

    LSSDA(D;θ)=1Nsamples∑i=1Nsamples∥yi−y^(xi)∥22+λL∑l=12L∥W(l)∥F2L_{\text{SSDA}}(\mathcal{D}; \theta) = \frac{1}{N_{\text{samples}}} \sum_{i=1}^{N_{\text{samples}}} \|y_i - \hat{y}(x_i)\|_2^2 + \frac{\lambda}{L} \sum_{l=1}^{2L} \|W^{(l)}\|_F^2

    where NsamplesN_{\text{samples}} is the number of training samples, xi∈RNx_i \in \mathbb{R}^N is the degraded input patch, yi∈RNy_i \in \mathbb{R}^N is the ground-truth clean patch, y^(xi)∈RN\hat{y}(x_i) \in \mathbb{R}^N is the final output of the deep autoencoder, W(l)W^{(l)} denotes the weight matrix of layer ll, ∥⋅∥F\|\cdot\|_F is the Frobenius norm, and λ\lambda is the weight decay hyperparameter.

    Explicit KL sparsity regularization is omitted during global fine-tuning because sparse feature representations are already established in the weights during layer-wise pre-training.

  5. Knowl 5 — Overlapping Patch-Based Inference and Image Reconstruction Algorithm

    algorithm

    Full images of arbitrary dimensions are enhanced and denoised by partitioning the input into overlapping 17×1717 \times 17 patches, executing model inference, and averaging reconstructed patches across overlapping pixel coordinates:

    Input: Input image I∈RH×WI \in \mathbb{R}^{H \times W}, patch spatial dimension P=17P = 17, stride S=3S = 3, trained model M\mathcal{M}
    Output: Enhanced and denoised image I^∈RH×W\hat{I} \in \mathbb{R}^{H \times W}
    Initialize intensity accumulator A∈RH×WA \in \mathbb{R}^{H \times W} with all zeros
    Initialize count accumulator C∈RH×WC \in \mathbb{R}^{H \times W} with all zeros
    for r=1r = 1 to H−P+1H - P + 1 step SS do
        for c=1c = 1 to W−P+1W - P + 1 step SS do
            Extract sub-patch p=I[r:r+P−1,c:c+P−1]p = I[r : r+P-1, c : c+P-1]
            Flatten pp into input vector x∈RP2x \in \mathbb{R}^{P^2}
            Compute reconstruction x^=M(x)\hat{x} = \mathcal{M}(x)
            Reshape x^\hat{x} into 2D patch p^∈RP×P\hat{p} \in \mathbb{R}^{P \times P}
            A[r:r+P−1,c:c+P−1]←A[r:r+P−1,c:c+P−1]+p^A[r : r+P-1, c : c+P-1] \leftarrow A[r : r+P-1, c : c+P-1] + \hat{p}
            C[r:r+P−1,c:c+P−1]←C[r:r+P−1,c:c+P−1]+1C[r : r+P-1, c : c+P-1] \leftarrow C[r : r+P-1, c : c+P-1] + 1
        end for
    end for
    I^←A/C\hat{I} \leftarrow A / C (element-wise division)
    return I^\hat{I}

    A stride size of 3×33 \times 3 achieves visual and quantitative equivalence to dense sliding windows (1×11 \times 1 or 2×22 \times 2) while maintaining computational tractability. On an NVIDIA TITAN X GPU, this sliding window reconstruction processes a 512×512512 \times 512 image in 0.420.42 seconds.

  6. Knowl 6 — Relative Patch Size Metric and Denoising-Sharpness Trade-Off

    definition

    The relative patch size rr is a dimensionless metric relating the spatial dimensions of an autoencoder patch to the overall dimensions of the test image:

    r=dpdi−1=wp2+hp2wi2+hi2r = d_p d_i^{-1} = \frac{\sqrt{w_p^2 + h_p^2}}{\sqrt{w_i^2 + h_i^2}}

    where dp,wp,hpd_p, w_p, h_p denote the diagonal length, width, and height of the patch in pixels, and di,wi,hid_i, w_i, h_i denote the diagonal length, width, and height of the test image in pixels. The relative patch size corresponds to the effective receptive field size of the model relative to the target image.

    Empirical evaluation on standardized resolution targets (such as the 1951 USAF resolution test chart conforming to MIL-STD-150A) demonstrates a fundamental trade-off governed by rr:

    • Small relative patch sizes provide finer-grained local enhancement with sharper edges and resolved structural lines, but leave more residual noise.
    • Large relative patch sizes achieve stronger noise suppression and higher Peak Signal-to-Noise Ratio (PSNR), but over-smooth edges and blur fine details.
    • Selecting the patch size that maximizes the Structural Similarity Index (SSIM) achieves a balanced trade-off between noise suppression and structural sharpness.

  7. Knowl 7 — Quantitative Benchmark Comparison on Synthetic Dark and Noisy Images

    data/table

    The table below compares Peak Signal-to-Noise Ratio (PSNR in dB) and Structural Similarity Index (SSIM) across five standard benchmark scenes under clean, darkened (γ=3\gamma = 3), and darkened plus Gaussian noisy (σ=18\sigma = 18 and σ=25\sigma = 25) conditions. LLNet and S-LLNet are evaluated against Histogram Equalization (HE), Contrast-Limiting Adaptive Histogram Equalization (CLAHE), Gamma Adjustment (GA with γ=0.3\gamma = 0.3), and a hybrid baseline of Histogram Equalization followed by BM3D denoising with noise parameter σ=25\sigma = 25 (HE+BM3D).

    Image Degradation HE CLAHE GA HE+BM3D LLNet S-LLNet
    Bird (Clean) 11.22 / 0.63 21.55 / 0.90 8.26 / 0.63 11.27 / 0.69 20.35 / 0.92 18.07 / 0.87
    Bird-D (γ=3\gamma=3) 11.28 / 0.62 15.15 / 0.52 26.06 / 0.84 11.35 / 0.71 18.43 / 0.60 16.18 / 0.53
    Bird-D+GN18 9.25 / 0.09 14.63 / 0.11 13.49 / 0.10 9.98 / 0.13 19.73 / 0.56 18.60 / 0.54
    Bird-D+GN25 9.04 / 0.08 13.60 / 0.09 12.51 / 0.08 9.72 / 0.11 21.17 / 0.50 22.11 / 0.61
    Girl (Clean) 18.24 / 0.80 17.02 / 0.70 10.52 / 0.78 18.23 / 0.69 17.33 / 0.84 14.57 / 0.76
    Girl-D (γ=3\gamma=3) 18.27 / 0.80 14.36 / 0.66 29.73 / 0.99 18.26 / 0.69 22.45 / 0.80 21.62 / 0.74
    Girl-D+GN18 16.07 / 0.26 12.95 / 0.17 16.90 / 0.30 19.28 / 0.53 20.04 / 0.60 22.07 / 0.66
    Girl-D+GN25 15.33 / 0.19 12.09 / 0.12 15.14 / 0.20 18.50 / 0.39 19.60 / 0.49 22.68 / 0.60
    House (Clean) 13.36 / 0.70 18.89 / 0.81 9.50 / 0.56 13.24 / 0.61 11.61 / 0.60 10.57 / 0.50
    House-D (γ=3\gamma=3) 12.03 / 0.65 16.81 / 0.60 26.79 / 0.82 11.92 / 0.54 21.10 / 0.64 18.73 / 0.49
    House-D+GN18 10.55 / 0.33 15.48 / 0.35 13.76 / 0.33 11.39 / 0.42 20.25 / 0.56 19.91 / 0.51
    House-D+GN25 10.09 / 0.29 14.08 / 0.29 12.67 / 0.28 10.94 / 0.37 19.71 / 0.52 20.76 / 0.51
    Pepper (Clean) 18.61 / 0.90 18.27 / 0.76 9.63 / 0.69 18.61 / 0.84 10.92 / 0.71 9.93 / 0.66
    Pepper-D (γ=3\gamma=3) 18.45 / 0.85 15.46 / 0.61 28.28 / 0.90 18.45 / 0.80 21.33 / 0.78 19.54 / 0.72
    Pepper-D+GN18 14.69 / 0.21 14.47 / 0.17 14.66 / 0.22 16.97 / 0.57 22.23 / 0.57 21.80 / 0.65
    Pepper-D+GN25 13.67 / 0.15 13.31 / 0.13 13.87 / 0.16 15.96 / 0.36 21.48 / 0.48 23.38 / 0.62
    Town (Clean) 17.55 / 0.79 16.35 / 0.69 9.41 / 0.74 17.70 / 0.76 17.91 / 0.90 16.87 / 0.84
    Town-D (γ=3\gamma=3) 17.55 / 0.79 15.00 / 0.65 28.63 / 0.97 17.72 / 0.76 22.47 / 0.81 20.31 / 0.71
    Town-D+GN18 14.85 / 0.25 13.34 / 0.17 15.12 / 0.24 17.51 / 0.41 20.00 / 0.60 21.02 / 0.62
    Town-D+GN25 14.22 / 0.20 12.40 / 0.13 13.73 / 0.17 16.62 / 0.32 20.11 / 0.51 24.27 / 0.61

    When noise is present (+GN18 and +GN25), LLNet and S-LLNet outperform all conventional methods by wide margins (up to 7–10 dB7\text{--}10\text{ dB} PSNR improvement and over 2–4×2\text{--}4\times higher SSIM). Traditional histogram equalization methods amplify noise variance, reducing SSIM below 0.350.35. At higher noise levels (σ=25\sigma = 25), the sequential S-LLNet achieves the highest PSNR across all test scenes.

  8. Knowl 8 — Adaptive Brightening and Natural Low-Light Enhancement Behavior

    empirical result

    When applied to naturally dark indoor images captured via mobile phone camera (using lights-on photographs as paired reference points), LLNet demonstrates adaptive local brightening without over-saturating bright regions or amplifying sensor noise.

    Key behaviors include:

    • Suppression of Blooming Artifacts: In scenes with high dynamic range (e.g., an active computer display in a dark room), Histogram Equalization (HE) and HE+BM3D cause severe blooming, oversaturating the screen into an illegible white glare while amplifying background noise. LLNet enhances dark background objects while preserving the contrast and intensity boundaries of the bright screen.
    • Adaptivity on Already-Bright Scenes: When tested on normally lit, noiseless images, LLNet leaves pixel intensities largely unchanged, avoiding the washed-out, over-exposed artifacts produced by non-adaptive global methods (such as fixed Gamma Adjustment with γ<1\gamma < 1).
    • Blind Inference: Unlike Gamma Adjustment (which requires prior tuning of γ\gamma) and BM3D (which requires the noise standard deviation σ\sigma), LLNet operates blindly without hand-tuned input parameters, generalizing across varying illumination and noise levels.

  9. Knowl 9 — Learned First-Layer Feature Representations in LLNet and S-LLNet

    empirical result

    Visualizing the input-to-hidden weight matrices of the first autoencoder layer illustrates how the models decouple and execute low-light enhancement:

    • Contrast Enhancement Module (S-LLNet Stage 1): First-layer weights converge to localized, blob-like structures. These filters evaluate local neighborhood context within the 17×1717 \times 17 patch (determining whether an object boundary or edge is present) to apply context-dependent brightening without uniform over-amplification.
    • Denoising Module (S-LLNet Stage 2): First-layer weights exhibit fine-grained, noise-like patterns rather than directional Gabor-like filters, reflecting task specialization for generic noise suppression across continuously varied noise variances.
    • Simultaneous Model (LLNet): First-layer weights converge to hybrid, coarse-grained textured blobs that combine contextual brightening filters and noise smoothing into a single layer of representation.

Coverage note — No substantial contributed material was omitted.

References

  1. 1.Agostinelli, Forest, Anderson, Michael R, and Lee, Honglak. Adaptive multi-column deep neural networks with application to robust image denoising. In Advances in Neural Information Processing Systems, pp. 1493–1501, 2013.
  2. 2.Bastien, Fr´ed´eric, Lamblin, Pascal, Pascanu, Razvan, Bergstra, James, Goodfellow, Ian J., Bergeron, Arnaud, Bouchard, Nicolas, and Bengio, Yoshua. Theano: new features and speed improvements. Deep Learning and Unsupervised Feature Learning NIPS 2012 Workshop, 2012.
  3. 3.Bergstra, James, Breuleux, Olivier, Bastien, Fr´ed´eric, Lamblin, Pascal, Pascanu, Razvan, Desjardins, Guillaume, Turian, Joseph, Warde-Farley, David, and Bengio, Yoshua. Theano: a CPU and GPU math expression compiler. In Proceedings of the Python for Scientific Computing Conference (SciPy), June 2010. Oral Presentation.
  4. 4.Burger, Harold C, Schuler, Christian J, and Harmeling, Stefan. Image denoising: Can plain neural networks compete with bm3d? In Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on, pp. 2392–2399. IEEE, 2012.
  5. 5.Chan, Raymond H, Ho, Chung-Wa, and Nikolova, Mila. Salt-and-pepper noise removal by median-type noise detectors and detail-preserving regularization. Image Processing, IEEE Transactions on, 14(10):1479–1485, 2005.
  6. 6.Chen, Tao, Ma, Kai-Kuang, and Chen, Li-Hui. Tri-state median filter for image denoising. Image Processing, IEEE Transactions on, 8(12):1834–1838, 1999.
  7. 7.Cheng, HD and Shi, XJ. A simple and effective histogram equalization approach to image enhancement. Digital Signal Processing, 14(2):158–170, 2004.
  8. 8.Couprie, C., Farabet, C., Najman, L., and LeCun, Y. Indoor semantic segmentation using depth information. In ICLR, 2013.
  9. 9.Dabov, Kostadin, Foi, Alessandro, Katkovnik, Vladimir, and Egiazarian, Karen. Image denoising by sparse 3-d transform-domain collaborative filtering. Image Processing, IEEE Transactions on, 16(8):2080–2095, 2007.
  10. 10.Dabov, Kostadin, Foi, Alessandro, Katkovnik, Vladimir, and Egiazarian, Karen. Image restoration by sparse 3d transform-domain collaborative filtering. In Electronic Imaging 2008, pp. 681207–681207. International Society for Optics and Photonics, 2008.
  11. 11.Dabov, Kostadin, Foi, Alessandro, Katkovnik, Vladimir, and Egiazarian, Karen. Bm3d image denoising with shape-adaptive principal component analysis. In SPARS’09-Signal Processing with Adaptive Sparse Structured Representations, 2009.
  12. 12.Elad, Michael and Aharon, Michal. Image denoising via sparse and redundant representations over learned dictionaries. Image Processing, IEEE Transactions on, 15 (12):3736–3745, 2006.
  13. 13.Gonzalex, Rafael and Woods, Richard. Digital image Processing. Number 0-201-28075-8 in 0-201-28075-8. Prentice Hall,, upper saddle Rivers,New Jersey, second edition, 2001.
  14. 14.Jain, Viren and Seung, Sebastian. Natural image denoising with convolutional networks. Neural information Processing Standard, pp. 1–8, 2008.
  15. 15.Kaur, Manpreet, Kaur, Jasdeep, and Kaur, Jappreet. Survey of contrast enhancement techniques based on histogram equalization. International Journal of Advanced Computer Science and Applications, 2(7):137–141, 2011.
  16. 16.Krizhevsky, A., Sutskever, I., and Hinton, G. E. Imagenet classification with deep convolutional neural networks. NIPS 2012: Neural Information Processing Systems, Lake Tahoe, Nevada, 2012.
  17. 17.Krutsch, Robert and Tenorlo, David. Histogram equalization. Application Note AN4318, Freescale Semiconductors Inc, June 2011.
  18. 18.Loza, Artur, Bull, David, Hill, Paul, and Achim, Alin. Automatic contrast enhancement of low-light images based on local statistics of wavelet coefficients. Elsevier Digital Signal Processing, 23(6):1856–1866, December 2013.
  19. 19.Pisano, Etta, Zong, Shuquan, Hemminger, Bradley, DeLuce, Maria, Johnston, Eugene, Muller, Keith, Braeuning, Patricia, and Pizer, Stephen. Contrast limited adaptive histogram equalization image processing to improve the detection of simulated spiculations in dense mammograms. Journal of Digital Imaging, 11(4):193–200, November 1998.
  20. 20.Pizer, Stephen M, Amburn, E Philip, Austin, John D, Cromartie, Robert, Geselowitz, Ari, Greer, Trey, ter Haar Romeny, Bart, Zimmerman, John B, and Zuiderveld, Karel. Adaptive histogram equalization and its variations. Computer vision, graphics, and image processing, 39(3):355–368, 1987.
  21. 21.Santoso, Albertus, Nugroho, Edi, Suparta, Bayu, and Hidayat, Risanuri. Compression ratio and peak signal to noise ratio in grayscale image compression using wavelet. International Journal of Computer Science and Technology, 2(2):1–10, June 2011.
  22. 22.Sarkar, S., Venugopalan, V., Reddy, K., Ryde, J., Giering, M., and Jaitly, N. Occlusion edge detection in rgbd frames using deep convolutional neural networks. Proceedings of IEEE High Performance Extreme Computing Conference, (Waltham, MA), 2015.
  23. 23.Schuler, Christian, Hirsh, Michael, Harmeling, Stefan, and Scholkpf, Hernhard. Learning to deblur. arXiv:1406.7444v1 [cs.CV], pp. 1–28, June 2014.
  24. 24.Trahanias, PE and Venetsanopoulos, AN. Color image enhancement through 3-d histogram equalization. In Pattern Recognition, 1992. Vol. III. Conference C: Image, Speech and Signal Analysis, Proceedings., 11th IAPR International Conference on, pp. 545–548. IEEE, 1992.
  25. 25.Vincent, Pascal, Larochelle, Hugo, and Bengio, Yoshua. Extracting and composing robust features with denoising autoencoders. Proceedings of the 25th International conference on Machine Learning-ICML ’08, pp. 1096–1103, 2008.
  26. 26.Wu, Xiaolin. A linear programming approach for optimal contrast-tone mapping. IEEE Transaction on Image Processing, 20(5):1262–1272, May 2011.
  27. 27.Xie, Junyuan, Xu, Linli, and Chen, Enhong. Image denoising and inpainting with deep neural networks. In Advances in Neural Information Processing Systems, pp. 341–349, 2012.
  28. 28.Z.Wang, Bovik, A., H.Sheik, and E.Simoncelli. Image quality assessment: From error visibility to structural similarity. IEE Trans. Image Process., 13(4):600–612, 2004.

Citation

MLA
Lore, K. G., et al. “LLNet: A Deep Autoencoder Approach to Natural Low-light Image Enhancement”. arXiv, 2015, http://arxiv.org/abs/1511.03995v3.
APA
Lore, K. G., Akintayo, A., & Sarkar, S. (2015). LLNet: A Deep Autoencoder Approach to Natural Low-light Image Enhancement. arXiv. http://arxiv.org/abs/1511.03995v3
Chicago
Lore, K. G., A. Akintayo, and S. Sarkar. 2015. “LLNet: A Deep Autoencoder Approach to Natural Low-light Image Enhancement”. arXiv. http://arxiv.org/abs/1511.03995v3.
Harvard
Lore, K.G., Akintayo, A. and Sarkar, S. (2015) “LLNet: A Deep Autoencoder Approach to Natural Low-light Image Enhancement”, arXiv [Preprint]. Available at: http://arxiv.org/abs/1511.03995v3.
Vancouver
1. Lore KG, Akintayo A, Sarkar S (2015) LLNet: A Deep Autoencoder Approach to Natural Low-light Image Enhancement. arXiv

BibTeX

@article{lore2015llnet,
  title = {LLNet: A Deep Autoencoder Approach to Natural Low-light Image Enhancement},
  author = {Lore, Kin Gwn and Akintayo, Adedotun and Sarkar, Soumik},
  year = {2015},
  journal = {arXiv},
  url = {http://arxiv.org/abs/1511.03995v3},
  eprint = {1511.03995}
}
Metadata:arXiv

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