High dynamic range image tone mapping technique based on the improved bitonic filter

Abstract

The field of High dynamic range (HDR) imaging technique tries to solve the problem of capturing images and display in the devices, which are limited in dynamic range. Tone mapping technique approximates the appearance of an HDR image based on the logarithm of the luminance value. The tone mapping technique leads to detail loss, artifacts and higher computation time in local tone mapping. The aim of the research is to increase the performance of the tone mapping and to increases the image quality. The new technique is proposed such as Improved Bitonic Tone Mapping (IBTM) for efficient tone mapping. The weight factor is added to the bitonic filter and edges having high weight, when compared to the flat surface. The proposed method is compared with different tone mapping algorithms (Reinhard, Drago, Durand, Mantiuk, Pattanaik, Exposure and Gamma, Local adaptation, Equalize Histogram, Adaptive reference value) with quality. The experimental result shows that higher performance can be achieved by the proposed technique in a) preserving the quality of the image which is measured by PSNR and TMQM metrics and in b) execution time taken to generate the image measured in seconds.

Keywords

Artifacts high dynamic range improved bitonic tone mapping luminance tone mapping and weighting factor

1. Introduction

Natural scenes always contain high dynamic range areas in comparison with the limited dynamic range capabilities of camerasor displays. The dynamic range is the ratio between the maximum and minimum light intensities of the scene [1]. Hence, HDR image/video has recently received significant attention in many multimedia applications, such as digital photography, ultra-high-definition movies and television, video games, remote sensing detection, and medical imaging [2]. HDR images, which differ from traditional low-dynamic-range (LDR) images that use a 24-bit per pixel data format (8-bit for each of the three-color channels), can adopt additional data bits to represent real-world scenes [3]. One key limitation of conventional 8-bit image representations is that the range of luminance available in the physical world substantially exceeds the dynamic range handled by traditional imaging pipelines, resulting in a loss of visual information in over/under-exposed image regions [4]. Global tone mapping algorithms apply a single operator to all the pixels in the image; thus, one input pixel value will result in a particular output value irrespective of its spatial location. Global methods are generally known to be easier to implement and consume low time, but they result in loss of local contrast due to global compression of the dynamic range [5].

Local tone mapping technique uses a symmetrical analysis-synthesis filer bank, and applies local gain control to the sub band. The number of LDR images used in HDR image generation is limited in consumer products due to the processing time [6]. These methods usually assume that the pixels of the different LDR images are well aligned, any motion or low edge preserving technique can lead to the annoying artifacts [7]. If a small-scale Gaussian filter is used, the local contrast and details in the resulting image are enhanced even though the artifacts are increased and vice versa [8]. Because of the cost and rarity of HDR cameras, many researches had been carried out in the global and local tone mapping, but limited in quality preserving and computation time. Generally, the quality of LDR images not only is determined by the HDR image, but is also affected by the tone mapping technique used [9, 10]. In this research, a new tone mapping technique has been proposed based on the improve bitonic filter to achieve the enhanced image quality. The bitonic filter doesn’t contain data-level-sensitivity parameters and can locally adapt to the pixel and noise levels in the images, which prevent smoothness of the image and reduce the noise level. The weight factor has been incorporated in the bitonic filter and higher weight is assigned for the edge of the objects in the image. The smoothness of the images is maintained and edge loss has been prevented in this technique to improve the quality of the image. The experimental result shows the efficiency of the tone mapping technique and has less computational time in local tone mapping method.

The paper has been organized as literature review in the Section 2, proposed tone mapping technique has been briefly explained in the Section 3 and different metrics evaluation is given in Section 4 and experimental result is shown in Section 5.

2. Literature review

The visual details of the image have been lost in the ordinary display due to bright and dark area of the image. Tone mapping technique has been used to minimize the dynamic range of the image. Several techniques have been proposed to improve the image quality in tone mapping technique. Recent technique used in the tone mapping has been surveyed in this section to analysis the robustness.

Gommelet et al. [11] analyzed the problem of tone mapping operator for rate distortion optimized backward compatible compression of HDR images. The distortion and rate are expressed as a function of the gradient in the HDR image. The proposed rate and distortion model are based on the function of the gradient in the HDR image. The experimental result shows that the proposed operator achieves optimal value for rate-distortion performance in global tone mapping. The second tone mapping operator has trade-off performance of the distortion rate and quality preservation. The quality of the image should be preserved and evaluated on various dataset.

Artusi et al. [12] proposed high order reconstruction (HOR) technique to preserve the edge in multi-scale edge aware tone mapping. The edge-aware technique increases the quality by smoothing an input images, while keeping its edges intact. The proposed HOR technique focus on various factors of the performance such as altering the image structure due to changes in contrast; removes artifacts around edges; as well as reducing computational complexity in terms of implementation and associated computational costs. The proposed technique aims to reduce the changes in the image structure by using an edge-stop mechanism, whose computational cost is compared to the state-of-art method. In the 18th evaluated image shows that the proposed HOR has the higher value for the loss of contrast. This technique is not much suitable to apply for the images with high contrast (blue) due to low amplification.

Kim et al. [13] proposed the weighted least squares (WLS) filter to preserve the detail and contrast of the image in HDR tone mapping, which preserve the global contrast in a scene by using a competitive learning neural network. The proposed method is applied before a tone reconstruction operator that preserves the color without shifting the lightness. According to the Helmholtz-Kohlrausch effect, the perception of brightness depends on the lightness, Chroma, and hue of a color. Based on the effect, the proposed method corrects the lightness of pixels according to the color of a tone mapped image. The filtering of the luminance in the image helps to measure the difference in luminance between each pixel, which is small in homogeneous regions and larger in edge pixels. The luminance value is used to preserve edge and reduce artifacts. The local contrast is low in the HDR technique and this filter is not reliable in local tone mapping.

Abebe et al. [14] developed a method to recover the lost details of the over-exposure image by considering channel cross-correlation in RGB image. The clipping threshold is applied for the image white point followed the optimal selection of parameter, which are included to automate the process. The input to the proposed method is 8-bit RGB video or image, which is linearized by removing and gamma correlation. The linear RGB values are used to encoded and this normally requires more than 8-bits, as luminance and chrominance values are often beyond 255. The corrected results are further compressed to 8-bit or expanded to a wider dynamic range depending on the target applications. The several images are processed through the proposed method and visual difference are quantized. This design helps to improve the performance of tone mapping. The over-exposure image having no information after applied this technique and find some difficulties in recovering some details.

Ok and Lee [15] measures the adaptive reference value from the images, which provides effective reconstruction the details of bright and shadow regions in HDR tone mapping. The tone mapping is processed for global and local form based on objective quality metrics. The new objective quality metric is developed for the input images and the proposed detailed metric measures shows good correlations with subjective quality. The proposed metric is hybridized with recent tone mapping technique such as Tone Mapped image Quality Index (TMQI), due to its detail loss. The evaluation of the proposed method has the better perceptual quality than existing methods. The guided filter is used in this method for preserving edge in the images. The halo artifacts are also reduced by guided filter while smoothing the image. Another guided filter is used to restore coarse edges by transferring the preserved edges of the initially smoothed image to the output of the Gaussian low pass filter. The proposed method is tested on 15 images set and preserve the quality of the image. The proposed adaptive reference measure is suitable for global function and higher computation time is needed for local tone mapping.

The several methods are proposed to improve the function of the tone mapping and some limitations are need to overcome. The tone mapping techniques are needed to preserve details of the image and minimize the artifacts. The computational time of the local tone mapping is needed to be decreased and preserve the quality of the image.

3. Tone mapping

Figure 1 is a flowchart of the proposed TMO based on the improved bitonic filter. The goal of tone reproduction is to compress the dynamic range of an image to the range that can be displayed on physical devices in case that the luminance range of the images is much broader than that of physical devices. A number of tone mapping techniques have been presented, and most of them can be categorized into two groups: global and local operators. Global operators apply the same transformation to every pixel of an image while local ones adapt their scales to different areas of an image. The luminance of an image is initially mapped by using a global tone mapping function to compress the range of luminance into the displayable range.

Figure 1.

Flow chart of the proposed architecture.

To enhance the quality of an image, a local adaptation is based on photographic “dodging and burning” technique which allows a different exposure for each part of the applied image. Then method is developed to operate automatically, freeing the user from setting parameters [16]. To automate processes, low contrast regions are found by a center-surround function at different scales. Then, a tone mapping function is locally applied. The automatic dodging and burning method enhance contrast and details in an image while preserving the overall luminance characteristics

Given an HDR image, the luminance component of the HDR image in Eq. (1) is decomposed as

$\displaystyle Y_{h}(p)=Y_{b}(p)Y_{d}(p),$ (1)

Where $Y_{b}$ and $Y_{d}$ are the base layer and the detail layer, respectively. The dynamic range of the detail layer $Y_{d}$ is small while the base layer $Y_{b}$ could vary a great deal, often much more than the detail layer $Y_{d}$ . The overall dynamic range is reduced by using a global tone mapping algorithm to scale down the base layer $Y_{b}$ as $\hat{y}_{b}$ . The detail layer $Y_{d}$ is preserved or even amplified as $\hat{y}_{d}$ to enhance local contrasts. The compressed luminance value $Y_{l}$ is the product of $\hat{y}_{b}$ and $\hat{y}_{d}$ . The HDR image $\{R,G,B\}_{h},Y_{h}$ and $Y_{l}$ are finally adopted to generate the output LDR image $\{R,G,B\}$ .

The log function approximates the transformation performed by the retina of the HVS [17] and from this decomposition is performed in the log domain in Eq. (2)

$\displaystyle L_{h}(p)=L_{b}(p)+L_{d}(p),$ (2)

Where $L_{b}(p)$ and $L_{d}(p)$ are $\log(Y_{b}(p))$ and $\log(Y_{d}(p))$ , respectively. Three major function of the proposed technique are: (1) the decomposition of $L_{h}$ into $L_{b}$ and $L_{d}$ , (2) the compression of $L_{b}$ through global tone mapping and (3) the amplification of $L_{d}$ . Improved bitonic filter is used to decompose the HDR images.

3.1 Improved bitonic filter

The median filter is commonly used to eliminate impulsive noise, whilst preserving edges well, at least if there is no more than one edge within the range of the filter. A rank filter is a generalization of the median where any centile can form the output. Such filters can be considered as monotonic in that they preserve signals which are monotonically increasing or decreasing, and indeed this leads naturally to impulsive noise reduction, since impulses are bitonic rather than monotonic. For two-dimensional (2D) data, the shape of the window used to form the set of ranked data, in morphology known as the ‘structuring element’, has some impact on which features can be preserved. Here we use a circular disk for 2D image data to ensure isotropic behavior [18]. Let $B$ be a Bitonic filter image and $\sigma_{B,1}^{2}(p^{\prime})$ be the variance of $B$ in the 3 $\times$ 3 window, $\Omega_{1}(p^{\prime})$ [19]. An edge-aware weighting $\Gamma_{B}(p^{\prime})$ is defined by using local variances of 3 $\times$ 3 windows of all pixels as follows:

$\displaystyle\Gamma_{B}(p^{\prime})=\frac{1}{N}\sum_{p=1}^{N}\frac{\sigma_{B,1% }^{2}(p^{\prime})+\epsilon}{\sigma_{B+1}^{2}(p)+\epsilon}$ (3)

where $\sigma$ is a small constant and its value is selected as $(0.001\times L)^{2}$ while $L$ is the dynamic range of the input image. All pixels in the bitonic images are used in the computation of $\Gamma_{B}(p^{\prime})$ . In addition, the weighting $\Gamma_{B}(p^{\prime})$ measures the importance of pixel $p^{\prime}$ with respect to the whole bitonic image.

The value of $\Gamma_{B}(p^{\prime})$ is usually larger than 1 if $p^{\prime}$ is at an edge and smaller than 1 if $p^{\prime}$ is in a smooth area. Clearly, larger weights are assigned to pixels at edge than those pixels in flat area by using the weight $\Gamma_{B}(p^{\prime})$ in Eq. (3). Applying this edge-aware weighting, there might be blocking artifacts in final images [20]. To prevent possible blocking artifacts from appearing in the final image, the value of $\Gamma_{B}(p^{\prime})$ is smoothed by a Gaussian filter. Clearly larger weights are assigned to pixels at edges than those pixels in flat areas. The proposed weighting matches one features of human visual system, i.e., pixels at sharp edges are usually more important than those in flat areas.

Using a rank filter of 100th centile (or maximum, known as a dilation) and immediately following this with another of 0th centile (minimum, known as an erosion) results in a morphological closing operation, which preserves signals with a local maximum, whilst rejecting any pixel with a local minimum. Reversing the order of these filters results in a morphological opening which has the opposite action. Such filters have many uses in processing the shape of data, particularly in granulometry. Figure 1 shows some examples of opening and closing operations on one-dimensional (1D) signals. Here a robust opening operation is used, with a small centile $c$ rather than the minimum, and ( $100-c$ ) in place of the maximum. This is not the same as either the slightly better-known rank-max opening or the soft-opening. The use of a small centile allows some control over impulsive noise rejection, since any impulse which takes up less of the filter range than $c$ will be rejected. The robust opening $O_{\omega,c}(x)$ and closing $C_{\omega,c}(x)$ of a pixel $x$ can be defined as:

$\displaystyle r_{\omega,c}(x)=c^{\text{th}}\mathop{\textit{centile}}_{i\in% \omega}\{x_{i}\}$ (4) $\displaystyle O_{\omega,c}(x)=r_{\{\omega\}},100-c(r_{\omega,c}(x))$ (5) $\displaystyle C_{\omega,c}(x)=r_{\omega,c}(r_{\omega},100-c(x))$ (6)

Where $r_{\omega,c}(x)$ is a rank filter, $\omega$ is the filter window (or structuring element in 2D), $|\omega|$ is the window length (or number of elements in 2D) and $c$ the chosen centile, which will generally be a fairly low percentage. However, robust opening and closing operations only do half of what is required, since they are not self-dual (symmetric in data value): they only preserve local minima or maxima respectively. In addition, it can beseen from Fig. 1 that the opening and closing operations do not preserve mean signal values in the case of a noisy signal, which would clearly be a vital property of a practical filter.

Figure 2.

Comparison of tone mapping results for SRC02 from [21]. (a) Reinhard [22], (b) Drago [23], (c) Durand [22], (d) Mantiuk [21], (e) Pattanaik [21], (f) Exposure and Gamma [15], (g) Local adaptation [15], (h) Equalize Histogram [15], (i) Adaptive reference value [15], and (j) IBTM.

Fortunately, both these drawbacks can be overcome by the same means. It is fairly clear, by comparison of the original signal with each of the opened and closed pixel, which is the most appropriate output for each part of the pixel. We can hence use such a comparison to weight a combination of the opening and closing operations. However, a weighting based on a point-wise comparison would simply return the original pixel, so instead the differences between the original signal and each of the rank-filtered pixel are smoothed with a linear filter. A Gaussian filter (i.e. a linear moving-window filter with Gaussian weights) is used for this purpose, since it is known to have good noise reduction properties. The filter length is determined experimentally to match the noise reduction from the rank filters, so that the standard deviation $\sigma=$ 0.33l where $l$ is the window length in 1D, or the diameter of the structuring element in 2D. This smoothed error can be seen in the middle column of Fig. 1.

The Gaussian linear filter as $G_{\sigma}(x)$ , this is used to weight the results of the opening and closing operations as follows:

$\displaystyle\epsilon_{O}(x)=|G_{\sigma}(x-O_{\omega,c}(x)|$ (7) $\displaystyle\epsilon_{C}(x)=|G_{\sigma}(C_{\sigma,c}(x)-x)|$ (8) $\displaystyle b_{\omega,c}(x)=\frac{\epsilon_{O}(x)C_{\omega,c}(x)+\epsilon_{C% }(x)O_{\omega,c}(x)}{\epsilon_{O}(x)+\epsilon_{C}(x)}$ (9)

Where $\epsilon_{O}(x)$ and $\epsilon_{C}(x)$ are smoothed opening and closing errors, and $b_{\omega,c}(x)$ is the output of the improve bitonic filter. This seems to have the required properties, i.e. the preservation of any bitonic signal, and reduction of noise in all regions, including across edges: the opening and closing operations effectively remove bitonic signals, leaving the Gaussian to reduce any residual noise signal everywhere to the same extent.

Figure 3.

Comparison of tone mapping results for SRC14 from [21]. (a) Reinhard [22], (b) Drago [23], (c) Durand [22], (d) Mantiuk [21], (e) Pattanaik [21], (f) Exposure and Gamma [15], (g) Local adaptation [15], (h) Equalize Histogram [15], (i) Adaptive reference value [15], and (j) IBTM.

4. Evaluation metrics

Datasets are collected from the research [21] and a perceptual quality assessment method is developed for HDR can perform akey role in adjusting the parameters of tone mapping algorithms to produce perceptually satisfactory quality under various conditions. Recently, a TMQI that provides a single quality score for TMIs was proposed [21]. The influence of detail loss in dark or bright regions can be diluted by patches under normal light conditions. Therefore, TMQI may not sufficiently reflect the detail loss in dark or bright regions whose areas may be relatively small compared to the areas under normal light conditions. To overcome this deficiency, we used an additional metric for detailness in the dark and bright regions. TMQM combines the detailness metric with TMQI. Detailness metric was calculated from under luminance (dark) and over-luminance (bright) regions by excluding the regions of normal light conditions.

The evaluation metrics of TQMI are followed from [21] and TMQM is followed from [15].

$\displaystyle\textit{TMQM}=\textit{TMQI}.\textit{DS}$ (10)

Where,

$\displaystyle\!\!\!\!\!\!\!\!\!\!\!\!\textit{DS}\!=\!\frac{\textit{DS}_{% \textit{under}}.\sum\limits_{(i,j)\in U}\!\!\textit{VS}(i,j)+\textit{DS}_{% \textit{over}}.\sum\limits_{(i,j)\in U}\!\!\textit{VS}(i,j)}{\sum\limits_{(i,j% )\in U}\textit{VS}(i,j)+\sum\limits_{(i,j)\in 0}\textit{VS}(i,j)}$

Table 1

PSNR Value for different technique in tone mapping

	PSNR (dB)
Image	Bilateral tone	Laplacian tone	Guided tone	IBTM
set	mapping [17]	mapping [17]	mapping [17]
SRC 1	29.00772	30.19475	38.19167	47.17179
SRC 2	32.53022	32.86842	35.88835	41.61905
SRC 3	31.9735	37.76288	42.08922	39.22318
SRC 4	25.45346	38.70908	39.2752	44.0556
SRC 5	26.46253	30.28261	41.97485	40.99554
SRC 6	31.91313	32.92083	39.17232	47.38188
SRC 7	26.5808	33.9287	33.57933	39.92773
SRC 8	32.93328	34.45061	35.8384	45.25411
SRC 9	34.29484	39.49938	44.61332	41.60167
SRC 10	25.54762	30.16148	43.96328	47.83824
SRC 11	34.42995	35.23552	42.0631	43.0499
SRC 12	26.54204	38.83729	34.73081	40.86063
SRC 13	30.62071	38.43677	42.67236	46.52206
SRC 14	27.00639	35.51884	36.5097	45.54858
SRC 15	28.45585	39.71937	41.30318	44.87408
Average	29.58347	35.2351	39.45767	43.72827

Table 2

TMQM evaluation on AFV and Bilateral filter

Image set	KRCC		SRCC		PCC
	TMQM measure
	ARV-GF [15]	IBTM	ARV-GF [15]	IBTM	ARV-GF [15]	IBTM
SRC 1	0.857	0.872	0.952	0.958	0.968	0.978
SRC 2	0.571	0.625	0.714	0.821	0.794	0.845
SRC 3	0.929	0.932	0.976	0.984	0.903	0.942
SRC 4	0.429	0.481	0.667	0.742	0.823	0.861
SRC 5	0.714	0.856	0.833	0.876	0.808	0.835
SRC 6	0.714	0.82	0.833	0.842	0.861	0.872
SRC 7	0.857	0.887	0.929	0.954	0.971	0.986
SRC 8	0.643	0.671	0.786	0.845	0.913	0.955
SRC 9	0.786	0.865	0.881	0.924	0.892	0.921
SRC 10	0.786	0.824	0.881	0.921	0.907	0.914
SRC 11	0.714	0.754	0.881	0.914	0.854	0.882
SRC 12	0.643	0.678	0.786	0.831	0.793	0.825
SRC 13	0.75	0.831	0.874	0.881	0.969	0.972
SRC 14	0.929	0.941	0.976	0.982	0.903	0.931
SRC 15	0.929	0.937	0.976	0.986	0.796	0.821
Average	0.750067	0.798267	0.863	0.8974	0.877	0.902667

Another old evaluation metric is PSNR (peek signal to noise ratio) represents quality of image and is given by

$\displaystyle\textit{PSNR}=10\log\left(\frac{d}{\textit{MSE}}\right)$ (11)

With $\textit{MSE}=\frac{1}{MN}([f(x,y)-r(x,y)])$

The MSE represents the average of the squares of the “errors” between our actual image and our noisy image. We can also use RMSE using the below formulae, since square root operation is an expensive operation, we use MSE for PSNR calculation.

$\displaystyle\textit{PSNR}=20\log\left(\frac{d}{\sqrt{\textit{MSE}}}\right)=20% \log\left(\frac{d}{\textit{RSE}}\right)$

5. Result and discussion

Tone mapped images are evaluated in terms of quality and computational time and compared with existing techniques. The 15-image set in the datasets are processed in the improved bitonic filter and results are compared with existing filters for tone mapping. The HDR image is tone mapped and images are validated.

The src02 image is processed from the database and the tone mapped image is compared to the existing method, displayed in Fig. 2. This proposed tone mapped method has the higher quality compared to the other technique. The higher luminous value is found in the Fig. 2c, which has loss of details in the cloud. The tone mapped image of IBTM has the higher quality and it is observed in cloud and other details of the image.

The SRC14 images from the database are evaluated in the several techniques and shown in the Fig. 3. The lower luminous value is found in the Fig. 3d in the Mantiuk technique, which has more detail loss. The higher luminous value is found in the Fig. 3c, which has loss of details in the books. The proposed tone mapping technique has the detailed quality of image compared to the other methods.

Figure 4.

Execution time of different global tone mapping.

Figure 5.

Execution time of different Local time mapping.

Table 1 shows that PSNR measure for the images processed for three filters namely bilateral, Laplacian filter and Guided filter. The Guided filter are highly used in the tone mapping technique due to its effectiveness. The PSNR value of Laplacian filter is present between 30 to 40 dB and guided filter has the higher performance compared to Laplacian. The guided filter shows the less efficiency in preserving the edges. The bitonic filter has higher performance compared to other filters due to its function. The average value of PSNR of proposed tone mapping technique is 43.72827. Then quality metrics of TMQM is measure for the proposed method along with the technique used in the research [15].

The bitonic tone mapping technique is tested with adaptive reference value method, used in the research [15]. The IBTM technique has the higher performance compared to the adaptive reference value method using guided filter. This technique is evaluated on the three-performance metrics: (1) Spearman’s rank-order correlation coefficient (SRCC); (2) Kendall’s rank-order correlation coefficient (KRCC); (3) Pearson’s correlation coefficient (PCC). The performance of TMQM for the three metrics is shown in the Table 2, where adaptive reference value using guided filter (ARV-GF) is compared with IBTM. The execution time of the proposed method is calculated and compared with other technique.

The execution has been calculated for global time mapping and local time mapping for different filter technique. The graphical representation of different method of tone mapping execution time is shown in Figs 1 and 2. The global tone mapping execution time is less compared to local tone mapping. The bitonic tone mapping technique has the lower computation time compared to the other techniques.

Hence, experiment result shows the performance of the proposed bitonic tone mapping technique. The proposed technique tone maps the HDR image to display in the normal computer or printer. The proposed method provides the better performance compared to the existing method in terms of quality and execution time.

6. Conclusion

Existing digital cameras are unable to capture full dynamic range of the scene. Several methods are introduced to capture full dynamic range of the scene. After capturing the full dynamic range, if display devices are unable to produce the high dynamic range details then it is simply useless to capture the full details. Hence tone mapping algorithms are introduced in order to display the images on the display devices. Detail preserving and high computational time are the major challenges for the existing tone mapping algorithms. We proposed a new tone mapping technique, Improved Bitonic Tone Mapping (IBTM) technique to overcome the challenges. The proposed IBTM reduce the computation time in local tone mapping compared to existing system. The weighting factor is added to the bitonic filter and higher weighting is adjust to the edge of the objects, which in turn preserve the edge of the object in the tone mapping technique. This edge preserving technique reduce the artifacts generated in the tone mapping technique.

References

Wong

Zhu

and Wang

, Gradient field multi-exposure images fusion for high dynamic range image visualization, Journal of Visual Communication and Image Representation 23(4) (2012), 604–610.

Bai

Jiang

Peng

and Chen

, Towards a tone mapping-robust watermarking algorithm for high dynamic range image based on spatial activity, Signal Processing: Image Communication 65 (2018), 187–200.

Song

Jiang

Peng

and Chen

, Quality assessment method based on exposure condition analysis for tone-mapped high-dynamic-range images, Signal Processing 146 (2018), 33–40.

Fan

Valenzise

Banterle

and Dufaux

, Fine-grained detection of inverse tone mapping in HDR images, Signal Processing (2018).

Ambalathankandy

Horé

and Yadid-Pecht

, An FPGA implementation of a tone mapping algorithm with a halo-reducing filter, Journal of Real-Time Image Processing (2016), 1–17.

Lee

J.W.

Park

R.H.

and Chang

, Noise reduction and adaptive contrast enhancement for local tone mapping, IEEE Transactions on Consumer Electronics 58(2) (2012).

Yan

Zhu

and Zhang

, Robust artifact-free high dynamic range imaging of dynamic scenes, Multimedia Tools and Applications (2018), 1–19.

Yun

B.J.

Park

Kim

H.D.

and Choi

H.H.

, Color correction for high dynamic range images using a chromatic adaptation method, Optical Review 20(1) (2013), 65–73.

Chiang

J.C.

Kao

P.H.

Chen

Y.S.

and Chen

W.R.

, High-dynamic-range image generation and coding for multi-exposure multi-view images, Circuits, Systems, and Signal Processing 36(7) (2017), 2786–2814.

10.

Chang

C.C.

Nguyen

T.S.

and Lin

C.C.

, A new distortion-free data embedding scheme for high-dynamic range images, Multimedia Tools and Applications 75(1) (2016), 145–163.

11.

Gommelet

Roumy

Guillemot

Ropert

and Le Tanou

, Gradient-based tone mapping for rate-distortion optimized backward-compatible high dynamic range compression, IEEE Transactions on Image Processing (2017).

12.

Artusi

Zhang

Drikakis

and Luo

, High-order wavelet reconstruction for multi-scale edge aware tone mapping, Computers and Graphics 45 (2014), 51–63.

13.

Kim

B.K.

Park

R.H.

and Chang

, Tone mapping with contrast preservation and lightness correction in high dynamic range imaging, Signal, Image and Video Processing 10(8) (2016), 1425–1432.

14.

Abebe

M.A.

Booth

Kervec

Pouli

and Larabi

M.C.

, Towards an automatic correction of over-exposure in photographs: Application to tone-mapping, Computer Vision and Image Understanding 168 (2018), 3–20.

15.

and Lee

, HDR tone mapping algorithm based on difference compression with adaptive reference values, Journal of Visual Communication and Image Representation 43 (2017), 61–76.

16.

Reinhard

, Parameter estimation for photographic tone reproduction, Journal of Graphics Tools 7(1) (2002), 45–51.

17.

and Zheng

, Visual-salience-based tone mapping for high dynamic range images, IEEE Transactions on Industrial Electronics 61(12) (2014), 7076–7082.

18.

Treece

, The bitonic filter: Linear filtering in an edge-preserving morphological framework, IEEE Transactions on Image Processing 25(11) (2016), 5199–5211.

19.

Dogra

Kadry

Goyal

and Agrawal

, An efficient image integration algorithm for night mode vision applications, Multimedia Tools and Applications (2018), 1–18.

20.

Goyal

Dogra

Agrawal

and Sohi

B.S.

, Two-dimensional gray scale image denoising via morphological operations in NSST domain and bitonic filtering, Future Generation Computer Systems 82 (2018), 158–175.

21.

Yeganeh

and Wang

, Objective quality assessment of tone-mapped images, IEEE Transactions on Image Processing 22(2) (2013), 657–667.

22.

Chen

Paris

and Durand

, Real-time edge-aware image processing with the bilateral grid, in: ACM Transactions on Graphics (TOG) 26(3) (2007), August, 103. ACM.

23.

Paris

Hasinoff

S.W.

and Kautz

, Local Laplacian filters: Edge-aware image processing with a Laplacian pyramid, ACM Trans Graph 30(4) (2011), 68–1.