WO2008019524A1

WO2008019524A1 - Method and apparatus for encoding video color enhancement data, and method and apparatus for decoding video color enhancement data

Info

Publication number: WO2008019524A1
Application number: PCT/CN2006/001699
Authority: WO
Inventors: Yongying Gao; Yuwen Wu; Ying Chen
Original assignee: Thomson Licensing
Priority date: 2006-07-17
Filing date: 2006-07-17
Publication date: 2008-02-21
Also published as: US8761249B2; JP4991851B2; US20090285283A1; DE602006019006D1; CN101507284B; EP2041983A1; JP2009544198A; CN101507284A; KR20090052849A; EP2041983B1; EP2041983A4; KR101196975B1

Abstract

A highly accurate reproduction of visual intensity and contrast rather than the conventional 8-bit color depth is more and more used, motivating the development of an enhanced dynamic range called high bit-depth. A method for encoding a first, low bit-depth image (IMBL) of M bit RGB pixels and a second, high bit-depth video image (IMEL) of N bit RGB pixels, M<N, both having same spatial resolution, comprises generating a first transfer function (LUTinter) based on color histograms (HistBL, HistEL) of the first and the second video image, generating a second transfer function (LUTfinal) based on the first transfer function (LUTinter), applying (TR) the second transfer function (LUTfinal) to the first video image, wherein a transformed first video image (IMTR, BL) is generated, calculating and encoding the residual (IMres) f encoding the first video image (IMBL) and transmitting the encoded first video image (IMBL), the second transfer function (LUTfinal) and the encoded difference.

Description

Method and apparatus for encoding video color enhancement data, and method and apparatus for decoding video color enhancement data

Field of the invention

This invention relates to a method and an apparatus for encoding video color enhancement data. Further, the invention relates to a method and an apparatus for decoding video color enhancement data.

Background

In recent years, highly accurate reproduction of visual intensity and contrast rather than the conventional 8-bit color depth is more and more used in many fields, such as medical imaging, high-quality video-enabled computer games and professional studio and home theatre related applications. This process motivates the development of an enhanced dynamic range, which is called high bit-depth herein, for the convenience of comparison with the conventional 8-bit color depth. On contrast to the fact that advances in electronic sensors, processors and storage devices have resulted in very high pixel resolutions for both capturing and display devices, the color capacities of digital imaging systems have evolved in a very slow pace. 8-bit color depth has been dominant for capturing and display devices since the first generation of digitalized visual contents emerged.

Color bit-depth scalability is potentially useful considering the fact that in a considerably long period in the future standard 8-bit and higher-bit digital imaging systems will simultaneously exist in consumer marketplaces . Different color bit-depths are of particular importance for example for terminal display devices during multimedia content deliveries .

Summary of the Invention

The present invention provides methods and device for enabling video color space scalability. According to one aspect of the invention, a method and a device for encoding a color enhancement layer is provided, which is encoded differentially. Another aspect of the invention is a method and a device for decoding a video signal to obtain either a conventional color bit-depth image or an enhanced color bit-depth image.

In principle, the encoding aspect of the invention comprises the following steps: generating a transfer function, for example in the form of a look-up table (LUT) , which is suitable for mapping input color values to output color values, both consisting of 2^M different colors, applying the transfer function to a first video picture with low or conventional color bit-depth, generating a difference picture or residual between the transferred video picture and a second video picture with higher color bit-depth (N bit, with N>M; but same spatial resolution as the first video picture) and encoding the residual. Then, the encoded first video picture, parameters of the transfer function (e.g. the LUT itself) and the encoded residual are transmitted to a receiver. The parameters of the transfer function may also be encoded. Further, the parameters of the transfer function are indicated as such. The first and second images can be regarded as being a color base layer and a color enhancement layer, respectively.

In particular, the transfer function may be obtained by comparing color histograms of the first and the second video pictures, for which purpose the color histogram of the first picture, which has 2^M bins, is transformed into a "smoothed" color histogram with 2^N bins (N>M) , and determining a transfer function from the smoothed histogram and the color enhancement layer histogram which defines a transfer between the values of the smoothed color histogram and the values of the color enhancement layer histogram. The described procedure is done separately for the basic display colors red, green and blue.

According to the decoding aspect of the invention, a method for decoding comprises extracting from a bit stream video data for a first and a second video image, and extracting color enhancement control data, furthermore decoding and reconstructing the first video image, wherein a reconstructed first video image is obtained having color pixel values with M bit each, and constructing from the color enhancement control data a mapping table that implements a transfer function. Then the mapping table is applied to each of the pixels of the reconstructed first video image, and the resulting transferred video image serves as prediction image which is then updated with the decoded second video image. The decoded second video image is a> residual image, and the updating results in an enhanced video image which has pixel values with N bit each (N>M) , and therefore a higher color space than the reconstructed first video image. The above steps are performed separately for each of the basic video colors red, green and blue. Thus, a complete video signal may comprise for each picture an encoded low color-resolution image, and for each of these colors an encoded residual image and parameters of a transfer function, both for generating a higher color-resolution image. Advantageously, generating the transfer function and the residual image is performed on the R-G-B values of the raw video image, and is therefore independent from the further video encoding. Thus, the low color-resolution image can then be encoded using any conventional encoding, for example according to an MPEG or JVT standard (AVC, SVC etc.) . Also on the decoding side the color enhancement is performed on top of the conventional decoding, and therefore independent from its encoding format.

Thus, devices with lower color bit-depth display capability (e.g. 8-bit displays) need only decode the color base layer having lower color bit-depth, while advanced devices with enhanced color bit-depth display capability (e.g. 12-bit displays) may decode also the color enhancement layer and the transfer tables for red, green and blue, and generate pictures with full color space utilization.

The invention is related to a new type of video scalability, namely color bit-depth scalability, and provides a color bit-depth prediction solution that can be applied in the inter-layer prediction of a color bit-depth scalable CODEC to improve the coding efficiency. The advantage of using a transfer function to generate a prediction image before generating a residual image is that the encoding is more efficient, because the prediction image matches the respective color enhancement layer image better. For example, a value for a particular green tone, which is described by an 8-bit value of 9Aheχ i-ⁿ the color base layer, may map to sixteen different 12-bit color values in the color enhancement layer, from 9A0_heχ to 9AF_hex- While in one picture one of these sixteen values may dominate in the color enhancement layer, it may be another value in another picture. Thus, the invention enables optimized encoding of the color enhancement layer.

Advantageous embodiments of the invention are disclosed in the dependent claims, the following description and the figures .

Brief description of the drawings

Exemplary embodiments of the invention are described with reference to the accompanying drawings, which show in

Fig.l the structure of color bit-depth scalable video coding and decoding;

Fig.2 a flow diagram of the encoding process;

Fig.3 a flow diagram of smoothed histogram specification;

Fig.4 an example of histogram smoothing;

Fig.5 a flow diagram of the decoding process; and

Fig.6 an overview of a complete enhancement layer encoder. Detailed description of the invention

Fig.l shows the general structure of color bit-depth scalable video coding and decoding. The encoder ENC receives each video image twice, in conventional color resolution of M bit and in enhanced color resolution of N bit (N>M) . Instead of generating two independent separate video streams, a color base layer and a color enhancement layer are generated, where the color enhancement layer is dependent on the color base layer. As shown in Fig.l, an inter-layer prediction is performed in both the encoder and the decoder for predicting colors of higher bit-depth.

The base layer encoder block in Fig.l contains all the conventional encoding steps, for example any spatial and/or temporal redundancy reduction such as generating separate luminance and chrominance values, DCT transformations etc. On the decoder side, the base layer decoder performs the respective decoding, "base layer" and "enhancement layer" refer only to color bit-depth. Advantageously, any video encoding and decoding method can be used for the base layer, independent from the color enhancement process enabled by the invention. The output of a base layer decoder is suitable for a conventional display device, for example with M bit color resolution, while the output of an enhancement layer decoder is suitable for a high-quality display that is capable of displaying 2^N (N>M) different color values for each of the red, green and blue partial image. Examples for N are 10, 12 or 16 bits if M is e.g. 8, but in principle all combinations of M and N (N > M) can be implemented. Like other type of scalability, such as quality scalability, spatial scalability and temporal scalability, the inter-layer prediction impacts significantly the coding efficiency.

The present invention proposes a spatially uniform approach for color bit depth prediction based on smoothed histogram specification. Consider two images that describe the same scene. For the two images, the corresponding pixels (here the "corresponding pixels" mean two pixels that belong to the two images respectively but have the same coordinates in the image coordinate system) refer to the same scene location. The only difference between the corresponding pixels is the color bit depth. Assume each color of one image is encoded with code words of M bit length while the other image it is encoded with code words of N bit length, with M < N. The task of inverse tone mapping is to generate a predicted version of the N-bit image from the M-bit image, following the criterion that the difference between the predicted N-bit image and the original N-bit image is minimized. The difference between the predicted N-bit image and the original N-bit image may be measured by any method, for example PSNR, which is widely accepted and used in the field of video compression. Further, in this case the most important aspect is how effective the predicted image works for the following residual data compression, rather than how the predicted image looks.

Fig.2 shows the general encoding procedure of one embodiment of the invention. A base layer image IM_BL is transformed TR using a look-up table LUT_finai that implements a mapping function. The result of the transformation TR is a transformed base layer image IM_TR,_BL- The look-up table LUT_fi_nai is generated from another, intermediate look-up table LUTinter, which is generated from color histograms of the base layer picture, preferably the reconstructed base layer picture, and of the enhancement layer picture. Due to the different color resolution of these pictures, their color histograms have not only a different amount of bins (one for each possible color) , but the distribution of the values among the bins of the enhancement layer picture may be different for each base layer color, and for each picture of a sequence. If for example the base layer has M bit and the enhancement layer has N bit, each bin of the base layer histogram corresponds to 2^N~M bins of the enhancement layer histogram, i.e. each base layer pixel value would conventionally serve to predict one out of 2^N~M pixel values. This prediction is optimized by considering the color histograms of the two pictures. The intermediate look-up table LUTi_nter implements a histogram transformation function describing a transformation between the color histogram Hist_EL of the enhancement layer picture and a color histogram SmHist_BL generated from the base layer picture IM_BL _r ^as described below.

The transformed base layer picture IM_TR,_BL is used to predict the enhancement layer picture IM_EL , and the difference or residual Δ is calculated, encoded IM_res and transmitted. All these processes are separate for R, G and B.

Fig.3 shows details of the proposed spatially uniform inverse tone mapping. For better understanding, the base layer (M-bit) image is named "input image" NI while the enhancement layer (N-bit) image is renamed "desired image" DI. One basic idea is to apply the algorithm of histogram specification¹ to NI and DI. A lookup table is generated to convert NI such that it has a particular histogram, as specified by the histogram of DI.

However, there are two major drawbacks when employing the classical histogram specification for color bit depth prediction. First, because the histograms have discrete values, the converted NI does not have exactly the same histogram as DI. Instead, the histogram of the converted NI is an approximation to that of DI. Second, the different bin size of the two histograms (caused by the different bit depth) deteriorates the matching between them (bin size means the number of the levels of each color component e.g. the bin size of 8-bit images is 256) . This is a particular drawback in the considered color bit depth prediction. For instance, in the case that NI is simply bit-shift from DI, the PSNR of the converted NI which was obtained by histogram specification can often be lower than that obtained by simple inverse bit-shift.

To overcome these drawbacks, we propose to "smooth" the histogram of NI before it is specified. The smoothed histogram SmHist_BL is of the same bin size as the histogram Hist_EL of DI, which serves as the desired histogram for the process of histogram specification. The classical histogram specification algorithm is applied on the smoothed histogram and the desired histogram. Finally, a post processing called "Probability mapping" is carried out to obtain the LUT. The flowchart of the smoothed histogram specification is shown in Fig.3. p_x(x_k),k = 0,1,---,2^ -1

¹ R. C. Gonzalez and R. E. Woods, "Digital image processing", 2^nd edition, pp. 94-102, Prentice Hall, 2002. represents the histogram of NI Hist_BL, and the histogram of DI Hist_EL is described by p_∑(z_k),k = 0,l,---,2^M -1 .

In Fig.3, for the low-bit (M-bit) image NI, the histogram of one color channel is p_x(x) . The histogram is smoothed by using eq. (5), and a smoothed histogram p_x ' (x) results, which has the same bin size as the histogram of the high- bit (N-bit) image DI does. The values are accumulated, resulting in s_k (k=l,...,16 for the example of Fig.3) . It can be expressed as a function F(x_k') .

The high-bit (N-bit) image DI has the histogram p_z(z) of the same color channel, which is accumulated to result in v_k (same k as above) . It can be expressed as a function G(z_k) . Then the distribution v-, is determined, which gives for each value of s_k the best-matching value vi (from the v_k values of the high-bit image DI) . This distribution v-, sets up a transformation from the values s_k (therefore x _k' ) to the values v_k, and it is the inverse of the transform G.

In the following, the process of histogram smoothing is described. The goal of histogram smoothing is to "stretch" the input histogram so that it will have the same bin size as the desired histogram. This is a prerequisite for the histogram specification process. Following the denotations in Fig.3, the value of p_x(x_k) is uniformly distributed into a corresponding range of [x_k,x_k ^' +1,■ ••,x_k + 2⁽-^N"M) -1] , as described in eq. (5) :

0, otherwise To ensure "uniform distribution", un-normalized histograms are used. A simple example of a smoothed histogram is shown in Fig.4. Fig.4 a) shows a color histogram of a base layer input picture with a color bit depth of 2 bit, i.e. 2²=4 possible color values of a particular color, e.g. red. The histogram smoothing process "stretches" the histogram to a color bit depth of 4 bit, i.e. 2⁴=16 possible color values. The number of values in a single bin of the input picture is the number of pixels within the input picture that have this particular color value. E.g. a value of 40 in the left-most bin corresponding to the value 0 means that 40 pixels of the input picture have a "red" value of 0 (in this histogram refers to the color "red") . The smoothing process distributes these values to 2^N~M corresponding bins of the smoothed histogram. In this example with 2^N~M=4 the first four bins of the smoothed histogram correspond to the first bin of the original histogram. According to eq.(5), the 40 values of the original histogram are evenly distributed to the four corresponding bins of the smoothed histogram, as shown in Fig.4 b) .

The idea behind the smoothing is that it improves the continuity of the input histogram, so that the histogram specification will be more efficient. In detail, we write down the alternatives of eq. (1) , (2) and (4) shown in Fig.3 for continuous gray level images:

s = F(x)= [p_x(μ)du (1')

y = G^~l(s) = G^~l[F(x)] (4') In the case of continuous gray levels, the input histogram can be specified as exactly the same as the desired histogram. However, as aforementioned, for discrete gray levels only an approximation to the desired histogram can be achieved. Furthermore, as the number of the gray levels is approaching infinity, the approximation is approaching exact match. Therefore, in theory the smoothness procedure is equivalent to increasing the sampling rate during the discretization of the input histogram, and it outperforms the direct histogram specification between two histograms that have different bin size.

The smoothed histogram is just an intermediate step in this algorithm. There is not an image corresponding to it.

The following describes the process of Probability Mapping. Once the classical histogram specification algorithm has been applied to the smoothed histogram p_x(x) and the desired histogram p~(z) , an intermediate LUT y_k - LUT_inter(x_k) is generated. The next problem is to choose the unique mapped value of x_k from the multiple mapped values of its corresponding range [x_k,x_k +!,^■■■,x_k +2⁽'^{N~M^>} -I] . Exemplarily, two different criteria are proposed as criterion for probability mapping, as described in eq.(6) and (7) below:

y_k e[x_k,x_k +l,- - -,x_k +2^{N~M) -1]}, ( 6)

y_k

-1]}J , ₍ 7 )

where y_k is the mapped value of χ_k . A final LUT

y_k = LUT_βm!(x_k), k = 0,l,-,2^M~l, y_k ≡{z_hl = 0X-,2^N -1} is generated to map the original histogram to the desired histogram. Eq. (6) says that among the 2^N~M values that x_k corresponds to, we select the one that has the highest value in the desired histogram p_z(y_/) •

Eq. (7) says that among the 2^N~M values that x_k corresponds to, we use the nearest integer less than or equal to their mean as the finally mapped value .

The LUT_mter is a "one-to-one" mapping, because it maps the smoothed histogram of the input image to the histogram of the desired image. However, if we consider the straight histogram of the input image, we can see that each x_k corresponds to 2^N~M values. The task of "Probability Mapping" is to choose only one value from the 2^N~M values as the finally mapped value of x_k. Hence, the LUT_finai is still a "one-to-one" mapping: it maps each value x_k to a value y_k. The mapping of the LUT is invertible because only 2^M values of the total values (2^N) of the desired image have the corresponding x_k.

When the above-described algorithm is applied on an input image and a desired image, the histograms of both images are calculated. Then the input histogram is smoothed, resulting in the "smoothed histogram". After the remaining steps (classical histogram specification and probability mapping) are finished, a final LUT is generated to map the levels of the input histogram to the levels of that of the desired histogram. Then the predicted image is generated by applying the LUT to each pixel of the input image .

Fig.6 shows an overview of a complete enhancement layer encoder, with separate branches for R, G and B images. The base layer is conventionally encoded. The invention is usable for video encoders and video decoders, and particularly for encoding / decoding a color base layer and a color enhancement layer of the same spatial resolution with optimized efficiency.

Claims

1. A method for encoding a first video image (IM_BL) consisting of RGB pixel values having M bit each and a second video image (IM_EL) consisting of RGB pixel values having N bit each, with M<N, wherein the first and second video image have the same spatial resolution, comprising the steps of generating a first transfer function (LUT_lnter) having 2^N possible input values and being based on color histograms (Hist_BL, Hist_EL) of the first and the second video image; generating a second transfer function (LUT_finai) based on the first transfer function (LUT_lnter) , wherein the second transfer function has 2^M possible input values and comprises a function for mapping each of its input values to one out of

2 ^(N-^M) _correSp_Onding values of the first transfer function; - applying (TR) the second transfer function

(LUTf_mai) to the first video image, wherein the second transfer function is used for mapping the pixels, and wherein a transformed first video image (IM_TRjBL) is generated; - calculating the difference (IM_res) between the transformed first video image (IM_TRfBL) and the second video image (IM_EL) ; encoding the difference (IM_res) ; encoding the first video image (IM_BL) ; and - transmitting the encoded first video image (IM_BL) , the second transfer function (LUT_flnal) and the encoded difference (IM_res) •

2. Method according to claim 1, wherein the step of generating a first transfer function (LUT_inter) comprises the steps of generating a first color histogram (Hist_BL) of the reconstructed first image, the first color histogram having 2^M bins; generating a second color histogram (Hist_EL) of the second image, the second color histogram having 2^N bins; and - generating (sm) from the first color histogram (Hist_BL) a smoothed color histogram (SmHist_BL) having 2^N bins, wherein the values of each bin of the first histogram are distributed to 2 ^(N~M> bins, and wherein the first transfer function (LUT_inter) defines a transfer between the smoothed color histogram (SmHist_BL) and the second color histogram (Hist_EL) .

3. Method according to claim 1 or 2, further comprising the initial steps of encoding the first video image (IM_BL); and decoding the encoded first video image to obtain a reconstructed first video image (IM_BL,_rec) , wherein the step of applying (TR) the second transfer function refers to the reconstructed first video image.

4. Method according to any of claims 1-3, wherein the function for mapping each of the input values of the second transfer function (LUT_fα.nai) to one out of 2 ^(N-^M) corresponding values of the first transfer function is a probability mapping function.

5. Method according to claim 4, wherein the probability mapping function selects for each of the 2^M values of the second transfer function one value from the 2^N~M corresponding values of the first transfer function, being the one that has the highest value in the second histogram (Hist_EL) ■

6. Method according to claim 4, wherein the probability mapping function selects for each of the 2^M values of the second transfer function one value from the 2^N~M corresponding values of the first transfer function, being the one that is the nearest integer less than or equal to the mean value of said 2^N~M corresponding values .

7. Method for decoding a first and a second video image received in a bitstream (bs), wherein the first and the second video image have the same spatial resolution, comprising the steps of - extracting from the bitstream (bs) video data for the first and the second video image, and color enhancement control data; decoding and reconstructing (dec_Bi) the first video image, wherein a reconstructed first video image (IM_BL) is obtained having color pixel values with M bit each; constructing (dec_LϋT) a mapping table (LUT_finai) from the color enhancement control data; applying (TR_dec) the mapping table (LUT_finai) to the pixels of the reconstructed first video image

(IM_BL) _r wherein a transformed first video image (IM_TR,_BL) is generated having a different color histogram than the reconstructed first video image; decoding the second video image (IM_res) i wherein the second video image comprises pixel difference values; and adding the decoded second video image (IM_res) to the transformed first video image (IM_TR,_BL) _r wherein a reconstructed second video image (IM_EL) is obtained having color pixel values with N bit each, with M<N .

8. Method according to claim 7, wherein the transformed first video image (IM_TR/BL) has color histograms for red, green and/or blue that differ from the respective color histograms of the reconstructed first video image (IM_BL) •

9. Method according to claim 7 or 8, wherein three separate mapping tables (LUT_finai) or mapping table areas are generated from the color enhancement control data, wherein each of the separate mapping tables or mapping table areas relates to one of the colors red, green and blue.

10. Apparatus for encoding a first video image (IM_BL) consisting of RGB pixel values having M bit each and a second video image (IM_EL) consisting of RGB pixel values having N bit each, with M<N, wherein the first and second video image have the same spatial resolution, comprising means for generating a first transfer function (LUTmter) having 2^N possible input values and being based on color histograms (Hist_BL, Hist_E]j) of the first and the second video image; means for generating a second transfer function (LUT_flnai) based on the first transfer function (LUTm_ter) r wherein the second transfer function has 2^M possible input values and comprises a function for mapping each of its input values to one out of 2 ^<N~M) corresponding values of the first transfer function; - means for applying (TR) the second transfer function (LUT_flnai) to the first video image, wherein the second transfer function is used for mapping the pixels, and wherein a transformed first video image (IM_TR,_BL) is generated; - means for calculating the difference (IM_res) between the transformed first video image (IM_TR,BL) and the second video image (IM_EL); means for encoding the difference (IM_res); means for encoding the first video image (IM_BL) ; and means for transmitting the encoded first video image (IM_BL) , the second transfer function (LUT_fmai) and the encoded difference (IM_res) •

11. Apparatus according to claim 10, wherein the means for generating a first transfer function (LUTmter) comprises means for generating a first color histogram (Hist_BL) of the reconstructed first image, the first color histogram having 2^M bins; means for generating a second color histogram (Hist_EL) of the second image, the second color histogram having 2^N bins; and means for generating (sm) from the first color histogram (Hist_BL) a smoothed color histogram (SmHist_BL) having 2^N bins, wherein the values of each bin of the first histogram are distributed to 2^(N~M) bins, and wherein the first transfer function (LUT_inter) defines a transfer between the smoothed color histogram (SmHist_BL) and the second color histogram (Hist_EL) .

12. Apparatus for decoding a first and a second video image received in a bitstream (bs) , wherein the first and the second video image have the same spatial resolution, comprising means for extracting from the bitstream (bs) video data for the first and the second video image, and color enhancement control data; means for decoding and reconstructing (dec_BL) the first video image, wherein a reconstructed first video image (IM_BL) is obtained having color pixel values with M bit each; means for constructing (dec_LUτ) a mapping table (LUTf_mai) from the color enhancement control data; means for applying (TR_dec) the mapping table (LUT_fλnai) to the pixels of the reconstructed first video image (IM_BL) , wherein a transformed first video image (IM_TR,_BL) is generated having a different color histogram than the reconstructed first video image; means for decoding the second video image (IM_res) , wherein the second video image comprises pixel difference values; and means for adding the decoded second video image (IM_res) to the transformed first video image (IM_TR,_BL), wherein a reconstructed second video image (IM_EL) is obtained having color pixel values with N bit each, with M<N .