Technical Field

The invention relates to the reconstruction of saturated regions of an image on the basis of color information coming from non-saturated color channels of these regions.

Background Art

Most existing cameras can only capture a limited range of illumination within a single exposure. Wrong selection of camera settings, combined with the physical constraints of camera sensors limit the maximum irradiance value which can be captured and stored as images for a given exposure time and amplifier gain. Therefore, scene irradiance values, which are beyond the maximum capacity of the camera sensor elements, cannot be accurately captured, resulting in over-exposed or color-saturated regions in the captured images.

Saturation of bright image regions of an image is of course undesirable as it visually degrades the image. Missing detail in bright image regions is particularly problematic in the context of high dynamic range (HDR) imaging, and in particular in reverse tone mapping. Recent developments in display capabilities suggest that consumer displays are likely to display HDR content in the near future, requiring solutions for preparing legacy content for this new display modality. As most current reverse tone mapping solutions detect and expand bright regions in the image, it is crucial that missing color information in these regions is recovered in a robust manner.

Image color saturation can take two forms, which are typically treated using significantly different methods. If all three color channels of colors of saturated parts of an image are clipped due to saturation, no robust color information is present in that part of the image and therefore correction methods either rely on probabilistic information obtained from the rest of the image to hallucinate the missing pixels or on filling in the missing information based on a global prior part of this image. On the other hand, if only one or two of the RGB components are saturated in a part of the image, it is possible to reconstruct missing information in clipped color channels using the color information in non- clipped channels or by combining it with information from nearby image regions. The invention concerns this second form of image color saturation.

Because of the strong correlation between color channels specifically in the RGB color space (see E. Reinhard and T. Pouli, "Colour spaces for colour transfer," in IAPR Computational Color Imaging Workshop, ser. Lecture Notes in Computer Science, R. Schettini, S. Tominaga, and A. Tremeau, Eds. Springer Berlin Heidelberg, 201 1 - invited paper, vol. 6626, pp. 1-15), most methods related to the second form above process images in the RGB color space. In the simplest case, the global correlation between color channels can be determined in the form of a global Bayesian model and used in the process of recovery of the missing color information due to saturation (see X. Zhang and D. H. Brainard, "Estimation of saturated pixel values in digital color imaging," JOSA A, vol. 21 , no. 12, pp. 2301-2310, 2004).

Correlation between color channels can be exploited by using information in undipped color channels within the same spatial region. In S. Z. Masood, J. Zhu, and M. F. Tappen, "Automatic correction of saturated regions in photographs using cross-channel correlation", Computer Graphics Forum, vol. 28, no. 7. Wiley Online Library, 2009, pp. 1861-1869, the Bayesian modelling method of Zhang et al.(see above) is extended and used the relationship between pixels and their neighbourhood to estimate ratios between RGB channels.

This approach relies on the minimization of two cost functions, which significantly increases its computational complexity. In addition, the use of neighbourhood information may result in incorrect hue values

An object of the invention is to avoid the aforementioned drawbacks.

Summary of invention

The method of conversion of a saturated image into a non-saturated image also relies on the correlation between color channels but opts for simple statistics in order to recover clipped color information and to enforce spatial consistency of colors across different lines of the image. Additionally, in contrast to previous methods, this method is designed to operate on line profiles of the images, making it well suited for hardware applications and display-side processing.

This method relies on the fact that in over-exposed image areas, the RGB components tend to not clip at the same spatial position. Based on the assumption that the color channels are spatially, strongly correlated, this method reconstructs the over-exposed areas of one color channel by considering the variation within undipped color channel(s) at the same spatial positions.

A subject of the invention is a method of conversion of a saturated image into a non-saturated image comprising:

- segmenting lines of adjacent pixels of said saturated image into segments of adjacent pixels the colors of which have at least one color channel in a linear RGB color space which is saturated, and,

for each of these saturated segments,

- among the different color channels providing color information in said linear RGB color space for the pixels of said segment, selecting at least one color channel that is not saturated over said segment,

- for each pixel of said segment, replacing the color information of said pixel that is provided by said at least one saturated color channel of said segment by a reconstructed non-saturated color information computed by using at least the color information provided by the selected at least one non-saturated color channel,

thus providing said converted non-saturated image.

Preferably, the method according to the invention is implemented line by line, one line with saturated pixels after the other.

Preferably, said reconstructed non-saturated color information replacing the color information of a pixel that is provided by said at least one saturated color channel is computed by adding said provided color information of said pixel and the difference between the color information provided by the selected at least one non-saturated color channel for said pixel and a baseline color value computed for the segment of said pixel. Preferably, said baseline color value of a segment is computed as the minimum color value provided by the reference color channel of this segment over all pixels in this segment and over pixels adjacent to this segment in the neighbouring lines the hues of which are closer than a proximity threshold to a representative average hue of this segment.

Preferably, said reconstructed color information replacing the color information of a pixel that is provided by said at least one saturated color channel is computed by using also color statistics based on color information related to other pixels belonging to the same line of the segment of said pixel and/or to neighboring lines. Preferably, these neighboring lines are adjacent one to another.

A subject of the invention is also a method of conversion of a saturated image into a non-saturated image, the color information of each pixel of said saturated image being provided through three different color channels such as to express this color information into an RGB color space, said method comprising the steps of:

- segmenting lines of adjacent pixels of said saturated image into segments of adjacent pixels the colors of which have at least one color channel which is saturated, and,

for each of these saturated segments,

- over the different color channels providing color information for the pixels of said segment, selecting at least one color channel that is not saturated,

- for each pixel of said segment, replacing the color information of said pixel that is provided by said at least one saturated color channel by a reconstructed color information computed by using at least the color information provided for the same pixel by said at least one not-saturated color channel, resulting in a replicated not-saturated color channel,

thus providing said converted non-saturated image.

The method can be summarized as follows:

- segmenting lines into segments of adjacent saturated pixels,

and, for each of these saturated segments,

- selecting at least one color channel that is not saturated,

- for each pixel of said segment, replacing the color information of said pixel that is color saturated by a reconstructed non-saturated color information computed by using the color information provided by the selected at least one non- saturated color channel.

A subject of the invention is also an image conversion device for the conversion of a saturated image into a de-saturated image, comprising:

a module configured to process each line of pixels of said saturated image having at least one pixel with saturated color, one line after the other,

said module being configured to segment said line into at least one segment of adjacent pixels the colors of which have at least one color channel in a linear RGB color space which is saturated,

said module comprising itself a module configured to process each saturated segment of a line, said module being configured to select at least one color channel that is not saturated over said segment among the different color channels providing color information in said linear RGB color space for the pixels of said segment,

said module comprising itself a module configured to process each pixel of a segment, said module being configured to replace the color information of said pixel that is provided by the at least one saturated color channel of said segment by a reconstructed non-saturated color information computed by using at least the color information provided by the at least one non-saturated color channel of said pixel selected by the module configured to process saturated segment.

Preferably, said module configured to process each pixel of a segment is configured to compute said reconstructed non-saturated color information replacing the color information of said pixel that is provided by said at least one saturated color channel by adding said provided color information of said pixel and the difference between the color information provided for this pixel by the selected at least one non-saturated color channel and a baseline color value computed for the segment of said pixel.

A subject of the invention is also an electronic consumer device comprising the image conversion device according to the invention. Preferably, this electronic consumer device is a TV set, a cell phone, a tablet, a set-top-box, a gateway, or a camera. Brief description of drawings

The invention will be more clearly understood on reading the description which follows, given by way of non-limiting example and with reference to the appended figures in which:

- figure 1 is a flow chart illustrating a main embodiment of the method according to the invention;

- figure 2 illustrates, for each color channel Red, Green and Blue, the mask S which is obtained after the first step of the embodiment of figure 1 ;

- figure 3 illustrates the segmentation of a line of the saturated image into different segments s ₂ , s ₃ , ... of color saturated pixel grouped based on their hue values after the second step of the embodiment of figure 1 ;

- figure 4 illustrates the variation of the color information provided by the three color channels R, G, B over all pixels of the same line as shown on figure 3, where the same saturated segments s^ s ₂ , s ₃ , ... are identified by the

saturation of the color information provided by at least one of the three color channels R, G and B ;

- figure 5 illustrates, in a segment of figure 3 and 4, how the difference between the color value provided by the reference color channel for a pixel of this segment and a baseline color value p _si computed for this segment is added to the color value of this pixel that is provided by the saturated color channel of this segment, according to the fourth step of the embodiment of figure 1 ;

- figure 6 illustrates the variation of the color information provided by the three color channels R, G, B over all pixels of the same line as shown on figure 3, resulting from the desaturation of all color channels as illustrated on figure 5, at the end of the fourth step of the embodiment of figure 1 .

Description of embodiments

The different steps of a main non-limiting embodiment of the method according to the invention will now be described in reference to figure 1 . These different steps may be provided through the use of dedicated hardware as well as hardware capable of executing software in association with appropriate software, forming together an image conversion device. The hardware may notably include, without limitation, digital signal processor ("DSP") hardware, read-only memory ("ROM") for storing software, random access memory ("RAM"), and non-volatile storage. The image conversion device can be part of an electronic consumer device as, for instance, a TV set, a cell phone, a tablet, a set-top-box, a gateway, or a camera.

An input 8-bit RGB saturated image, denoted I, of size M x N pixels, and R,G,B color channels, is provided to the image conversion device. If this image conversion device is part of a cell phone, a tablet or a camera, this image may be directly provided through the capture of an image. If this image conversion device is part of a TV set or a set-top-box, the image may be provided by the image receiver integrated in them.

If necessary, the gamma correction is first removed in a manner known per se from the saturated image (we assume a 2.2 gamma) such as to obtain an image with color data expressed in the linear RGB color space of a display device. This display device could be any kind of actual or virtual device, as a LCD, an OLED display, a projector, or any device complying with a standard as, for instance, ITU 709. To each color of a pixel of the image, correspond three color values that represent the color coordinates of this color in the linear RGB space. Each color value is related to a color channel, R, G or B.

In the first step of the embodiment, a saturation mask S is constructed for each color channel of the image in a manner known per se in order to mark over-exposed areas of the image, more precisely to mark each pixel of this image having at least one color value which is saturated in at least one of the three color channels. These masks are illustrated on figure 2. For each color channel, the saturation mask S is for instance calculated through the following equation 1 using a saturation color value t which may be common to all color channels, or specific to each color channel:

1 otherwi se

.... where l _m (n, c) is the value of a color in the color channel c for a pixel belonging to a line m of the image and to a column n of the image, where m belongs to the interval [1 ;M] and indicates a line, where n belongs to the interval [1 ;N] and indicates a column, and where c belongs to the group of color channels R, G, and B and indicates the color channel. For a 8-bit RGB image, the threshold t is for instance common to all color channel and set to 235 (instead of 2 ⁸ ) to account for noise.

According to the above definition of the value assigned to each pixel of a saturation mask S of a color channel c:

- the value 0 for a pixel at position m,n means that the value of the color channel c associated with this pixel is saturated and that a color information related to this color channel needs to be reconstructed for this pixel;

- the value 1 for this pixel means that the value of the color channel c

associated with this pixel is not saturated and that no color information needs to be reconstructed.

In the second step of the embodiment of the method according to the invention, the lines of adjacent pixels of the image are segmented into segments of adjacent pixels for which the corresponding values of the mask are 0. All pixels of any of those segments are clipped, i.e. have colors with at least one color channel being saturated.

In an ideal scenario, both the segmentation into segments of adjacent saturated pixels and the computation as described of a reconstructed non- saturated color information for each saturated saturated pixel (see fourth step below) would be performed on the whole image. However, in most hardware- based applications, notably when the image processing device is integrated into an electronic consumer device, processing a saturated image as a whole would be computationally impractical. To resolve this, it is preferred, as described below, to process separately the different lines in which at least one segment of adjacent saturated pixels has been identified, i.e. one line after the other. Such an iterative process advantageously requires only access to the color values of the line to process, plus, as explained below, the preceding and following a lines, i.e. all lines in the interval [m-a,m+a]. When having low computational resources, a may be set to 0 or 1 , while larger values are preferred to increase the robustness of the method. Of course, lines of pixels of the image are generally horizontal and, then, columns are vertical. But lines of pixels may also be vertical and, then columns are horizontal. More details will now be given about one way to segment lines of pixels of the saturated image into segments of adjacent color saturated pixels using the saturation masks built in the first step above. Without departing from the invention, other ways of segmentation of lines into saturated segments can be used to implement the invention, as for instance those that do not require a mask.

In this embodiment, to avoid introducing discontinuities and to ensure that similar clipped regions of the image are treated in a consistent way, in each line of the image, clipped pixels are grouped into segments based on their hue values and the pixels of each group (i.e. segment) are assigned a same pixel label as described below.

The hue H _m (n) for a pixel belonging to a line m and to a column n is defined in a usual way in the CIELab color space as H _m (n) = arctan(b _m (n); ^a m( ⁿ )) _> where a and b are the CIELab chromatic components in the Lab color space corresponding to the above color values in the RGB color space of this pixel. The segmentation process of a line m of pixels of the image is then performed such that clipped and adjacent pixels with hue values H _m (n) closer than a given threshold t _c (for instance, t _c = 30°) are considered as part of a same segment of this line and are therefore assigned a same label s _{m i} . Therefore, the label s _{m i} of a pixel belonging to a line m and to a column n can be defined according to equation 2 as follows:

S _{m i} if ΔΗ(τη, ) < t _c and S _m (n, c)≡ 1

Sm,(i+i) if ΔΗ(τη, η)≥ t _c and S _m n, c)≡ 1

0 otherwise

where i belongs to the interval [1 ,L _m ], L _m being the number of clipped segments in line m,

where the function Δ _Η (ηη,η) for a pixel position m,n is defined in equation 3: AH(m, n) = min(|H _m (n) - H _m (n + 1) L 360 - \H _m (n) - H _m (n + 1) | ) eq.(3) where H _m (n) is the hue of this pixel, as defined above.

At the end of this segmentation step of a line m comprising color saturated pixels, as illustrated on figure 3 and 4, there are L _m segments s ₂ , s ₃ , s _Lm of color saturated pixels in this line, the color values of them need to be reconstructed for a same at least one color channel. On figure 4, the color channel of the RGB color space that is saturated can be identity in each segment s ₂ , s ₃ , .... The grouping process of clipped pixels with similar hue values in each line of the image is optional but preferred, because it ensures that pixels of similar hues will be treated similarly, and therefore will likely obtain a more coherent appearance after reconstruction of color information for desaturation.

In a third step of the embodiment, for each of the saturated segments s^ s ₂ , s ₃ , s _Lm of the line m, at least one color channel that is not saturated is selected among the different color channels providing color information for the pixels of this segment. This selected color channel is named reference color channel c _ref because it will be used as described below for the reconstruction of color information for the pixels of this segment. This reference channel c _ref is selected among the undipped color channels of a segment, according to a rule defined in table 1 below, based on the identification of the saturated color channel of this segment, using for instance the color information provided by the different channels as shown on figure 4.

Table 1

According to this rule, if only one color channel is clipped in a segment (cases 1 , 2 and 3 in table 1 ), one of the undipped color channels is chosen as the reference color channel for the desaturation correction. When possible, the green channel is preferably favoured in the choice, as this channel was found to be a better predictor for the remaining channels. This stems from its high correlation with image luminance due to the color filter array typically used in cameras. In cases where two color channels are saturated (cases 4, 5 and 6 in table 1 ), the only undipped color channel is selected as the reference color channel to be used for the desaturation correction. Finally, if all color channels are saturated in the segment, a simple luminance expansion is applied in accordance to the corrections applied to the rest of the image (see below).

In a fourth step of the embodiment, in each saturated segment of a line as defined above, the color value of each pixel that is related to the saturated color channel idenfified for this segment is replaced by a reconstructed non- saturated color value that is computed by using at least the color value of the same pixel that is related to the reference color channel of this segment as selected in the third step above.

In this step, the saturated colors of pixels of each detected segment are corrected according to the information given by the reference color channel c _ref selected for this segment.

The color correction for the desaturation of the pixels of a segment S of a line m having one or two saturated color channels will now be described for one of these saturated color channel(s) of this segment. For this correction, some color information provided by the reference color channel c _ref of this segment will be transferred to the saturated color channel of this segment as described below. To achieve this correction without introducing inconsistencies between similar regions of the image (e.g. multiple highlights on the same object), a hue consistency constraint is used in addition to the hue-based segmentation into segments of pixels of the second step above. This type of hue constraint is optional but preferred.

The hypothesis is that variations of color values related to the reference color channel c _ref of a segment Sj can be used to reconstruct missing color information in another color channel for pixels within this segment s^ However, color values cannot be directly transferred from a reference color channel to a saturated color channel, as the saturated channel along this this segment and the non-saturated channel along the same segment do not provide the same levels of values. One is higher than the other. Therefore, as described below, a scaling operation is preferably performed for this transfer.

To solve this scaling issue, a baseline color value p _si related to the reference color channel c _ref of the segment S to be desaturated is computed. Such a computation is detailed below. Then, the transfer is performed as illustrated in figure 5, i.e. for each pixel of this segment s^ the difference between the color value of the reference color channel c _ref for this pixel and the baseline color value p _si computed for this segment is added to the color value of the saturated color channel for the same pixel.

To compute the baseline color value p _si used for the scaling and related to the reference color channel c _ref of a saturated segment S of a line m, a representative average hue H _si of this segment S is first calculated in the CIELab color space for all pixels of this segment s^ Then, the difference AH(k,n) between this average hue H _si of this segment S and the average of hues H _k (n) of pixels at positions k,n , where k e {m-a, m+a}, i.e. in the line m as well as in the a preceding and following lines is computed in a similar way as in equation 3 above. Then, the baseline color value p _si is then computed according to the following equation 4:

p _s . = m _{k e[m} _ _{a m + a]} (I _k (n, c _ref )) Vk\ H(k, n) \ < t _hue (Eq.4)

where k e {m-a, m+a}, where preferably, t _hue equal approximately 20°.

It means that, when a = 0, the baseline color value p _si is computed as the minimum color value provided by the reference color channel c _ref over all pixels in the segment S whose hue is close enough to H _si according to a criteria based on a proximity parameter t _hue defined by \AH m, n) \ < t _hue .

It means that, when a > 0, the baseline color value p _si is computed as the minimum color value provided by the reference color channel c _ref over all pixels in the segment S of the line m, and over pixels adjacent to this segment in the in the neighbouring lines k, where k e {m-a, m+a}, whose hues are close enough to H _si according to a criteria based on a proximity parameter t _hue defined by vk\AH(k, n) \ < t _hue . The above color correction for the desaturation of the color values of a saturated color channel for all pixels of a segment of a line of the image is then repeated for another color channel, if any. In segments where all three color channels are clipped, color values of the pixels of these segments are expanded according to corrected adjacent color values of pixels on the same line in order to obtain the local maximum color value. This ensures that fully clipped pixels are consistent with the corrections applied to one or two color channels of the other clipped segments of the same line.

The above color correction process is then reiterated for each segment of this line. The variation of color information provided by the three color channels R, G, B over all pixels of this line that is then obtained is illustrated on figure 6.

Then, the process described above for a line of pixels is reiterated for all lines of the image comprising pixels with saturated colors.

After all pixels of the image are desaturated, a desaturated image Γ is obtained, which is then gamma corrected in a reverse manner compared to the above beginning of the process.

Optionally, to still improve the robustness of the method as described above, the image is processed both horizontally and vertically. It means that, after a first processing as described above, the lines of the saturated image becomes the columns, and the columns of this saturated image becomes the lines, and the desaturation process is then repeated a second time on this saturated image, providing a second desaturated image. The final image is computed by taking, over the two desaturated images, the maximum color value in each color channel for each pixel. To avoid creating discontinuities near saturation boundaries, the corrected image is preferably alpha-blended with the uncorrected input using a simple ramp for input pixel values in the interval [t; 255].

It is to be understood that the invention may be implemented by an image conversion device in various forms of hardware, software, firmware, special purpose processors, or combinations thereof. The invention may be notably implemented as a combination of hardware and software. Moreover, the software may be implemented as an application program tangibly embodied on a program storage unit. The application program may be uploaded to, and executed by, a machine comprising any suitable architecture. Preferably, the machine is implemented on a computer platform having hardware such as one or more central processing units ("CPU"), a random access memory ("RAM"), and input/output ("I/O") interfaces. The computer platform may also include an operating system and microinstruction code. The various processes and functions described herein may be either part of the microinstruction code or part of the application program, or any combination thereof, which may be executed by a CPU. In addition, various other peripheral units may be connected to the computer platform such as an additional data storage unit and a monitor display device to display the images before and after desaturation.

In reference to figure 1 , this image conversion device comprises notably: - a module 1 configured to process each line of pixels of said saturated image having at least one pixel with saturated color, one line after the other,

said module being configured to segment said line into at least one segment of adjacent pixels the colors of which have at least one color channel in a linear RGB color space which is saturated, and,

said module comprising itself a module 2 configured to process each saturated segment of a line, said module being configured to select at least one color channel c _ref that is not saturated over said segment among the different color channels providing color information in said linear RGB color space for the pixels of said segment,

said module comprising itself a module configured to process each pixel of a segment, said module being configured to replace the color information of said pixel that is provided by the at least one saturated color channel of said segment by a reconstructed non-saturated color information computed by using at least the color information provided by the at least one non-saturated color channel c _ref of said pixel selected by the module 2 configured to process saturated segment.

This image conversion device can be notably part of an electronic consumer device according to claim 9 wherein said consumer device is a TV set, a cell phone, a tablet, a set-top-box, a gateway, or a camera As a conclusion, the method of conversion of a saturated image into a de- saturated image that has been described takes advantage of the spatial RGB channel correlation present in natural images. Instead of relying on complex image analysis, the method according to the invention employs simple statistics computed from a small set of lines of pixels of the image to enforce hue consistency between recovered regions. By considering hue information throughout the correction, the method produces natural results. Both the segmentation and the correction step of the method operate on a small number of lines of pixels, making this method well-suited for real-time, hardware applications, such as correcting images on the display side before expanding the dynamic range. It means that the correction device used to implement the invention can be incorporated into a TV set, a set-top-box, a gateway, a tablet, a cellphone, or any consumer device having a display. Despite the computational simplicity, this method out-performs existing partial saturation correction solutions according to several image quality and color difference metrics.

While the present invention is described above with respect to a particular embodiment, it is understood that the present invention is not limited to this embodiment. The present invention as claimed below therefore includes variations from this embodiment described herein, as will be apparent to one of skill in the art.