Correction of halos in digital images and a device for performing such correction

Abstract

The object of the present invention is a method (400) for correcting halos (H) in digital images (1) captured using photogrammetry in a three-dimensional modeling studio. The halos are generated by the interaction of light emanating from light sources (L3, L4, L5, L6) in the studio with the optical system of the capture device and appear as local brightness in the digital image. The method comprises the steps of generating (410) a light intensity map (M) characterizing the light source in terms of spatial distribution and light intensity, providing (420) a convolution kernel specific to the capture device, calculating (430) the convolution product of the light intensity map and the kernel to obtain a correction value map (CVM), and removing the correction value map pixel by pixel from the digital image to generate a corrected image (Icorr) in which the halo is absent.

Description

Technical Field

[0001] The present invention relates to a method for correcting halos appearing in digital images captured during photographing or video recording, and an image capture device for implementing the halo correction method.

[0002] In particular, the present invention relates to a method for correcting the halo effect generated by a light source from the perspective of a photographing device, and intended to reduce the quality of an image captured by this device due to a non-uniform change in its brightness.

Background Art

[0003] Conventionally, an artificial lighting system has been used to illuminate a scene in order to take a shot of the scene under good light conditions.

[0004] Due to the intensity of the light generated by these lighting systems and the interaction between this light and the photographing device, artifacts can be generated in the images captured by these devices, and in particular, they can reduce their contrast, generate non-uniform changes in brightness such as whitening, flare factors or diffraction patterns, and appear as diffused or localized lighting that adversely affects the quality.

[0005] For example, in the case of conventional photography using a single camera as a photographing device, the light source is generally arranged outside the field of the camera to avoid the generation of such artifacts for artistic purposes.

[0006] Furthermore, visual corrections such as contrast enhancement are also routinely performed for artistic purposes, that is, according to subjective criteria.

[0007] In the case of a three-dimensional scene modeling studio using multi-viewpoint photogrammetry described in Patent Documents FR3016028B1 and WO2019 / 166743A1, light sources and cameras are dispersed around the scene to be modeled. Generally, it is impossible to avoid the presence of light sources within the field of the cameras, leading to the above-mentioned problem of a decrease in the quality of the images captured by these cameras.

[0008] Therefore, three-dimensional scene modeling based on three-dimensional analysis of images is also adversely affected by the above-mentioned artifacts.

Summary of the Invention

[0009] One object of the present invention is to correct artifacts caused by light sources used during shooting. These artifacts are all designated by the term "halo" and to make the digital images captured by the shooting device in a state almost the same as when there is no luminance distortion caused by the halo.

[0010] For this purpose, an object of the present invention is a method for correcting a halo in a digital image to be corrected in a three-dimensional modeling studio captured using photogrammetry by means of a capturing device having a capturing field for the scene, the halo being generated by an interaction between light emitted by a light source and an optical system of the capturing device and appearing as the brightness of a pixel of the digital image to be corrected, the light source forming part of an illumination system of the scene, the method comprising generating a digital light intensity map characterizing the light source with respect to a spatial distribution with respect to the capturing field and with respect to a light intensity perceived from the capturing device during the capture of the image to be corrected, the map forming a first data matrix, providing a convolution kernel specific to the capturing device, forming a second data matrix, calculating a convolution product of the first matrix and the second matrix to obtain a third data matrix corresponding to a correction value map of the halo in the digital image to be corrected, and removing the correction value map pixel by pixel from the digital image to be corrected to obtain a corrected image free of halos.

[0011] A first advantage of the present invention is the correction of a halo that unevenly modifies the brightness of an image captured by a capturing device, the halo appearing as one or more local brightening or whitening in the image if not corrected.

[0012] Thus, the method according to the present invention greatly improves the contrast, color, and reproducibility of light intensity levels not only within the same image but also between images captured by these different devices in a studio including a plurality of capturing devices.

[0013] Furthermore, in the case of three-dimensional modeling of a scene by multi-view photogrammetry, the modeling identifies and tracks the movement of a particular element of a scene, inter alia, based on the color identity of that element from images captured by a plurality of capturing devices.

[0014] Therefore, a better evaluation of the colors in the image processed by the method according to the invention enables a better three-dimensional modeling of the scene.

[0015] The method according to the invention may have the following specific features. - The step of generating a digital light intensity map may include the step of generating a preliminary digital light intensity map characterizing the light source with respect to the spatial distribution over the imaging field and with respect to the light intensity perceived from the imaging device when the light source is fully visible to the imaging device, and the step of generating the digital light intensity map from the preliminary digital map and the digital image to be corrected by determining the pixels belonging to the light source in the preliminary digital map that are hidden from the imaging device in the digital image to be corrected; - The convolution kernel may be a matrix generated from the sum of a one-dimensional function representing the contribution from the kernel scattering phenomenon and a two-dimensional function representing the contribution from the kernel diffraction pattern, having a steadily decreasing envelope; - The convolution kernel may be a matrix generated from an isotropic function that decreases steadily; - The convolution kernel may be generated by the imaging device in the steps of acquiring first and second digital training images, the two images each comprising a training light source that is switched off and the training light source that is switched on, respectively, generating a training light intensity map of the training light source from the first digital training image by assigning light intensity values to the pixels of the second digital image included in the light source, and calculating the kernel from the two digital training images and the light intensity map.

[0016] The invention extends to an image capture device for a three-dimensional modeling studio comprising a plurality of imaging devices functionally connected to a data processing unit, the data processing unit being particularly adapted to implement the method for correcting the halo according to the invention.

Brief Description of the Drawings

[0017] Reading the detailed description of the non-limiting embodiments taken as examples with reference to the accompanying drawings briefly described below will better understand the present invention and clarify other advantages.

[0018]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Mode for Carrying Out the Invention

[0019] This embodiment is described with reference to FIGS. 1 to 10 and relates to the application of the present invention to a three-dimensional scene modeling studio by multi-viewpoint photogrammetry.

[0020] The studio includes cameras C1 to C8 used as photographing devices, and is arranged around an area A where a scene to be modeled from digital images captured by these cameras is arranged.

[0021] The studio further includes light sources L1 to L8 regularly arranged around the same region A.

[0022] The field V of camera C1 surrounds light sources L4 and L5, which, due to their high light intensity and interaction with the optical elements (lenses, possibly diaphragms) of camera C1, are collectively considered as halos that non-uniformly brighten the image captured by camera C1, causing artifacts in the captured image.

[0023] This embodiment aims to correct the digital image of the scene captured by camera C1.

[0024] Image correction The method for correcting an image according to the present invention, as shown by FIG. 400 in FIG. 4, consists of modeling the halo generated by the light source in the image to be corrected captured by camera C1, and then removing it from this image to obtain a corrected image without halos.

[0025] Here, for simplicity, the case of a grayscale image is considered, but this method can be similarly applied to color images, and it is considered that different color channels such as the red channel, green channel, and blue channel only need to be processed in parallel.

[0026] Step 405 consists of capturing the digital image I to be corrected by camera C1 and storing it in the computer memory, and the halo H brightens a specific region of the image, as shown by the hatched area in FIG. (b) of FIG. 2.

[0027] Step 410 consists of generating a digital light intensity map M showing the spatial distribution and intensity of the light sources located in the field of view of camera C1.

[0028] In the first example of this embodiment, only the halos generated by the emission surfaces of light sources L4 and L5 directly visible by camera C1 are corrected.

[0029] Each visible light source of the camera can be regarded as a set of basic light sources visible to the eye in the image corrected in the form of saturated pixels.

[0030] A "saturated pixel" means a pixel that displays a brightness level of 255 in the case of the maximum value and the brightness encoded in 8 bits of a digital image.

[0031] Figure (c) of FIG. 2 shows the light sources L4 and L5 located in the field V of the camera C1 as seen by the camera C1 when capturing the image to be corrected pixel by pixel. The light source L4 is partially hidden from the camera C1 by the character CE of the scene shown in FIG. (b).

[0032] Here, this figure (c) exactly corresponds to the image of the field V of the camera C1 according to the horizontal range E corresponding to the horizontal range E of the field of the camera C1.

[0033] Step 410 can include sub-steps 412, 414, and 416.

[0034] Sub-step 412 is to determine the saturated pixels of the image of the empty studio with the light sources L4 and L5 switched on so as to reach a preliminary digital light intensity map PM, and then generate a digital light intensity map of the empty studio, that is, a state where the light sources are completely visible from the camera C1, by assigning the light intensity of the light sources to these pixels.

[0035] Figure (a) of FIG. 2 shows a preliminary map PM corresponding to the image of the field V of the range E of the camera C1 including only the light sources L4 and L5 as seen from the camera C1. These light sources are rectangular illumination surfaces here.

[0036] This light intensity map is the first matrix of corresponding pixels with respect to the position of the pixels of the digital image I to be corrected, and the luminance value of the light source is assigned to the pixels of the map belonging to the light emitting surface as seen by the camera C1, and the other pixels have a value that is essentially zero.

[0037] The light intensity here is related to the maximum luminance achievable by the saturated pixels in the image to be corrected.

[0038] In any photometric reference consisting of assigning 1 as the light intensity value just sufficient to saturate the pixels of the image to be corrected, the luminance of the light source reaches several thousand, for example, 10,000 and can be measured by conventional means.

[0039] For example, by adjusting the light sensitivity range of the camera C1, the camera can directly measure the light intensity of the light source for each pixel while avoiding pixel saturation, and generate a light intensity map PM that constitutes a reference digital image of the field of the imaging device.

[0040] By proceeding in this way, only the light sources located in the field V of the camera C1 are considered, and these are also generally the light sources that most degrade the image to be corrected.

[0041] To generate the light intensity map of an empty studio, the position and intensity of the light source can also be manually entered into a computer file.

[0042] Sub-step 414 consists of identifying, by means of conventional image processing, the pixels of the map PM belonging to the light sources found as unsaturated pixels in the image I to be corrected, as shown by figure (b) of figure 2.

[0043] This step makes it possible to identify the pixels of the illuminated surface of the preliminary map PM hidden from the camera C1 by elements of the scene such as the character CE in figure (b) of figure 2.

[0044] Sub-step 416 consists of assigning an essentially zero value to the pixels identified in step 414 so as to obtain a light intensity map M, represented by FIG. (c) of FIG. 2, which represents the light sources directly visible by camera C1 during the acquisition of the image I to be corrected, and which shows the respective positions and light intensities in the field of the camera.

[0045] The light sources can have a spatially non-uniform intensity in the light intensity map M in order to reflect, for example, a situation where the light sources consist of panels of very bright light-emitting diodes arranged behind a diffusing panel that only imperfectly diffuses the light from the diodes, and they can appear in the form of small local surfaces having a stronger light intensity than the surroundings.

[0046] Step 420 of FIG. 400 consists of providing a convolution kernel K adapted to camera C1. This kernel can be regarded as a halo generation kernel specific to camera C1, and this kernel forms a second data matrix.

[0047] This kernel is a matrix that has been captured by the digital image sensor of the camera and that converts the influence of the points of the light source on the luminance of each point of the image to be corrected.

[0048] The characteristics of the camera are the optical system (lens, aperture) and the sensor formed by the matrix of photosensitive pixels.

[0049] The kernel can be defined as a matrix that converts the influence of the pixels of the digital light intensity map on the pixels of the digital image to be corrected, as will be explained below.

[0050] It is recalled that the digital image is a data matrix and each data item represents the light intensity of the corresponding pixel of the image.

[0051] The halo appears significantly only when the light source is stronger compared to the light intensity level of the scene in which the image is being captured, mainly in the case of studio light sources.

[0052] Furthermore, each pixel of the light intensity map belonging to the light source can be regarded as a basic light source that itself generates a halo.

[0053] However, the overall halo of the image to be corrected results from an additional phenomenon where the effects of each light source or each region of the light source are added to the effects of other light sources or regions of the light source.

[0054] Therefore, the halo of the digital image to be corrected is the addition of all the halos specific to each pixel of the digital map M, and can be calculated by convolution or convolution product of the light intensity map with the convolution kernel K that models the influence of the basic light source on the capture of the image by the camera.

[0055] For this reason, even if the halo generated by the basic surface of the light source is local and hardly perceptible, the halo generated by the light source considered as a whole has a much greater influence and range and needs to be considered in the kernel.

[0056] Therefore, since the inventors have the experience that a light source placed within the image to be corrected generates a significant halo phenomenon over a radius corresponding to one quarter of the width of the image, in such a situation, the kernel has to be twice the normal size, and thus has a size corresponding to half of the size of the pixels of the image.

[0057] In the case of a square kernel and an image with a width of 2,000 pixels, the kernel corresponds to a two-dimensional matrix of 1,000×1,000 elements, and the effect of the halo extends over a radius of 500 pixels from the considered light source, such as the pixels of the image of the illuminated surface of the artificial lighting system.

[0058] The halo having the effect to be corrected is specific to the optical system of the camera used because it is generated by one or more light sources and the interaction between the optical system of the camera that captures the image to be corrected, in particular, the lens and the aperture.

[0059] One can consider a simple but widely applicable case of an isotropic halo generated by multiple scattering of light received by the lens of a camera optical system, which results in non-uniform brightness of the image captured thereby.

[0060] The effect of the isotropic halo on the pixels of the image decreases rapidly with the apparent distance of this pixel from a particular pixel of the light source, and here, since it is limited to the isotropic effect, it depends only on this distance.

[0061] A kernel suitable for correcting an image affected by this isotropic halo can be determined empirically based on an isotropic parametric function that initially decreases rapidly and then tends to asymptotically approach 0, which transforms the effect of the local, isotropic, and rapid decrease given by the pixels from the light source to the image being corrected.

[0062] FIG. 5 generally shows such a parametric function PF, which is defined using parameters and varies as a function of a radius r representing the distance from the light source considered in the case of the kernel.

[0063] FIG. 6 shows a convolution kernel K obtained from the function of FIG. 5 by spatial digitization for a number of points (1,000 points in the above example) corresponding to the distance from the sensor of camera C1 where the halo has a perceptible effect, i.e., a matrix that models the effect of the light source on the image of a particular camera.

[0064] The elements of this matrix are defined such that their values vary according to the distance from the center of the matrix according to the parametric function PF, and each data of the matrix corresponds to a pixel and is considered to be at a distance equal to the pixel pitch of the camera sensor away from its nearest neighbor.

[0065] The kernel K of FIG. 6 shows this structure, and the data of the values decrease with each distance from the center of the matrix according to the decreasing parametric function PF.

[0066] For the sake of explanation, here the central data of the kernel has a value of 1, and the data farthest from the center has a value of 10 -15 , which of course represents a specific case taken as an example.

[0067] Apply this kernel to the image correction described below, manually adjust the parameters of the parametric function PF considering the correction results to adapt to the imaging system used, and improve the correction. By doing so, the kernel is changed by successive iterations until a satisfactory level of correction is obtained.

[0068] Due to the specific situations represented by each imaging system, it is impossible to define one or more exact parametric functions applicable to all systems. Therefore, the above adjustment phase is essential unless the kernel training method described below is applied.

[0069] Each expert can select one or more types of parametric functions according to the type of camera used, personal experience, and preferences regarding calculations.

[0070] Step 430 of the image correction method consists of calculating a correction value map CVM shown by the image in Figure (d) of Figure 2. Each pixel of it corresponds to a pixel of the image to be corrected by obtaining the convolution product of the light intensity map M and the convolution kernel K. This map forms a third data matrix.

[0071] The map CVM includes a correction region Corr having non-zero values corresponding to the halos generated by the light sources of the light power map M, and the pixels in other regions of the map CVM have non-corrected values of zero.

[0072] Step 440 of this method consists of calculating a corrected image Icorr shown by FIG. (e) of FIG. 2 by removing the correction value map CVM pixel by pixel from the image I to be corrected, so as to obtain a corrected image without halos.

[0073] The expression "obtaining a corrected image without halos" actually represents an ideal goal that cannot be fully achieved, that is, the goal is to completely eliminate halos. However, it is understood that the implementation and advantages of this method do not require completely removing halos from the image to be corrected.

[0074] In this document, removing the first image from the second image is considered equivalent to subtracting the luminance value of the first image from the luminance value of the second image for each pixel.

[0075] Finally, step 450 consists of recording the corrected image calculated in step 440 in the computer memory.

[0076] FIG. 3 shows a modification in the establishment of the light intensity map M'. Different from the method of FIG. 2, the light intensity map is preferably an image of any size in the range E' that covers a field wider than the field of the camera C1 and surrounds the latter.

[0077] For example, it is possible to manually input the positions of the light sources included in this range E with respect to the field of view of the camera C1, insert them into an image having the surface of the pixels surrounding the image to be corrected, and assign the light intensity measured by the conventional method to these positions to obtain a preliminary light intensity map PM' shown by FIG. (a) of FIG. 3.

[0078] For the convenience of calculation, each pixel of the image I to be corrected preferably corresponds exactly to one pixel of the preliminary map PM'.

[0079] From a practical perspective, the map PM’ can also be obtained by changing the orientation of the camera C1, capturing views of the entire range E’ with an appropriate range of light sensitivities, and then combining these views in a conventional manner to form a preliminary map PM’.

[0080] Subsequently, by performing sub-steps 414 and 416, it is possible to establish a light intensity map M’ shown in Figure 3(b) that is adapted to the correction of the corrected image I in Figure 2(b).

[0081] This map takes into account not only the light sources L4 and L5 located within the field of the camera C1 in a range E’ larger than the range E, but also the light sources L3 and L6 located outside the field V of the camera C1, which therefore do not appear in the corrected image I but can significantly contribute to the halos that degrade the quality of the latter.

[0082] By using this light intensity map in the correction method 400, it is possible to correct not only the halos generated by the light sources included in the image to be corrected but also those generated by the light sources located outside the image to be corrected, thus improving the quality of the correction.

[0083] Kernel - Training Method The above image correction procedure can use a convolutional kernel obtained empirically as previously described or a kernel obtained by calculation through a training procedure.

[0084] An overview of two possible training procedures will be described. Each makes it possible to calculate a convolutional kernel for generating the halos of a specific imaging device using two training images, each representing a light source that is captured by this imaging device with the switch on and the same light source with the switch off.

[0085] These procedures enable the determination of the light intensity map defined above by either of these two images, and are based on the principle that the difference between these images enables the determination of the halo created by the light source interacting with the imaging device under consideration.

[0086] If it is known how the halo is generated from these data and the target convolution kernel of the optical power map, the conventional calculation method can be used to find this kernel.

[0087] The first training procedure is shown by FIGS. 700 of FIG. 7 and FIGS. (a)-(f) of FIG. 8, and performs deconvolution.

[0088] Step 710 consists of capturing a first training digital image LI1 that includes the switched-off light source Soff identified by Soff in FIG. 8(a) of FIG. 8 using the imaging system for which the halo generation kernel is to be calculated.

[0089] This can be the imaging device used to capture the image to be corrected, or another device of the same model equipped with the same optical system.

[0090] Step 720 consists of capturing a second training digital image LI2 that includes the same light source but is now switched on, under the same conditions as the image captured when the switch was off and identified by Son in FIG. 8(b) of FIG. 8.

[0091] Note that due to the halo caused by the light from the light source, the area occupied by the image captured when the switch is on is larger than when the switch is off.

[0092] Step 730 consists of removing the first image LI1 from the second image LI2, thereby obtaining a third digital image LI3 that includes a portion representing the halo generated by the interaction between the light source and the imaging device.

[0093] In FIG. 8(c), the regions Hint and Hext represent the halo regions located inside and outside the switched-off light source in FIG. 8(a), respectively.

[0094] Step 740 consists of extracting a curve LC from the third image LI3, and this curve LC represents the luminance difference ΔL between the images LI1 and LI2 along the segment Seg crossing the light source, as shown by FIG. 8(d).

[0095] Next, the portion P of the curve LC that corresponds to the outer periphery of the light source, crosses the region Hext, and includes the maximum value of the curve LC is considered to represent the halo and enables finding the kernel.

[0096] Conversely, the region Hint is not considered to provide information that can actually be used to determine the kernel.

[0097] Step 750 consists of generating, for example, from the first image LI1, a light intensity training map LM as shown by FIG. 8(e) by assigning the light intensity values of the light source measured by a conventional method to the pixels considered to be part of the light source.

[0098] For example, if a pixel exceeds a given light intensity level, the pixel can be considered part of the light source, and this given light intensity level is selected by the operator to distinguish the switched-off light source from the background of the image LI1.

[0099] The training light intensity map LM can also be obtained manually as described above to obtain a preliminary light intensity map PM.

[0100] From the portion P of the curve LC and the training light intensity map LM, it is possible to determine the kernel generation function KF, and then to determine the desired convolutional kernel KM itself in the form of a matrix that will be used later in the image correction method outlined above.

[0101] More specifically, considering that the function represented by the curve LC in the region P is the result of the convolution of the training light intensity map LM and the kernel to be obtained, the inverse convolution operation 760 of the curve LC by the map LM makes it possible to determine the kernel generation function KF, and the inverse convolution is performed by a digital data processing unit such as a computer processor.

[0102] Finally, step 770 for the spatial digitization of the function KF at several points corresponding to the pixel-by-pixel distance from the sensor of the imaging device where the halo has a perceptible effect makes it possible to determine the elements of the kernel, i.e., the matrix KM.

[0103] The elements of the matrix KM are defined from the function KF as the kernel K in FIG. 6 derived from the function PF in FIG. 5, and the data having the values b to f decreases in this order at respective distances from the center of the matrix occupied by the element a so that the matrix KM in FIG. 8(g) is obtained, according to the decreasing function KF.

[0104] The second training procedure is shown by FIG. 900 in FIG. 9 and makes it possible to approximate the kernel generation function by a conventional mathematical method of approximation by regression such as approximation by polynomial regression.

[0105] For the steps in FIG. 900 having the same identifiers as the steps in FIG. 700, reference can be made to the foregoing description.

[0106] The image LI3 obtained from step 730 is considered to correspond to the halo generated by the switched-on light source Son.

[0107] However, this halo is modeled by the convolution product of the required convolution kernel and the light intensity map LM obtained at the end of step 750.

[0108] Therefore, it is understood that the convolution kernel can be obtained by successive approximation of the kernel aimed at converging the result of the convolution product and the image LI3.

[0109] Specifically, an initial kernel is provided in step 910, its convolution product with the light intensity map LM is obtained in step 920, and then the convolution product obtained in the test step 930 can be compared with the image LI3.

[0110] If it is determined during the test step that the image LI3 obtained from step 730 and the result of the convolution product from step 920 are too far apart according to the criteria selected by the implementer, a new kernel is calculated by the data processing unit according to the conventional regression method in step 940, and it is re-input into the convolution product instead of the kernel in step 910, and steps 920 to 940 are looped and repeated until the test step 930 indicates sufficient convergence to the image LI3 of the convolution product.

[0111] When sufficient convergence is obtained, in step 950, the last calculated kernel is recorded in the computer memory as the desired convolution kernel, similar to the kernel KM obtained by the method shown in FIGS. 700, 7, and 8.

[0112] The above example is limited to the case of an isotropic halo generated by a kernel that is isotropic in turn. The kernel is calculated by a one-dimensional function representing the light intensity change in only one direction, the radial direction, and is sufficient to characterize the isotropic contribution to the halo.

[0113] Such an isotropic kernel is generally sufficient. This is because, for example, in the case of an aperture using a large opening, the diffraction effect of the aperture, which is practically negligible, is often ignored.

[0114] In the case of non-negligible diffraction from the aperture, a two-dimensional function having the same degree of symmetry as the aperture is added to a one-dimensional function representing the isotropic contribution to the kernel to represent the diffraction contribution of the aperture to the kernel, and for empirical determination or determination by training of the kernel, it can be performed in the same manner as before.

[0115] This principle applies to any type of optical effect involved in the halo generation of the image to be corrected.

[0116] Generally, the convolution kernel can be generated from the sum of a one-dimensional function having a steadily decreasing envelope representing the isotropic contribution to the kernel and a two-dimensional function representing the anisotropic contribution to the kernel.

[0117] The isotropic contribution to the kernel can arise, in particular, from the scattering phenomenon by the lens of the imaging device being considered.

[0118] The anisotropic contribution to the kernel can originate, in particular, from the diffraction phenomenon and is visualized, for example, in the form of a diffraction pattern of the image to be corrected.

[0119] In such a case, the two-dimensional function can represent the diffraction pattern and can correspond to a function having the same degree of symmetry as this pattern or to the sum of such functions.

[0120] FIG. 10 shows an image capture device 100 for a three-dimensional modeling studio adapted to implement the halo correction method according to the present invention.

[0121] This device includes a plurality of cameras C used as imaging devices, each of which is connected to a digital data processing unit DTU that includes a data concentration and calculation unit CU and a distributed unit DU that forms an interface between the data concentration unit CU and each camera C.

[0122] The control monitor MON connected to the concentration unit enables viewing of the images captured by the system, and the digital data input unit KB such as a numeric keyboard enables input of commands into the data processing unit.

[0123] The monitor MON and the unit KB can be used, for example, to improve the kernel adopted by the halo correction method when an operator visually estimates the quality of halo correction in an image and empirically determines the kernel by manually modifying the parameters of the kernel generation function.

[0124] In this example, each of the distributed units DU includes a memory in which a convolution kernel adapted to the model of the camera of the device is stored.

[0125] The light intensity maps are also stored in these distributed units, each being specific to a camera, associated with the configuration of the modeling studio in which the device is integrated, and depending on the arrangement of the associated camera and lighting system.

[0126] These distributed units are configured to process the images captured by the halo correction method according to the present invention using the light intensity maps and kernels stored in the memory.

[0127] Of course, this system can also be adapted to implement the training procedure of the convolution kernel.

[0128] It should be noted that the images referred to in this description may or may not have been subjected to digital processing for the purpose of improving, for example, contrast and sharpness, and can be understood as a single digital image captured at a specific moment, or as the average of a plurality of digital images captured at different moments, respectively.

[0129] Needless to say, the present invention is not limited to the above-described embodiments, and can be modified without departing from the scope of the present invention.

Claims

1. A method for correcting halos in a digital image to be corrected of a scene captured in a three-dimensional modeling studio using photogrammetry by a photographing device having a photographing field, wherein the halos are generated by an interaction between light emitted by a light source and an optical system of the photographing device, appear as the brightness of pixels of the digital image to be corrected, and the light source forms part of an illumination system of the scene, - generating a digital light intensity map characterizing the light source with respect to the spatial distribution of the light source with respect to the photographing field and with respect to the light intensity perceived from the photographing device during capture of the digital image to be corrected, the map forming a first data matrix, - providing a convolution kernel specific to the photographing device that forms a second data matrix, - calculating the convolution product of the first matrix and the second matrix to obtain a third data matrix corresponding to a correction value map of the halos in the digital image to be corrected, - removing the correction value map from the digital image to be corrected pixel by pixel to obtain a corrected image free of halos, characterized in that it comprises the steps of A method for correcting halos.

2. The step of generating the digital light intensity map comprises - generating a preliminary digital light intensity map characterizing the light source with respect to the spatial distribution of the light source with respect to the photographing field and with respect to the light intensity perceived from the photographing device when the light source is fully visible to the photographing device, - generating the digital light intensity map from the preliminary digital light intensity map and the digital image to be corrected by determining the pixels belonging to the light source in the preliminary digital light intensity map hidden by the elements of the scene from the photographing device in the digital image to be corrected, characterized in that it comprises the steps of The method for correcting halos according to Claim 1.

3. The convolution kernel is a matrix generated from the sum of a one-dimensional function representing the contribution from the kernel scattering phenomenon and a two-dimensional function representing the contribution from the kernel diffraction pattern, having an envelope that decreases steadily, The method for correcting halos according to Claim 1 or 2.

4. The convolutional kernel is a matrix generated from an isotropic function that decreases steadily, characterized in that A method for correcting a halo according to claim 1 or 2.

5. The convolutional kernel is - A step of acquiring first and second digital training images by the imaging device, wherein the first digital training image includes an image of a training light source that is switched off, and the second digital training image includes an image of the training light source that is switched on; - A step of generating a training light intensity map of the training light source from the first digital training image by assigning the light intensity value of the training light source to the pixels of the region occupied by the image of the training light source in the first digital training image; - A step of calculating a kernel from two digital training images and the training light intensity map; characterized by being generated by the steps of A method for correcting a halo according to any one of claims 1 to 4.

6. An image capture device for a three-dimensional modeling studio, comprising a plurality of imaging devices functionally connected to a data processing unit, wherein the data processing unit is specifically adapted to implement the method for correcting a halo according to any one of claims 1 to 5, characterized in that it is an image capture device for a three-dimensional modeling studio.

Images