(19)
(11)EP 2 922 294 B1

(12)EUROPEAN PATENT SPECIFICATION

(45)Mention of the grant of the patent:
10.05.2017 Bulletin 2017/19

(21)Application number: 15167198.9

(22)Date of filing:  21.11.2011
(51)International Patent Classification (IPC): 
H04N 13/00(2006.01)
H04N 17/00(2006.01)
G06T 7/00(2017.01)
G06T 5/00(2006.01)

(54)

SYSTEM AND METHODS TO EFFICIENTLY MEASURE STEREOSCOPIC DISPARITY AND TO GENERATE RELATED MISMATCH IMAGES

SYSTEM UND VERFAHREN ZUR EFFIZIENTEN MESSUNG DER STEREOSKOPISCHEN DISPARITÄT UND ZUR ERZEUGUNG DER ZUGEHÖRIGEN NICHT ÜBEREINSTIMMENDEN BILDER

SYSTÈMES ET PROCÉDÉS POUR LA MESURE EFFICACE DE LA DISPARITÉ STÉRÉOSCOPIQUE ET POUR GÉNÉRER LES IMAGES DÉSADAPTÉES CORRESPONDANTES


(84)Designated Contracting States:
AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR

(30)Priority: 28.12.2010 US 201061427743 P
11.01.2011 US 201113004124

(43)Date of publication of application:
23.09.2015 Bulletin 2015/39

(62)Application number of the earlier application in accordance with Art. 76 EPC:
11189904.3 / 2472882

(73)Proprietor: Tektronix, Inc.
Beaverton, Oregon 97077-0001 (US)

(72)Inventor:
  • Ferguson, Kevin
    Beaverton, Oregon 97008 (US)

(74)Representative: HGF Limited 
Fountain Precinct Balm Green
Sheffield S1 2JA
Sheffield S1 2JA (GB)


(56)References cited: : 
WO-A1-03/036992
US-A1- 2009 041 336
WO-A1-2010/015957
  
  • BOEV A ET AL: "Towards compound stereo-video quality metric: a specific encoder-based framework", IMAGE ANALYSIS AND INTERPRETATION, 2006 IEEE SOUTHWEST SYMPOSIUM ON DENVER, CO MARCH 26-28 2006, PISCATAWAY, NJ, USA,IEEE, 26 March 2006 (2006-03-26), pages 218-222, XP010917472, DOI: 10.1109/SSIAI.2006.1633754 ISBN: 978-1-4244-0069-0
  • SCHARSTEIN D ET AL: "A taxonomy and evaluation of dense two-frame stereo correspondence algorithms", STEREO AND MULTI-BASELINE VISION, 2001. (SMBV 2001). PROCEEDINGS. IEEE WORKSHOP ON KAUAI, HI, USA 9-10 DEC. 2001, LOS ALAMITOS, CA, USA,IEEE COMPUT. SOC, US, 9 December 2001 (2001-12-09), pages 131-140, XP010582733, DOI: 10.1109/SMBV.2001.988771 ISBN: 978-0-7695-1327-0
  
Note: Within nine months from the publication of the mention of the grant of the European patent, any person may give notice to the European Patent Office of opposition to the European patent granted. Notice of opposition shall be filed in a written reasoned statement. It shall not be deemed to have been filed until the opposition fee has been paid. (Art. 99(1) European Patent Convention).


Description

BACKGROUND



[0001] Three-dimensional or stereoscopic video is a proven technology that is in the midst of a renaissance due to improved techniques for simultaneously delivering slightly different images, known as stereoscopic images, to a pair of offset eyes. When each of the left and right eyes view their respective portions of a stereoscopic image having a horizontal disparity, the human brain generates an illusion of depth perception from the disparity. Early stereoscopic cinema relied on different-colored filters worn over each eye to create the depth perception effect, but the overall color of the screen images appeared muted. Newer technology uses polarization filters to separate the images each eye sees. Such techniques are reaching larger audiences because of the widespread adoption of digital projectors in movie theaters, which make it easier to project stereoscopic images than did legacy equipment. With reference to Figure 1, the mechanics of viewing a stereoscopic image are illustrated in a scene 10. Although the left and right eyes 2, 4, are focused on the same screen 6, filters 12 and 14 use filtering techniques such as those described above, or others, causes each eye to see a slightly different image. The brain then translates spatial, especially horizontal disparity of the same or similar objects in the image as the object having depth, when, in fact, both objects are projected on the same screen 6. There is accordingly a need to accurately measure disparity in stereoscopic images and to convey the measurement to a user in an informative way.

[0002] Although there are methods to measure spatial disparity of three-dimensional images, such as those described in a paper "A Taxonomy and Evaluation of Dense Two-Frame Stereo Correspondence Algorithms," by Daniel Scharstein and Richard Szeliski, Microsoft Technical Report MSR-TR-2001-81, available at www.research.microsoft.com, which is referred to as the "Technical Report", they are typically complex and require a large amount of computing power, making them unsatisfactory to use for real-time analysis due to the large expense of processing required.

[0003] Embodiments of the invention address these and other limitations in the prior art.

[0004] Document WO 03/036992 discloses a method and apparatus for performing correspondence estimation between pixels of a stereo image pair and to obtain matching information for corresponding pixels in each image. Document US 2009/041336 discloses a system and method for stereo image matching, using a disparity estimation method. Document BOEV A ET AL, "Towards compound stereo-video quality metric: a specific encoder-based framework", IMAGE ANALYSIS AND INTERPRETATION, 2006 IEEE SOUTHWEST SYMPOSIUM ON DENVER, CO MARCH 26-28, PAGE 218 - 222, discloses a compound full-reference stereo-video quality metric. Document WO2010015957 discloses a similarity measure in image registration that is computed of first and second images offset at a plurality of relative axial offsets, and where a starting relative axial offset between the first and second images is identified based on the computed similarity measures.

SUMMARY



[0005] There is provided a method of generating difference data from a stereoscopic image having left and right image components as set out in claim 1.

BRIEF DESCRIPTION OF THE DRAWINGS



[0006] 

Figure 1 is a block diagram of a 3D delivery system according to the prior art.

Figure 2A is a test image as seen by the left eye and Figure 2B is the test image as seen by the right eye used in explaining embodiments of the invention. Together they are referred to as a stereoscopic image.

Figure 3 is a flow diagram of an example method of efficiently measuring stereoscopic disparity according to embodiments of the invention.

Figures 4 and 5 are each graphs of average luminance values of a respective line of a vertically downsampled stereoscopic image from video frames according to embodiments of the invention.

Figure 6 is a graph of a continuously local mean of the luminance values of the left image illustrated in Figure 4 according to embodiments of the invention.

Figure 7 is a graph of a continuously local mean of the luminance values of the right image illustrated in Figure 4 according to embodiments of the invention.

Figure 8 is a graph of difference values created from comparing the downsampled image lines to their own local mean, for both left and right images, according to embodiments of the invention.

Figure 9 is a two-dimensional data plot image illustrating an example output generated by calculating difference magnitudes of local mean unsharp masked values.

Figure 10A is a graph illustrating measuring of matching of the data plot for the vertical sub-image illustrated in Figure 9.

Figure 10B is a graph illustrates the measure of matching for a particular pixel of the stereoscopic image illustrated in Figure 9.

Figure 11 is a mismatch image created from the generated difference values according to embodiments of the invention.

Figure 12 is a disparity image created from the generated difference values according to embodiments of the invention.

Figure 13 is a functional block diagram illustrating an example system for performing embodiments of the invention.


DETAILED DESCRIPTION



[0007] Embodiments of the invention are directed to a system and methods for determining a magnitude of left to right disparity of an image, as well as an indication of potential discomfort of a viewer simultaneously looking at the left and right images. For images having little or no horizontal disparity, the potential discomfort may be determined along the vertical axis.

[0008] As enumerated in the above-referenced Technical Report on correspondence algorithms of left/right images, the general processing steps when generating a disparity measurement are: a) quantifying a measurement indicative of spatial matching of portions of the respective images on a local basis; b) aggregating the local results from such quantification; c) disparity computation / optimization; and d) disparity refinement. Embodiments of the invention include each of the major steps to create the correspondence, as well as variations that allow a much better result while using a fraction of the resources of the prior art methods.

[0009] Embodiments of the invention may begin with one or more pre-processing steps that modify the images to be compared. With reference to Figures 2A, 2B, and Figure 3, in a preferred embodiment both the left and right images (Figures 2A and 2B) of a stereoscopic image are first down-sampled in the vertical direction in a process 102. For example, if the original stereoscopic image is 1080 pixels in the vertical direction and 1920 pixels in the horizontal direction, as a standard HD video frame, each image may be downsampled into a 10 x 1920 pixel image. This creates a pair of images (left and right) where each image has exactly 10 horizontal lines, each 1920 pixels long. In this example each line represents a combination of the 108 lines of the original stereoscopic image used to make the line, and is referred to as a vertically downsampled image line. Although grayscale images are illustrated, embodiments of the invention work in the color domain as well.

[0010] Various techniques can be used to create the vertically downsampled image line from the 108 lines. In one embodiment the first pixel from each of the 108 lines in a section is averaged to create a single mean pixel value. Then the second pixel of all 108 lines is averaged to create the second pixel in the line, and so on, until the vertically downsampled image line of 1920 pixels is created. Figure 4 is a graph plotting average luminance values for the first vertically downsampled image line in each of Figures 2A and 2B, and Figure 5 is a graph plotting average luminance values for the second vertically downsampled image line of Figures 2A and 2B. In Figure 4 the reference 20 charts the average luminance value for the first vertically downsampled image line from vertically downsampling Figure 2A from 1080 lines to 10. The reference 22 charts the same value for the downsampled Figure 2B. Likewise, in Figure 5 the reference 30 refers to the average luminance value for the second vertically downsampled image line resulting from vertically downsampling Figure 2A, and reference 32 refers to the same for Figure 2B.

[0011] In another pre-processing procedure, the vertically downsampled image lines, which in this example are 1920 pixels long, may likewise be horizontally downsampled into a shorter line in a process 104 (Figure 3). This may be appropriate if the display for a monitor for evaluating the stereoscopic images is less than 1920 pixels wide, for instance. Other embodiments may downsample in the horizontal direction so that later processing is performed on fewer pixels, which may be important if computing power is limited. Rather than downsampling a vertically downsampled image line into a specific number of pixels, it may be downsampled by a given percentage, for instance to one having a length of 70% of the original length. Such downsampling is a tradeoff between computing speed of the system vs. accuracy of the overall measurements, however. In some embodiments the vertical and horizontal downsampling factors may be inputs controllable by the user, who may wish to vary such factors until a balance between compute speed and accuracy is achieved.

[0012] In some embodiments an anti-alias filter may be used prior to sub-sampling. The selection of which anti-alias filter to use may be determined by available computing resources. For example, if hardware and software resources are readily available, a Seagull filter, such as that described in US Patent 7,406,493 may be used. If instead computational speed is available at a premium, the system may use bi-directional IIR low pass filters described in US Published Application 2008/0152257, entitled "Measurement Apparatus and Method of Measurement of Video Spatial Scale, Offset and Cropping," and US Published application 2010/0226592 entitled "Methods and Systems for Image Registration,". These bi-directional IIR filters would be used prior to sub-sampling. Another method of computationally efficient downsampling that may be used includes averaging, using similar methods to those described above. Further, if no computing resources are available for anti-alias filtering, aliasing may be forced and the mean representative line from the vertical downsampling described above is simply decimated, and other values are not used. All of the data need not be used in all cases.

[0013] Next, in some embodiments the resulting vertically downsampled image lines for each of the left and right images may be unsharp masked by subtracting the respective bidirectional IIR low pass filtered version from each, using bi-directional IIR filters as described in the above-referenced patent publications but using different techniques. These are illustrated in the flow of Figure 3, as 106. The filter coefficients of the bi-directional IIR filter used may be a1 = 1 - b0. This results in a generated DC (low frequency) line for each of the vertically downsampled image lines. Figures 6 and 7 each include two lines superimposed on a single graph. The reference line 20 of Figure 6 is the same as reference line 20 of Figure 4, except that Figure 6 has a longer x-axis. The line referenced as 21 in Figure 6 is the output of reference line 20 after being passed through the bidirectional IIR low pass filter, as described above. Similarly, the line referenced as 23 in Figure 7 is the output of reference line 22 of Figure 4 after being passed through the bidirectional IIR low pass filter to create a continuously local mean. Passing the data through the bidirectional IIR low pass filter effectively generates a local mean, which is noticeable in Figures 6 and 7 by comparing the lines 21, 23 to the respective original data lines of 20 and 22.

[0014] Next, differencing values made from modifying the left and right images are then put through the standard four correspondence measurement steps described in the Technical Report above, with some differences that further reduce computing resources while still producing effective results. The first step of the disparity measurement process in the Technical Report is referred to as "matching cost computation," in which a measurement of the spatial matching of portions of the stereoscopic images is produced. The Technical Report states that differences may be computed using an absolute value of difference, squared difference, cross-correlation, phase or wavelet phase of the left and right images. In effect, the left and right images are compared at different relative shifts, with local matching quantification results of each shift.

[0015] In embodiments of this invention, with reference to Figures 3 and 6, a first portion of this step is accomplished by generating a difference value between the vertically downsampled image lines and their respective local mean as shown in process 110. In other words, the original luminance values calculated for the original vertically downsampled image line (reference 20 of Figures 4 and 6) is compared to an IIR filtered version of the same reference (reference 21 of Figure 6) to yield a difference value, one difference value per pixel in the line. This difference value is illustrated in Figure 8, in which the plot 40 is the difference value for the reference lines 20 and 21, originating from the left side of the stereoscopic image of Figure 2A, and in which the plot 42 is the difference value for the reference lines 22 and 23 of Figure 7, originating from the right side of the stereoscopic image of Figure 2B. These difference values that are plotted as reference lines 40 and 42 of Figure are referred to as AC (high frequency) portions of the respective signal, and are also referred to as the AC lines.

[0016] The second portion of the disparity measurement is processed in embodiments of the invention by comparing the difference values calculated in the first portion to one another. Some embodiments include another computing savings efficiency by not comparing every pixel to every other pixel, but instead skips a certain number of pixels in the lines being compared. For example, each pixel in line 40 of Figure 8 need not be compared to all of the pixels in line 42. This is equivalent to decimation for the above-mentioned downsampling process, without the need for making new copies of the data.

[0017] The second step of calculating a correspondence in the Technical Report is aggregation of the local results from the first step. These are computed with a support region defined by a window such as square, Gaussian, and others having varying size, phrase, etc., typically using summation or FIR (Finite Impulse Response) filters.

[0018] Embodiments of the invention, exhibiting improved computational efficiency and accuracy, address the aggregation step by passing the difference values generated above through the bidirectional IIR low pass filter. Thus, aggregation is accomplished by filtering in this example embodiment. The same bi-directional IIR filter coefficients for performing this filtering action may be identical to the IIR filter coefficients used above, a1 = 1 - b0. This is illustrated as a process 112 of Figure 3. The output of the low pass filter is a two-dimensional output where each column is formed of data representative of a shift value. An example output is illustrated as the data plot 200 in Figure 9. For a 1080 x 1920 stereoscopic image that is downsampled into 10 horizontal bands in the process 102 of Figure 3, ten data plots 200 are produced in the process 112, one for each vertically downsampled image line.

[0019] More specifically, with reference to Figure 9, the data plot 200 is an error image that shows pixel localized error against shift values. For instance, in some embodiments the top line of the data plot 200 may be formed by taking a vertical downsampled image line for the left image and comparing it to a vertical downsampled image line for the right image. Such a comparison may be the absolute value difference of each pixel. In Figure 9 example left and right image portions of the original stereoscopic image illustrated in Figures 2A and 2B are shown. These left and right image portions 202, 204 are compared to each other at a first shift value, and each pixelwise comparison is recorded as a data element of the data plot 200. As typical in comparing lines having 1920 pixels, there would be 1920 - |shift| (the absolute value of the number of pixels of relative shift) comparisons for each line in the data plot 200.

[0020] The second line in the data plot is created by shifting the left and right image portions 202, 204 relative to each other by a given value. In systems where computing resources are readily available, the left and right image portions 202, 204 may only be shifted one pixel for each set of comparisons. More typical, though, would be for the pixels to shift at a larger incremental shift. After all of the desired comparisons are made, and their comparison values recorded, the data plot 200 is complete.

[0021] The third step of calculating a correspondence in the Technical Report is disparity computation/optimization. Typically this can be quickly performed by taking the best match per image portion. For example, the disparity computation may be performed per pixel if the stereo correspondence is being measured at that density. However, after all pixel shifts are calculated, the group of shifts may be evaluated for "smoothness," or discontinuities deemed unlikely in some applications. Filtering by noise, median or other filters may be used. Embodiments of the invention assume smoothness is valid for all video, and detecting non-matched regions in the stereoscopic image are more highly valued than generating the disparity value for image portions with no good match.

[0022] At this stage embodiments of the invention generate two maps or images from the data plots 200. First, for each column of the data plot 200, the minimum error value is selected and its value is coded as a pixel of a mismatch image 210, as illustrated in Figure 11. This process appears in Figure 3 as process 114. The top line of the mismatch image 210 is made from selecting the minimum values in each column of the first data plot 200, and is replicated an additional 107 times to compensate for the vertical downsampling process 102 (Figure 3). That is, the first 108 lines of the mismatch image 210 are identical. This means that, in this embodiment, the mismatch image 210 and the original stereoscopic images illustrated in Figures 2A and 2B have the same dimensions. Similarly, the second 108 lines in the mismatch image 210 are made from a second data plot 200, which was made by comparing the left and right images in the second vertically downsampled image line to one another, as described above. Each of the data plots 200 created in the comparison process is used to make a corresponding horizontal line of the mismatch image 210. After all of the minimum error values are selected from all of the data plots 200, the values may be normalized for maximum image contrast for the mismatch image. For instance, the minimum error value of 0 may be coded to appear black on the mismatch image 210, and a maximum error value of 210, which may be the largest of the minimum values found in any column of any data plot 200, may be coded to appear white.

[0023] This mismatch image 210 serves as an indication of whether a viewer will be successful in generating the three-dimensional depth perception of the particular section of the stereoscopic image. In other words, higher levels of mismatch, which may appear as light areas, indicate that the viewer may have problems with successful fusion of the left and right images of the stereoscopic image in the viewers mind.

[0024] Also from the data plot 200 of Figure 9, a disparity image 220 is generated, as illustrated in Figure 12 and shown as process 116 of Figure 3. Each pixel in the disparity image 220 represents the number of pixels that each pixel must be shifted along its horizontal line to make the minimum error determined above. In other words, differently to where the mismatch image of Figure 11 was made by taking the actual minimum comparison value in every column, the disparity image of Figure 12 is made by plotting in the disparity image of Figure 12 the shift corresponding to the line number of each column in which the minimum value appeared. Shifts are related to line number in this embodiment by shift = LineNumber * ShiftIncrement + InitialShift. Like the mismatch image 210, the disparity image 220 is made by replicating the line an additional 107 times. Also like the mismatch image, each horizontal line of the disparity image corresponds to a single one of the data plots 200.

[0025] Figure 10A is a graph 206 that illustrates the measure of matching for the sub-image illustrated in Figure 9. In this graph, the maximum values illustrate the minimum difference magnitude, which is also the compliment to the local absolute mean difference. The top curve family in Figure 10A shows how matching curves of match (vertical) vs. pixel (horizontal) where each new curve represents a new offset (disparity) used for calculation. Figure 10B is a graph 208 that illustrates the measure of mapping for a single pixel, the 550th left-most pixel of the second sub-image illustrated in Figure 9. Again, the value graphed is the compliment to local absolute mean difference, meaning the minimum difference magnitude is shown as a maximum. There are clearly two nearly equal magnitude peaks, which may be further selected, if desired, using known methods.

[0026] The fourth step of calculating a correspondence in the Technical report is refinement of the disparity computation/optimization step (step 3). Refinement is conventionally performed by interpolating between disparity map locations, such as the disparity map 220 of Figure 12. Embodiments of the invention perform different refinement processes.

[0027] A first option for refinement in embodiments of the invention is to perform an extended horizontal search when searching for minimum values. For instance, a threshold value may be established that must be reached in order to be a qualified minimum for the minimum selection process used to create the mismatch and disparity images in processes 114 and 116 of Figure 3. Recall that these searches are performed in windows around the value being searched. In the extended search, a larger window is created and searched, which may include vertical as well as horizontal searching.

[0028] A second option of refinement includes performing a sub-region search. For instance, if multiple identical displacements are found when searching the data plot 200 of Figure 9, or if the mismatch image generation process indicates a poor mismatch, then the particular vertical segment may be divided into two (each made from averaging 54 lines each, for example, instead of 108) and the process repeated until better matches are found. This may occur multiple times, if necessary, to satisfy the threshold minimums. This may occur especially if the original stereoscopic image includes multiple distinct depth planes.

[0029] A third option of refinement includes a temporally adaptive search for stereoscopic images that are a portion of a video. For video, successive frames use previous frame disparity information and differential disparity (disparity motion) to predict disparity in the current frame, which has the effect of creating better matches at the outset with less refining. More specifically, after the first left and right frames of the stereoscopic image are processed to get the shifts for matching (or mismatch) of each pixel, this information is used as a starting point for matching each pixel of the following frame. This has the benefit of being able to use a smaller search range than in systems that do not use the prediction, which, in turn, allows for a better result while simultaneously reducing computing resource needs. Ranges may be limited to those corresponding to human eye tracking rate (measured in visual degrees/second), and vergence adjustment tracking rates. Techniques other than smaller search range may also be used in conjunction with embodiments of the invention, such as object tracking, linear predictive coding, and motion compensation methods in video compression algorithms known in the prior art, for example.

[0030] Figure 13 is a functional block diagram of an example system to efficiently measure stereoscopic disparity and make measurements of 3-D images according to embodiments of the invention. An analysis device 300 may be a stand-alone device or may be part of a larger system. The device 300 may be implemented on a general purpose computer running a commercial operating system and having the various illustrated functions embodied as software processes. These processes may be separate processes or may all be sub-components of a larger program. Alternatively, the device 300 may be formed in a programmable device, such as an FPGA (Field Programmable Gate Array), as is well known in the art. If production costs warrant, the device 300 may be embodied in an ASIC (Application Specific Integrated Circuit). Most typically, though, the analysis device 300 is formed of several of these components amalgamated into a single device. For example, an FPGA may be used for the bulk of processing, while a software routine running on general purpose computer hardware may run a menu system that allows the user to make selections for the analysis device 300 to perform. It is important to note that the details of how the processes are embodied are implementation specific and generally left to the system designer, who is well versed in all of these various implementation methods and the benefits and detriments of each.

[0031] A stereoscopic video input including left and right images is input to the analysis device 300. A downsample processor 310 is structured to perform vertical 312 and/or horizontal 314 downsample processing, depending on operating details. The more the stereoscopic images are downsampled, the fewer computing resources are necessary to perform analysis, but at a cost of decreased precision and accuracy. The vertical and horizontal downsampling may be dissimilar, where images are downsampled vertically much more than horizontally.

[0032] Following the downsampling, a processor 320 generates an unsharp masked image output by comparing the output of the downsample processor 310 to a low-pass filtered version of itself. As described above, the low-pass filter 322 may include an IIR bi-directional filter. The unsharp masked image output is then passed to a difference measurer 330, which performs the difference processing between the unsharp mask-processed image lines for both the left and right images at various relative shifts. The difference measurer 330 may use an absolute value measurement facility 332 as described above to make its initial difference calculation, or any of a number of variations. Then, the difference signal is passed through another process, such as another IIR low pass filter 334 to generate an output of local mean unsharp masked difference magnitudes. This output was described above as the two-dimensional data plot that is generated for each vertically downsampled image line. This data plot serves as the basis for the image creator 340, which makes both a mismatch image and a disparity image from the data plot. More specifically, both the disparity image and the mismatch images are indicative of how matched the components in the original stereoscopic image are to one another. The disparity image provides information to the viewer regarding how far each section of the images are apart relative to one another, while the mismatch image is informational regarding the actual minimum error at this spatial shift, corresponding to the amount of mismatch between left and right images. The amount of mismatch can also be a good indicator of likely visual discomfort experienced from viewing the images. These images are presented to the user as images that can be displayed or stored for analysis.

[0033] Consequently, in view of the wide variety of permutations to the embodiments described herein, this detailed description and accompanying material is intended to be illustrative only, and should not be taken as limiting the scope of the invention. What is claimed as the invention, therefore, is all such modifications as may come within the scope of the following claims.


Claims

1. A method of generating difference data from a stereoscopic image having left and right image components, the method comprising:

vertically downsampling (102) each of the left and right images into a set of downsampled image lines;

creating an unsharp mask (106) for at least one of the downsampled image lines by:

filtering the at least one downsampled image line with an IIR low pass filter to create a filtered downsampled image line; and

removing the filtered downsampled image line components from the at least one downsampled image line;

generating a set of difference data (110) for the at least one of the downsampled image lines in the left image by comparing the downsampled image line to its own unsharp mask and storing the comparison;

generating a set of difference data (110) for the at least one of the downsampled image lines in the right image by comparing the downsampled image line to its own unsharp mask and storing the comparison;

comparing the stored sets of comparison data from the left image and the right image to one another;

quantifying a difference value between the stored sets of comparison data; and

creating a set of local mean unsharp masked difference magnitudes (112) from the quantified difference value.


 
2. The method of claim 1, further comprising:

generating a mismatch image (114) from the set of local mean unsharp masked difference magnitudes.


 
3. The method of claim 2, further comprising:

generating a disparity image (116) from the set of local mean unsharp masked difference magnitudes.


 
4. The method of claim 1, further comprising horizontally downsampling (104) each of the left and right images into a set of downsampled image lines, or horizontally downsampling the set of downsampled image lines created by the vertical downsampling.
 
5. The method of claim 1 in which the IIR low pass filter has coefficients a1 = 1 - b0.
 
6. The method of claim 1 in which generating a set of difference data comprises filtering the downsampled image line with a bi-directional IIR filter having coefficients a1 = 1 - b0.
 
7. An image processing device (300) comprising means for carrying out the method of any of claims 1 to 6.
 
8. A computer readable medium comprising instructions which when executed by one or more processors results in the one or more processors carrying out the method of any of claims 1 to 6.
 


Ansprüche

1. Verfahren zum Erzeugen von Differenzdaten von einem stereoskopischen Bild, das linke und rechte Bildkomponenten aufweist, wobei das Verfahren Folgendes umfasst:

Vertikales Unterabtasten (102) von jedem der linken und rechten Bilder zu einem Satz von unterabgetasteten Bildzeilen;

Erstellen einer Unschärfemaske (106) für zumindest eine der unterabgetasteten Bildzeilen durch:

Filtern der zumindest einen unterabgetasteten Bildzeile mit einem IIR-Tiefpassfilter, um eine gefilterte unterabgetastete Bildzeile zu erstellen; und

Entfernen der gefilterten unterabgetasteten Bildzeilekomponenten von der zumindest einen unterabgetasteten Bildzeile;

Erzeugen eines Differenzdatensatzes (110) für die zumindest eine der unterabgetasteten Bildzeilen in dem linken Bild

durch Vergleichen der unterabgetasteten Bildzeile mit ihrer eigenen Unschärfemaske und Speichern des Vergleichs;

Erzeugen eines Differenzdatensatzes (110) für die zumindest eine der unterabgetasteten Bildzeilen in dem rechten Bild durch Vergleichen der unterabgetasteten Bildzeile mit ihrer eigenen Unschäfemaske und Speichern des Vergleichs;

Vergleichen der gespeicherten Vergleichsdatensätze von dem linken Bild und dem rechten Bild miteinander;

Quantifizieren eines Differenzwertes zwischen den gespeicherten Vergleichsdatensätzen; und Erstellen eines Satzes von lokalen mittleren unscharf maskierter Differenzgrößen (112) von dem quantifizierten Differenzwert.


 
2. Verfahren nach Anspruch 1, ferner umfassend:

Erzeugen eines nicht übereinstimmenden Bildes (114) aus dem Satz von lokalen mittleren unscharf maskierter Differenzgrößen.


 
3. Verfahren nach Anspruch 2, ferner umfassend:

Erzeugen eines Disparitätsbildes (116) aus dem Satz von lokalen mittleren unscharf maskierter Differenzgrößen.


 
4. Verfahren nach Anspruch 1, wobei das Verfahren ferner horizontales Unterabtasten (104) von jedem der linken und rechten Bilder zu einem Satz von unterabgetasteten Bildzeilen, oder horizontales Unterabtasten des durch das vertikale Unterabtasten erstellten Satzes von unterabgetasteten Bildzeilen umfasst.
 
5. Verfahren nach Anspruch 1, wobei der IIR-Tiefpassfilter Koeffizienten a1 = 1 - b0 aufweist.
 
6. Verfahren nach Anspruch 1, wobei das Erzeugen eines Differenzdatensatzes das Filtern der unterabgetasteten Bildzeile mit einem bidirektionalen IIR-Filter, der Koeffizienten a1 = 1 - b0 aufweist, umfasst.
 
7. Bildverarbeitungsvorrichtung (300), umfassend Mittel zur Durchführung des Verfahrens nach einem der Ansprüche 1 bis 6.
 
8. Computerlesbares Medium, umfassend Anweisungen, die, wenn sie durch einen oder mehreren Prozessoren ausgeführt werden, in den einen oder mehreren Prozessoren resultieren, die das Verfahren nach einem der Ansprüche 1 bis 6 durchführen.
 


Revendications

1. Procédé de génération de données de différence à partir d'une image stéréoscopique ayant des composants d'image de gauche et de droite, le procédé comprenant :

la réduction de fréquence d'échantillonnage verticale (102) de chacune des images de gauche et de droite pour obtenir un ensemble de lignes d'image de fréquence d'échantillonnage réduite ;

la création d'un masque flou (106) pour au moins l'une des lignes d'image de fréquence d'échantillonnage réduite par :

le filtrage de ladite au moins une ligne d'image de fréquence d'échantillonnage réduite avec un filtre passe-bas IIR pour créer une ligne d'image de fréquence d'échantillonnage réduite et filtrée ; et

la dépose des composants de ligne d'image de fréquence d'échantillonnage réduite et filtrée à partir de ladite au moins une ligne d'image de fréquence d'échantillonnage réduite ;

la génération d'un ensemble de données de différence (110) pour ladite au moins une des lignes d'image de fréquence d'échantillonnage réduite dans l'image de gauche

en comparant la ligne d'image de fréquence d'échantillonnage réduite avec son propre masque flou et en stockant la comparaison ;

la génération d'un ensemble de données de différence (110) pour ladite au moins une des lignes d'image de fréquence d'échantillonnage réduite dans l'image de droite en comparant la ligne d'image de fréquence d'échantillonnage réduite avec son propre masque flou et en stockant la comparaison ;

la comparaison des ensembles stockés de données de comparaison à partir de l'image de gauche et de l'image de droite l'une avec l'autre ;

la quantification d'une valeur de différence entre les ensembles stockés de données de comparaison ; et

la création d'un ensemble de magnitudes (112) de différence masquées floues moyennes locales à partir de la valeur de différence quantifiée.


 
2. Procédé selon la revendication 1, comprenant en outre :

la génération d'une image d'inadéquation (114) à partir de l'ensemble de magnitudes de différence masquées floues moyennes locales.


 
3. Procédé selon la revendication 2, comprenant en outre :

la génération d'une image de disparité (116) à partir de l'ensemble de magnitudes de différence masquées floues moyennes locales.


 
4. Procédé selon la revendication 1, comprenant en outre la réduction de fréquence d'échantillonnage verticale (104) de chacune des images de gauche et de droite pour obtenir un ensemble de lignes d'image de fréquence d'échantillonnage réduite, ou la réduction de fréquence d'échantillonnage horizontale de l'ensemble de lignes d'image de fréquence d'échantillonnage réduite créé par la réduction de fréquence d'échantillonnage verticale.
 
5. Procédé selon la revendication 1, dans lequel le filtre passe-bas IIR présente des coefficients al = 1 - b0.
 
6. Procédé selon la revendication 1, dans lequel la génération d'un ensemble de données de différence comprenant le filtrage de la ligne d'image de fréquence d'échantillonnage réduite avec un filtre IIR bidirectionnel présentant des coefficients al = 1 - b0.
 
7. Dispositif de traitement d'image (300) comprenant des moyens pour exécuter le procédé selon l'une des revendications 1 à 6.
 
8. Support lisible par ordinateur comprenant des instructions qui, lorsqu'elles sont exécutées par un ou plusieurs processeurs, ont pour résultat qu'un ou plusieurs processeurs vont exécuter le procédé selon l'une des revendications 1 à 6.
 




Drawing





























Cited references

REFERENCES CITED IN THE DESCRIPTION



This list of references cited by the applicant is for the reader's convenience only. It does not form part of the European patent document. Even though great care has been taken in compiling the references, errors or omissions cannot be excluded and the EPO disclaims all liability in this regard.

Patent documents cited in the description




Non-patent literature cited in the description