AU2012202191B2 - Reducing noise in digital images - Google Patents

Abstract

A target digital image is received from an image sensor. The image is contaminated by noise of unknown magnitude that is represented by a reference digital image. A process 5 is applied that uses statistical analysis of the target digital image and of the reference digital image to estimate a magnitude of the noise for at least some pixels of the target digital image. 1292757 1 (GHMItters) P80185.AU.1 OPTICS CMVOS SENSOR ARRAY INPUT DIGITAL IMAGE (T) 10 22 NOISES-I -STO1RAGE, 38)REDUCTION .2 T, 10 SOFTWARE FUGPUT DIGITAL IMAGE, () QUANTITATIVE DETERMiNATIO.N, 30

Description

AUSTRALIA Patents Act 1990 COMPLETE SPECIFICATION Standard Patent Applicant(s): MELA Sciences, Inc. Invention Title: REDUCING NOISE IN DIGITAL IMAGES The following statement is a full description of this invention, including the best method for performing it known to me/us: REDUCING NOISE IN DIGITAL IMAGES This description relates to reducing noise in digital images. When features that appear in digital images, for example in digital images produced by a CMOS image sensor, are to be to quantitatively analyzed (for example, in 5 medical applications), it is useful to reduce or remove the noise from the images before the analysis. It also is useful to reduce or remove the noise from the images in cases for which the signal that produces the digital images is low relative to the noise. One kind of noise in the pixel values that make up the image (called dark current noise, see figure 1) represents random noise levels that are produced by respective 10 pixels of the CMOS sensor array whether or not light is being received by the sensor. (We sometimes use the word pixel in two different senses, one to refer to the photosensitive element that resides at a particular location on the sensor array and the other to refer to the picture element that resides at a particular location on the image and has a value that corresponds to brightness. We sometimes use the term digital 15 image to refer to the array of pixel values that make up the image.) Aside from temperature, the physical characteristics of each sensor pixel that govern its dark current level do not change over time. The pixel's dark current level does, however, depend on the temperature of the pixel. And the pixel's dark current causes charge to build up over time, so that the effect of dark current on a pixel value depends on the 20 duration of exposure of the pixel. Other artifacts in the images include vertical patterns (also called fixed pattern noise), offset, and shot noise. Vertical patterns (see figure 2) are due to unintended differences in the operations of the respective readout circuits of different columns of the array and generally do not change over time. Offset represents differences in 25 overall signal level (brightness) from image to image that result from variations in certain electrical properties of the readout circuitry. Each pixel value generated by the sensor array includes random shot noise with variance proportional to the signal value. Digital images produced by other kinds of image sensors can also be subject to dark 30 current noise. I A SUMMARY In general, a target digital image is received from an image sensor. The image is contaminated by noise of unknown magnitude that is represented by a reference digital image. A process is 5 applied that uses statistical analysis of the target digital 5 image and of the reference digital image to estimate a magnitude of the noise for at least some pixels of the target digital image. Embodiments of the present invention may include one or more of the following features. The 10 sensor is a CMOS sensor or a CCD sensor. The noise comprises dark current noise. The process estimates the dark current magnitude for every pixel of the target digital image. The 10 process comprises program instructions. The statistical analysis includes a decorrelation analysis with respect to the target digital image and the reference digital image. The dark current magnitude estimates are produced without requiring information about a temperature of the sensor or a 15 duration of exposure. The reference digital image is based on a dark current digital image that is substantially free of vertical patterns and has been generated from a grey digital image and a black digital image acquired respectively using different exposure periods. The reference digital image is based on a corrected dark current digital image that has been processed to reduce the effect of low-frequency spatial trends across the pixels of the CMOS sensor. The reference 20 digital image is based on a de-correlation of the black digital image with the de-trended dark current image. The process subtracts vertical patterns, pixel by pixel, from the target digital image to produce a vertical pattern corrected digital image. The process applies a dark current removal function to the vertical pattern corrected digital image to produce a dark current corrected digital image. The noise in every pixel of the target digital image is reduced using the 25 estimated dark current levels. The process applies an offset estimation and subtraction function to the dark current corrected digital image to remove offset. The noise reduced target digital image is provided to a processor for use in analyzing features of an image captured by the CMOS sensor. The target digital image includes possibly malignant lesions. 30 These and other features and aspects may be expressed as apparatus, methods, systems, program products, and in other ways. In a first aspect of the present invention, there is provided a method comprising: receiving, from an image sensor, a target digital image contaminated by noise of 35 unknown magnitude that is represented by a reference digital image; and 2 3289929_1 (GHMalers) P80185.AU applying a process that uses a variance minimization analysis with respect to the unknown magnitude associated with the target digital image and the reference digital image, to determine the magnitude of the noise for at least some pixels of the target digital image, wherein the determination using the variance minimization analysis is statistically 5 equivalent to a determination of the magnitude of the noise made from a decorrelation condition: Correlation (the target digital image - the magnitude * {the reference digital image}, {the reference digital image}) = 0. io In a second aspect of the present invention, there is provided an apparatus comprising: a software processor configured to receive, from an image sensor, a target digital image contaminated by noise of unknown magnitude that is represented by a reference digital image; and apply a process that uses a variance minimization analysis with respect to the i5 unknown magnitude associated with the target digital image and the reference digital image, to determine the magnitude of the noise for at least some pixels of the target digital image, wherein the determination using the variance minimization analysis is statistically equivalent to a determination of the magnitude of the noise made from a decorrelation condition: 20 Correlation (the target digital image - the magnitude * {the reference digital image}, {the reference digital image}) = 0. In a third aspect of the present invention, there is provided An apparatus comprising: means for receiving a target digital image contaminated by noise of unknown 25 magnitude that is represented by a reference digital image; and means for applying a process that uses a variance minimization analysis with respect to the unknown magnitude associated with the target digital image and the reference digital image, to determine the magnitude of the noise for at least some pixels of the target digital image, wherein the determination using the variance minimization analysis is statistically 30 equivalent to a determination of the magnitude of the noise made from a decorrelation condition: Correlation (the target digital image - the magnitude * {the reference digital image}, {the reference digital image}) = 0. 35 Other advantages and features will become apparent from the following description of Embodiments of the invention. 3 3289929_1 (GHMatters) P80185 AU EMBODIMENTS OF THE INVENTION 5 Figure I is an image of dark current. Figure 2 is an image of vertical patterns. Figures 3, 4A, and 4B are schematic flow diagrams. 5 Figures 5, 6, and 7 are images at stages of the calibration process. Figures 8 and 9 are graphs. 10 As shown in figure 3, an input digital image 10 (which we call T, for target) generated by a CMOS sensor array 12 in response to light 14 received through optics 20 from a target scene 16 (for example, skin with a pigmented lesion 18) can be processed (after temporary storage in storage 38) by noise reduction software 22 (run by a processor 24) to produce an output digital image 26 (which we call 0) for use in quantitative analysis 28 15 (for example, to determine 30 whether the lesion is a malignant melanoma, using the MelaFind@ melanoma detection product of Electro-Optical Sciences, Inc., of Irvington, NY). Even though we describe an example in which the noise reduction is performed on a digital image in the context of medical diagnosis, the noise reduction process is applicable 20 broadly to any digital image produced by a any image sensor array for any purpose and in any context, including any in which the noise-reduced digital image may be subjected to later analysis for which noise in the digital image would be problematic. Although we also describe specific ways to reduce vertical patterns and offset, we note that the 25 technique for reducing dark current (and other noise that can be characterized by a reference digital image) in the digital image described here can be used in a wide variety of applications in the absence of vertical patterns and offset or in which other kinds of noise may or may not be reduced and, if reduced, in which the reduction may or may not be done in the way described in the example described here. 30 Other kinds of processing may also be required with respect to digital images produced by sensors in various applications, including processing to correct optical 3a 3289929_ 1 (GHMatters) P80185 AU effects associated with a specific lens and illuminator. The noise reduction techniques described here thus have applications not limited by any optical correction or optical correction of a particular kind. As shown in figures 4A and 4B, prior to noise-reduction processing of the input 5 digital image, reference information is acquired and processed. In one step in developing reference information, multiple independent sets of a grey digital image 34 are acquired in the dark at a known or unknown sensor temperature and stored in storage 38. The exposure time is in an intermediate range to avoid saturation from long exposures and yet have a reliable measurement of dark current. 10 Figure 5 shows an example of a grey image acquired in the dark (intensities multiplied by 15 for display purposes.) The multiple grey digital images are averaged (39) to produce an average grey digital image 40 (which we call G) in which the level of shot noise at the individual pixels is reduced. 15 Also prior to noise-reduction processing of the input digital image, multiple independent sets of a black digital image are acquired in the dark, if fixed pattern noise such as vertical patterns is present. The exposure time for the black sets is shorter than for grey sets (for example, as short an exposure time as the hardware permits) to have a reliable measurement of the fixed pattern noise and to minimize the 20 dark current level in the black sets. The multiple black images are averaged (39) to produce an average black digital image 44 (which we call B) in which the effect of shot noise at the individual pixels is reduced, as it was for the average grey digital image. The average black digital image B is subtracted (46) from the average gray digital 25 image G to produce a dark current digital image 48 (D). The subtraction of the black digital image from the gray digital image produces an image of pure dark current (free of vertical patterns). Figure 6 shows the image of figure 5 after subtraction of the vertical patterns (intensities multiplied by 15 for purposes of display). Next, a de-trending function is applied (50) to the dark current digital image 48 to 30 remove low-frequency spatial trends from the pixels of the data 48, because the trend 4 in the dark current digital image can be correlated with the target digital image. This de-trending is done by subdividing the entire array of the dark current digital image 48 into sub-arrays of N pixels by N pixels. Within each sub-array, the dark current digital image values are fit to a quadratic function of two variables, using a least 5 squares fit. The value of this quadratic function for each pixel is then subtracted from the actual dark current value in that pixel. In practice, good results have been obtained with N = 3. The result is a de-trended dark current digital image 52 (S). Next, a pure vertical pattern digital image V 56 is generated by first applying a dark current removal function (54) to B. The removal function in general returns a digital 10 image that represents the difference, pixel by pixel, between (i) an input digital image B and (ii) a product of the dark current digital image D times a factor Al V = B - A1* D, where the de-correlation function finds a factor Al with respect to two sets of digital image, B and S: 15 Correlation (B - A l * S, S) = 0, where the correlation is computed over all pixels in a specified region of the image. In other words, the de-correlation function determines a factor Al that de-correlates the digital image B from the digital image S over some region of the image. With respect to the particular step 54 in figure 4, the de-correlation determines the 20 magnitude (Al) of dark current D in the black image B. The de-correlation function is an example of a statistical analysis that enables the dark current noise to be determined from target pixels and from reference pixels without the need to know the temperature of the sensor or the period of exposure. Other statistical approaches could also be used, such as a variance minimization analysis. 25 The result of step 54 in the figure is the pure vertical pattern digital image V 56. Figure 7 shows the image of figure 6 after subtraction of dark current noise (intensities multiplied by 15 and 100 levels were added to each pixel for display purposes). 5 Figure 8 provides a graphical illustration of cross sections of the images of figures 5, 6, and 7 without intensity adjustments. In figure 8, VRem is the intensity after subtraction of fixed pattern noise; DCRem is the intensity after subtraction of fixed pattern noise and of dark current; OffRem is the intensity after subtraction of fixed 5 pattern noise, of dark current, and of offset. In the original image of figure 5, shot noise is about 7-8 levels, and it is not removed by the calibration process. Reduction of the shot noise would require either spatial or temporal averaging of images. The steps described above need only be performed once, e.g., during factory calibration, and the resulting calibration digital images can be stored and used for a 10 large number of target images over a long period of time. It is not necessary to develop the calibration image information again each time a target image is captured. As shown in figure 4B, to reduce noise in T, the pure vertical pattern digital image V is subtracted (58), pixel by pixel, from T to yield a vertical pattern corrected digital image Ti 60: 15 TI = T - V. Next, using the de-correlation function (54) over some region of the image, the magnitude A2 of dark current D in that region of the image TI is determined from Correlation (TI - A2 * S, S) = 0, and then the dark current is removed from TI to produce a dark current corrected 20 digital image T2 64: T2= TI - A2 * D. The CMOS sensor can be arranged to have a black region (for example, in a corner or along one of the edges) of the array which is screened from light (including any light from a target). The digital image from the black region can be used to correct for 25 offset in the image. The black region digital image is first processed by the vertical pattern removal and dark current removal steps described earlier and the resulting processed black region data are averaged 66 to produce an average black region value NB 68. The average black region value is subtracted 70 from every pixel of the dark current corrected digital image T2 of an image to eliminate offset from the target 6 image. The resulting offset corrected digital image T3 72 can be subjected to additional processing depending on the circumstances. For example, if it is of interest to remove artifacts from the digital image T3 due to non-uniformities imparted by the imaging system (such as non-uniform target 5 illumination or non-uniform sensor response) or to determine actual reflectances in the digital image T3, a calibration digital image W may be acquired by imaging a uniform white target with a known diffuse reflectance. This white target digital image is then subjected to a series of operations that include vertical pattern removal (subtracting V, if applicable), determination of the magnitude of dark current by 10 applying de-correlation to W - V and S, dark current removal, and offset removal, to produce WI. A reflectance calibration 78 may be applied to the digital image T3 to produce a reflectance digital image T4 80 by the following computation (performed pixel by pixel): T4 = (T3/W) * (E(W)/E(T3)) * p 15 in which E is the exposure time, and p is reflectance of a white calibration target. The reflectance calibrations 78 removes from the digital images T3 non-uniformities imparted by the imaging system. Additional processing as needed can be performed on T4 to yield the output image 0. The process described above assumes that the temperature of the sensor is uniform at 20 all locations across the sensor array. To accommodate the fact that the temperature may vary across the sensor, dark current could be estimated independently at different parts of the sensor and the independent estimates applied separately to the corresponding portions of the input digital image. The process described above can be applied to monochromatic digital images 25 provided by a sensor. In some examples, the process can be applied to multiple digital images in different spectral ranges that are produced simultaneously by the sensor (e.g., red, green, and blue-RGB). In such cases, the digital images in different spectral ranges may be processed independently as described above. 7 The process may take advantage of a statistical analysis to reduce the need, for some sensors, to control the temperature or duration of exposure as a way to reduce the effects of dark current noise. The processes described above could be implemented in hardware, software, or 5 firmware, or any combination of them. Validation of the dark current estimation technique was performed by comparing the dark current level retrieved from dark images specially taken at various exposure times in a climate-controlled environment against the dark current level predicted by the estimation technique described above. It was demonstrated that a de-correlation based 10 estimator is able to predict accurately the actual dark current level in individual images even in the presence of unavoidable shot noise, as illustrated in figure 9. Other implementations are within the scope of the claims. For example, a variance minimization analysis could be substituted for the de-correlation analysis. The techniques described here may be useful not only for sensors operating in the is visible and infrared ranges but also for x-rays and possibly ultrasound, that is, for any sensors for which removal of dark current noise or effects similar to dark current noise would be useful. Although the discussion above is directed to dark current noise correction, similar techniques could be applied in other contexts in which any noise for which a reference 20 image is known or can be obtained and in which the magnitude of the noise in the target digital image is unknown. In the claims which follow and in the preceding description of the invention, except where the context requires otherwise due to express language or necessary implication, the word "comprise" or variations such as "comprises" or "comprising" is used in an 25 inclusive sense, i.e. to specify the presence of the stated features but not to preclude the presence or addition of further features in various embodiments of the invention. 8 31 8882.1 (GHMatters) P80185.AU

Claims (16)

1. A method comprising: receiving, from an image sensor, a target digital image contaminated by noise of 5 unknown magnitude that is represented by a reference digital image; and applying a process that uses a variance minimization analysis with respect to the unknown magnitude associated with the target digital image and the reference digital image, to determine the magnitude of the noise for at least some pixels of the target digital image, wherein the determination using the variance minimization analysis is statistically 10 equivalent to a determination of the magnitude of the noise made from a decorrelation condition: Correlation (the target digital image - the magnitude * {the reference digital image}, {the reference digital image}) = 0. is 2. The method of claim I wherein the determination is made over pixels in some regions of the target digital image and the reference digital image to estimate the magnitude of the noise for each of at least some pixels of the target digital image.

3. The method of claim I or 2 wherein the sensor comprises a CMOS sensor. 20

4. The method of claim I or 2 wherein the sensor comprises a CCD sensor.

5. The method of any one of claims I to 4 wherein the noise comprises dark current noise. 25 6. The method of any one of the preceding claims also comprising subtracting fixed vertical patterns, pixel by pixel, from the target digital image, to produce a vertical-pattern-corrected target digital image.

7. The method of claim 6 also comprising applying a dark current removal function to the 30 vertical-pattern-corrected target digital image to produce a dark-current-corrected target digital image.

8. The method of claim 7 also comprising applying an offset estimation and subtraction function to the dark-current-corrected target digital image to remove offset. 35 9

9. The method of claim 2 also comprising reducing noise in every pixel of the target digital image using the estimated noise magnitudes for all pixels of the target digital image to produce a noise-reduced target digital image. 5 10. An apparatus comprising: a software processor configured to receive, from an image sensor, a target digital image contaminated by noise of unknown magnitude that is represented by a reference digital image; and apply a process that uses a variance minimization analysis with respect to the 10 unknown magnitude associated with the target digital image and the reference digital image, to determine the magnitude of the noise for at least some pixels of the target digital image, wherein the determination using the variance minimization analysis is statistically equivalent to a determination of the magnitude of the noise made from a decorrelation condition: 15 Correlation (the target digital image - the magnitude * {the reference digital image}, {the reference digital image}) = 0. I1. The apparatus of claim 10 further comprising the image sensor. 20 12. The apparatus of claim I I wherein the image sensor comprises a CMOS sensor.

13. The apparatus of claim I 1 wherein the image sensor comprises a CCD sensor.

14. The apparatus of any one of claims 10 to 13 wherein the noise comprises dark current noise. 25

15. The apparatus of any one of claims 10 to 14 wherein the target digital image comprises an image of one or more lesions on a patient.

16. The apparatus of any one of claims 10 to 15 wherein the determination is made over pixels 30 in some regions of the target digital image and the reference digital image, to estimate the magnitude of the noise for each of at least some pixels of target digital image.

17. The apparatus of any one of claims 10 to 16 wherein the software processor is also configured to produce a noise-reduced version of the target digital image using the determined 35 noise magnitude. 10

18. The apparatus of claim 17 wherein the software processor is also configured to provide the noise-reduced version of the target digital image to a processor for use in analyzing features of an image captured by the image sensor. 5 19. The apparatus of claim 18 wherein the image captured by the image sensor includes one or more lesions.

20. The apparatus of claim 19 wherein the software processor is also configured to report a malignancy state of the one or more lesions to a user of the apparatus. 10

21. An apparatus comprising: means for receiving a target digital image contaminated by noise of unknown magnitude that is represented by a reference digital image; and means for applying a process that uses a variance minimization analysis with respect to 15 the unknown magnitude associated with the target digital image and the reference digital image, to determine the magnitude of the noise for at least some pixels of the target digital image, wherein the determination using the variance minimization analysis is statistically equivalent to a determination of the magnitude of the noise made from a decorrelation condition: 20 Correlation (the target digital image - the magnitude * {the reference digital image}, {the reference digital image}) = 0.

22. A method or an apparatus substantially as hereinbefore described with reference to one or more of the accompanying drawings. 25 SI 1909747 i rlHMntterei P801iaAU 1