WO2017153410A1

WO2017153410A1 - Method for generating a noise-reduced image based on a noise model of multiple images, as well as camera system and motor vehicle

Info

Publication number: WO2017153410A1
Application number: PCT/EP2017/055328
Authority: WO
Inventors: David Hurych; Pavel Krizek; Jiri Kula
Original assignee: Connaught Electronics Ltd.
Priority date: 2016-03-07
Filing date: 2017-03-07
Publication date: 2017-09-14
Also published as: DE102016104043A1

Abstract

The invention relates to a method for generating a noise-reduced image (l_D), in which a noisy image (l_N) is captured by a camera (3) of a motor vehicle (1) and the following steps are performed: a) capturing an image sequence (10) of images (9) by the camera (3); b) determining a noise model (NLF) based on multiple images (9) of the image sequence (10); c) determining a weighting value (w_xy) for weighting the noisy image (l_N) as well as at least one image (l_B) depending on the noisy image (l_N) based on the noise model (NLF); d) generating a blurred image (l_B) based on the noisy image (l_N); e) generating a weighted blurred image by weighting the blurred image (l_B) depending on the weighting value (w_xy); f) generating a weighted noisy image by weighting the noisy image (lN) depending on the weighting value (w_xy); and g) generating the noise-reduced image (l_D) based on the weighted blurred image and the weighted noisy image.

Description

Method for generating a noise-reduced image based on a noise model of multiple images, as well as camera system and motor vehicle

The invention relates to a method for generating a noise-reduced image. A noisy image is captured by a camera of a motor vehicle and an image sequence of images is captured by the camera. The invention also relates to a camera system for a motor vehicle as well as to a motor vehicle with a corresponding camera system.

Methods for generating a noise-reduced image are known from the prior art. Thus, US 2014/0126808 A1 describes an approach to denoise a digital camera image. The approach is based on direct measurement of the local statistic structure of natural images in a large training set, which are impaired by faked digital camera noise.

One designates the deterioration of a digital or electronically captured image by disturbances, which do not have any relation to the actual image content, the image signal, as image noise. The disturbing pixels deviate from those of the actual image in color and brightness. Therein, a so-called signal to noise ratio is a measure of the noise component. The appearance of the image noise is not directly comparable to the so-called "grain" in the photography on conventional film material, but has similar effects on the technical image quality, in particular the detail resolution.

In electronic image sensors, such as for example CCD (charge-coupled device) sensors or CMOS (complementary metal-oxide-semiconductor) sensors, the image noise is dark noise in a large part. Thus, it occurs without light being incident on the sensor. The reason for this image noise is the dark current of the individual light sensitive image sensor elements on the one hand, also noise of a read-out amplifier of the image sensor on the other hand (read-out noise).

It is the object of the invention to provide a method, a camera system as well as a motor vehicle, by which a more noise-reduced image can be generated in more detail conserving manner.

According to the invention, this object is solved by a method, by a camera system as well as by a motor vehicle having the features according to the respective independent claims. In a method according to the invention, a noise-reduced image is generated. A noisy image is captured by a camera of a motor vehicle and the following steps are performed: a) capturing an image sequence of images by the camera;

and as an essential idea of the invention

b) determining a noise model based on multiple images of the image sequence;

c) determining a weighting value for weighting the noisy image as well as at least one image depending on the noisy image based on the noise model;

d) generating a blurred image based on the noisy image;

e) generating a weighted blurred image by weighting the blurred image depending on the weighting value;

f) generating a weighted noisy image by weighting the noisy image depending on the weighting value; and

g) generating the noise-reduced image based on the weighted blurred image and the weighted noisy image.

The invention is based on the realization that the weighting value can be more precisely determined by determining the noise model based on the images of the image sequence. The weighting value in turn provides the basis for combining the noisy image with the blurred image. Thereby, more informative details are conserved in the noise-reduced image than it would be the case in combining the noisy image with the blurred image without the weighting value.

The noisy image has image noise. Thus, the noisy image is for example generated such that an image sensor of the camera is impaired by undesired thermal influence during the capture of the noisy image or the image sensor is exposed to undesired electromagnetic radiation during the capture of the noisy image.

In order to reduce the image noise in the noisy image, now, a global smoothing filter, as for example a Gaussian filter, can for example be applied to the entire noisy image in unchanged manner - as in known methods. This is disadvantageous in that informative details and structures of the noisy image, such as for example edges, thus for example contours of objects, are herein also impaired. In order to prevent this, a locally adapted noise reduction is performed by the method according to the invention. Thus, the weighting value can be determined in particular for various areas of the noisy image, whereby the various areas can then be smoothed in differently intense manner.

Preferably, the area of the image is respectively described as an individual pixel, in particular as an individual intensity value of a color channel of the pixel. In particular, thus, denoising of the noisy image can thereby be effected with the accuracy of one pixel. Thus, the noisy image is thereby preferably individually adapted or denoised for each pixel of the image, in particular each intensity value of an image channel of the noisy image.

Preferably, it is provided that the noise-reduced image is generated by a linear

combination, in particular by summation, of the weighted blurred image and the weighted noisy image. Thus, the weighting value can for example be between 0 and 1 , wherein the combination of the weighted blurred image and the weighted noisy image then together preferably again arrives at a weighting value of 1 . In particular, the linear combination is performed on the level of the intensity values of the image channels of the weighted blurred image and the weighted noisy image, respectively. Thus, the intensity values are then proportionally added to each other depending on the weighting value to generate a new intensity value of the noise-reduced image. By the linear combination, the noise- reduced image is generated adapted to situation respectively for different image areas in simple manner. For example, this can then be formulated as follows:

I_D (x, y) = (l - w_xy )I_N (x, y) + w_xyI_B (x, y) , with l_D = noise-reduced image;

l_N = noisy image;

l_B = blurred image;

w_xy = weighting value;

x = column coordinate of an intensity value of the respective image;

y = line coordinate of the intensity value of the respective image.

Furthermore, it is preferably provided that before step b) at least each one image area of the images of the image sequence is selected based on at least one predetermined selection criterion and the noise model is accordingly only determined based on the respective image areas in step b). In particular, the steps c) to g) are preferably also only performed for the respective image areas. The respective image area preferably includes a plurality of pixels, in particular 1 00 pixels. However, the image area can also include only one pixel. The noise model is then preferably determined for each of the pixels. In particular, the noise model is determined for each of the pixels with a different intensity value. Thereby, the noise model can be provided for each of the intensity values. Thereby, the noisy image can be pixel-precisely weighted and thus be pixel-precisely combined with the blurred image. In particular, image areas of the images of the image sequence are determined by the selection criterion, which substantially show the same location of the environmental region unchanged over the capture period of time of the image sequence. Thus, by the selection criterion, in particular those image areas are therefore selected, which have a static background and a static foreground. This is advantageous since thereby variations of the intensity values of the image, which relate to a movement of the camera and/or a movement of an object in the images, can be effectively excluded. As the image areas, only those pixels remain, which should not have any intensity variation apart from the noise. Thereby, the noise model can then be determined in step b) only from intensity values, which are associated with static pixels and thus are substantially varied by the noise to be modeled by the noise model. Thus, the noise model can be more precisely and more faultlessly determined by the predetermined selection criterion.

Preferably, it is provided that a variation rate of an average intensity value, in particular an arithmetic average, of locally corresponding intensity values of the images of the image sequence over the temporal progress of the image sequence is determined as the selection criterion of the image areas, and the image areas are selected if the variation rate is less than a predetermined variation limit value. Thus, if there should neither move the camera nor an object in the image, thus, it can be assumed that the average intensity value of the locally corresponding intensity values has a variation rate, which is less than the predetermined variation limit value. Thus, the locally corresponding intensity values are pixels of consecutively captured images, which depict the same location of a scene, for example of an environmental region of the motor vehicle. By the variation rate of the average intensity value, the image areas are simply and reliably selected to the effect that the respective intensity value can be incorporated in the noise model without being affected by a movement of the camera and/or a movement of an object in the image.

Furthermore, it is preferably provided that a distribution curve of differences of respective locally corresponding intensity values of the images of the image sequence with an average intensity value, in particular an arithmetic average, of the locally corresponding intensity values of the images is determined as the selection criterion of the image areas, and the image areas are selected if the distribution curve of the differences has an average value of 0 and/or the distribution curve of the differences is symmetrical to the ordinate axis extending to the maximum of the distribution curve. Thus, it can be recognized based on the distribution curve if the respective image area is captured from a static region or if objects move in the image area or the camera moves in capture of the image area. Hereto, the differences of the respective locally corresponding intensity values are considered as distribution, wherein the distribution is represented by the distribution curve. The distribution curve in turn can then be assessed based on the average value and/or the symmetry to select the image areas.

Preferably, it is provided that the steps b) to g) are respectively performed for at least one intensity value of a color channel of the noisy image, in particular an RGB color channel of the noisy image or a YUV color channel of the noisy image. In that the steps b) to g) are performed for the intensity value of the color channel of the noisy image, the noise- reduced image can be locally generated in locally precise manner. In contrast to global or coarse image improvement methods, in the method according to the invention, an intensity value-precise generation of the noise-reduced image can thereby be effected, whereby the noise in the noisy image can be considerably reduced, but at the same time contours of objects can further be obtained in sharp manner or intense in contrast in the noisy image. However, the noisy image can also be present in other color channels than the RGB color channel or the YUV color channel.

Furthermore, it is preferably provided that the noise model is determined in step b) based on intensity values of locally corresponding pixels of the images of the image sequence. Thereby, the noise model can be locally precisely applied for weighting the noisy image and/or for weighting the blurred image. Thereby, the noise-reduced image is generated precisely and sharply for an observer and free of noise at the same time.

Furthermore, it is preferably provided that the locally corresponding pixels of the images of the image sequence are combined with respect to respective average intensity values in determining the noise model in step b) and a noise characteristic, in particular a standard deviation, for at least one value from a range of values of the intensity values of the images of the image sequence is respectively provided by the noise model. Thus, the noise model is first determined preferably for each intensity value of the images of the image sequence. Then, the noise model can provide the noise characteristic preferably for each value from the range of values of the intensity values of the images, thus for example for an 8 bit image from 0 to 255. The noise characteristic is in particular described by the standard deviation of the respective value. Thereby, a noise

characteristic, which is formed over multiple pixels of the images of the image sequence, can then for example be assigned to each intensity value of the noisy image.

Furthermore, it can preferably be provided that an in particular arithmetic average is respectively formed from the locally corresponding pixels of the images of the image sequence and the noise model is determined depending on all of the corresponding pixels with the same average value. Thus, the noise characteristic for a single intensity value is no longer determined only based on a pixel of the images of the image sequence, but preferably based on all of the pixels of the images of the image sequence with the same average value. By combining the noise characteristic with respect to the average value, a particularly reliable and significant noise characteristic for the value from the range of values of the intensity values is provided by means of the noise model. Thus, influences of erroneous individual standard deviations of a single intensity value can thereby for example be recognized and compensated for by the amount of the intensity values with the same average value. Thus, the effect of outliers in the determination of the noise model can be reduced. Thereby, the noise model can be still more precisely and flawlessly applied for weighting the blurred image and/or the noisy image.

In a further embodiment, it is preferably provided that after step a) a further image is captured by the camera depending on an update criterion and the noise model is again determined additionally considering the further image. Thus, by the update criterion, in particular a so-called online update of the noise model is performed. Thereby, the noise model can for example then be adapted from time to time. For example, a predetermined period of time or else a predetermined number of captured images can be preset as the update criterion. However, the update criterion can for example also be determined by how a brightness state in an environmental region of the camera varies. Thus, the update criterion can for example depend on a value of automatic exposure setting of the camera. For example, the update criterion can be preset as a variation of a signal gain value of the camera or an exposure time variation of the camera. The update criterion is advantageous in that the noise model can thereby always be provided currently adapted to the noise causing circumstances and thus the noise-reduced image can always be generated in a detail conserving and informative manner.

Preferably, it is provided that the weighting value is determined depending on a magnitude of an edge of the noisy image, and the blurred image is the less severely taken into account in weighting in step e) the higher the magnitude of the edge is and/or the noisy image is the more severely weighted in weighting in step f) the higher the magnitude of the edge is. By weighting the blurred image and the noisy image depending on the magnitude of the edge of the noisy image, it is allowed that areas of the noisy image are maintained, which for example contain a contour of an object, but homogeneous areas of the image, which then in particular do not contain contours of objects, are again considerably smoothed. Thus, the noise reduction of the noisy image can thereby be particularly accurately effected in locally adapted manner. By the magnitude of the edge, it is indicated how severely the edge differs from the adjacent intensity values.

Furthermore, it can be provided that step a) or step b) is performed at a point of time, at which an automatic exposure setting of the camera is performed and/or at which the motor vehicle is statically operated. Thus, the image sequence of the images is preferably captured when at least the camera of the motor vehicle is unmoved and thus the respective selection criterion can be more easily satisfied. In particular, the noise model is then respectively newly determined when the automatic exposure setting of the camera is newly performed. For example, this can be the case in a variation of a brightness state in the environmental region of the camera. Because after the variation of the brightness state in the environmental region of the camera, another noise characteristic can also be assumed, the determination of the noise model at this point of time is particularly effective. At this point of time, the new noise model then considerably differs from the old noise model. However, step a) and/or step b) can also be performed directly after activating the camera, for example upon starting the motor vehicle.

Preferably, it is provided that the blurred image is determined in step d) by convolution of the noisy image with a Gaussian kernel, in particular with an average value of 0. Thus, the blurred image in step d) is determined by the application of the Gaussian kernel or a Gaussian filter to the noisy image. The Gaussian kernel is in particular present as a two- dimensional filter and can be applied in various pixel sizes, for example 3x3, 5x5 or 7x7. A standard deviation of the Gaussian kernel can also be preset with various values.

The invention also relates to a camera system for a motor vehicle with a camera and an evaluation unit, which is formed to perform a method according to the invention. The camera system can include multiple cameras. Therein, the camera has a motor vehicle fixing element for fixing to the motor vehicle.

For example, the camera system can be a component of a driver assistance system of the motor vehicle.

Furthermore, the invention relates to a motor vehicle with a camera system according to the invention. The preferred embodiments presented with respect to the method according to the invention and the advantages thereof correspondingly apply to the camera system according to the invention as well as to the motor vehicle according to the invention.

Further features of the invention are apparent from the claims, the figures and the description of figures. The features and feature combinations mentioned above in the description as well as the features and feature combinations mentioned below in the description of figures and/or shown in the figures alone are usable not only in the respectively specified combination, but also in other combinations without departing from the scope of the invention. Thus, implementations are also to be considered as encompassed and disclosed by the invention, which are not explicitly shown in the figures and explained, but arise from and can be generated by separated feature combinations from the explained implementations. Implementations and feature combinations are also to be considered as disclosed, which thus do not have all of the features of an originally formulated independent claim. Moreover, implementations and feature combinations are also to be considered as disclosed, in particular by the explanations set out above, which extend beyond or deviate from the feature combinations set out in the relations of the claims.

Below, the embodiments of the invention are explained in more detail based on schematic drawings.

There show:

Fig. 1 a schematic plan view of an embodiment of a motor vehicle according to the invention with a camera system;

Fig. 2 a schematic representation of a noisy image captured by a camera of the camera system;

Fig. 3 a diagrammatic representation of weighting values of a noise model for generating a noise-reduced image based on the noisy image and a blurred image generated based on the noisy image;

Fig. 4 a schematic representation of an image of an image sequence provided for determining the noise model, which is captured by the camera; a schematic representation of image areas of the image of the image sequence, which are selected based on two predetermined selection criteria; a schematic representation of a first image, a second image and a third image of the image sequence with locally corresponding intensity values for determining the second selection criterion; a schematic representation of a distribution curve of differences of respective locally corresponding intensity values of the images of the image sequence to determine the second selection criterion; a schematic representation of the images of the image sequence and a locally corresponding intensity value of the images of the image sequence, which is registered in a vector; a diagrammatic representation of noise characteristics of the noise model, which is determined based on the images of the image sequence; a diagrammatic illustration of exemplary standard deviations of the noise characteristic of the noise model; a schematic representation of identical intensity values of the images of the image sequence for an intensity value of 100, which are combined and collectively contribute to the noise characteristic of an intensity value of the noise model; and a schematic representation analogously to Fig. 1 1 , but for an intensity value of 200.

In the figures, identical or functionally identical elements are provided with the same reference characters. In Fig. 1 , a plan view of a motor vehicle 1 with a camera system 2 is schematically illustrated. The camera system 2 includes a camera 3 and an evaluation unit 4. According to the embodiment, the camera 3 is disposed at a front 5 of the motor vehicle 1 . However, the arrangement of the camera 3 is variously possible, however, preferably such that an environmental region 6 of the motor vehicle 1 can at least partially be captured by the camera 3. The arrangement of the evaluation unit 4 is also variously possible, however, preferably such that it can be connected to the camera 3. For example, the evaluation unit 4 can be integrated in the camera 3 or be formed separately from the camera 3. The camera system 2 can for example also include multiple cameras 3. Then, the evaluation unit can 4 for example be connected to multiple cameras 3 or else also be formed by multiple partial evaluation units.

In particular, the camera 3 is formed as a CMOS (complementary metal-oxide- semiconductor) camera or as a CCD (charge-coupled device) camera or else as a versatile image capturing device. The camera 3 is in particular formed as a video camera, which continuously provides an image sequence of frames.

The camera 3 has an objective 7 and an image sensor 8. Therein, the image sensor 8 can add noise to a picture in generating the picture of the environmental region 6 by different impairment of heat. However, the noise added to the picture can for example also be added to the picture by electromagnetic influences on the image sensor 8.

Fig. 2 shows a noisy image l_N. The noisy image l_N is captured from the environmental region 6 of the motor vehicle 1 . In the noisy image l_N, noise in particular caused by the image sensor 8 is contained. A noise-reduced image l_D is now generated from the noisy image l_N. Hereto, a blurred image l_B is generated based on the noisy image l_N. In order to generate the noise-reduced image l_D, then, the noisy image l_N and the blurred image l_B are weighted by a weighting value w_xy and linearly combined. The weighting value w_xy is provided by a noise model NLF. The linear combination and thus the generation of the noise-reduced image l_D can be mathematically represented as follows. i_D ( > y) = (i - ^wxy )i_N ( > y) + ^w _Xy ( > y) 0 )

A position of an intensity value of a color channel or of a pixel of the respective image l_D, l_N, IB is described by the line coordinate x and the column coordinate y. The blurred image l_B is generated by a convolution of the noisy image l_N by means of a discrete two-dimensional Gaussian kernel with an average value of 0, a filter dimension [sxs] of the Gaussian kernel and a standard deviation of σ_Β. By applying the Gaussian kernel to the noisy image l_N, intense gradients are reduced. Thus, the noisy image l_N is smoothed or rendered blurred by the Gaussian kernel.

The filter dimension [sxs] and σ_Β of the blurred image l_B as well as a gradient magnitude normalization constant g are preferably determined based on a number C of images with a respective image dimension [XxY]. This is effected by optimization of the following mathematical expression.

C X Y ²

arg min∑∑∑ (/, (x, y) - f (V_N (x, y), s, σ_Β , g\\ (2)

S,<7_B ,g j₌l _X=l y₌l

Herein, f is a denoising function, I, is an ideal image without noise and I_N ^] is the j-th noisy image, which is captured by the camera 3. These parameters are in particular only once determined and can then for example be used for all of the noise models NLF of any camera.

The weighting value w_xy can be mathematically described as follows:

The maximization of the term is provided in case that a normalized gradient is greater than NLF(l(x,y)) and thus a result would have a negative value. The weighting value in particular has a range of values of w_xy e (0,1) . The partial derivatives are estimated by the following equations. di_N (x, y)

+ l, y) - I_N (x - l, y) (4) dx dl_N (x, y)

= I_N (x, y + l) - I_N (x, y -l) (5) In the equation (4) and the equation (5), the normalized gradient magnitude describes a local intensity of a texture of the noisy image l_N. The more intensely the texture is formed, the less intensely the blurred image l_B is weighted in equation (1 ) and the more intensely the noisy image l_N is weighted. In particular, each intensity value of the noisy image l_N and of the blurred image l_B obtains an own weighting value w_xy, which indicates how severely it is denoised or smoothed.

Hereto, Fig. 3 shows, which weighting values w_xy are applied for the noise-reduced image

Preferably, the noisy image l_N is present in an RGB color space or in a YUV color space. Thereby, the method is preferably performed for each intensity value l_N(x,y) of the noisy image l_N. In particular, the method is applied to provide the noise-reduced image l_D in real time, thus in the time, in which a new noisy image is not yet captured by the camera.

The noise model NLF, based on which the weighting value w_xy is determined, is now determined as follows. First, locally corresponding image areas p_a of images 9 of an image sequence 10 captured by the camera 3 are determined. Therein, the image area p_a preferably does not overlap with adjacent image areas and can for example be described as follows. p_a = (x, y,W, H)^T (6)

Herein, the position of an upper left corner coordinate is described by x and y. A width of the image area p_a is described by W and a height of the image area p_a is described by H. The index a describes an ordinal number for the linkage to a certain image 9 from the image sequence 10. The x and y coordinates are linked to a certain image area p_a while the width W and the height H are the same for all of the image areas p_a. The weight W can for example be preset with a parameter value of ten pixels and the height H can for example also be preset with a parameter value of ten pixels. The image areas p_a are in particular separately examined for each image channel of the respective color space. In the further, the description is only presented for a respective intensity value l(x,y) of one of the images 9 of the image sequence 1 0.

In order to determine if the respective image area p_a is suitable to be incorporated in the noise model NFL, the respective image area p_a is examined with respect to a first selection criterion 1 1 and a second selection criterion 12. The first selection criterion 1 1 implies that an average intensity value δ_α' over a time t has a variation rate below a predetermined variation limit value ξ . This can be mathematically described as follows. The average intensity value δ_α' can be described as follows. W-l H-l

f (7)

Whereby the first selection criterion 1 1 can then be defined as follows.

Wherein it is stated by the equation (8) that the maximum variation rate of the average intensity value δ_α' over a number A of all of the images 9 is to be less than the

predetermined variation limit value ξ , which is for example one to two percent of the value range of intensity values of the images 9. By the number A, for example, ten images 9 are specified. For this purpose, an overall average intensity value δ_α ^} is subtracted from the average intensity value δ_α' . The image areas p_a, which satisfy the condition from equation (8) of the first selection criterion 1 1 , are forwarded for examining the second selection criterion 12. For the second selection criterion 12, it is assumed that the noise in the noisy image l_N substantially follows a Gaussian distribution. Thus, it can be assumed that if differences of locally corresponding intensity values of the images 9 do not have a distribution curve, which substantially corresponds to a Gaussian distribution, thus, the image area p_a is not suitable with respect to the second selection criterion 12.

Thus, the second selection criterion 12 implies that differences A_j, of locally corresponding intensity values Vji should have a distribution curve over the time t, which shows an average value of 0 or else is formed symmetrically to an ordinate axis extending to the maximum of the distribution curve. The locally corresponding intensity values Vji are mathematically represented as follows. v_jl (t) = I'(p_ax + j,p_ay + l) t e {l,...,A} (9)

Hereby, a vector is preset, in which the intensity values of a pixel of an image area are preset over the changing time and within an interval j and an interval I. Therein, the interval j corresponds to j e {0,...,W - 1} and the interval I therein corresponds to l e {0,...,H - l} . The differences A_j, of the locally corresponding intensity values Vj, are obtained by subtracting an average vector b.

A_jl (t) = v_jl (t) ^~ _jv_jl (b) te {l,...,A} (10)

-A b=l

Thereby, a sum of the distributions results, which is described as follows.

Herein, a distribution curve Λ is provided, which represents the distribution of the differences in particular by a discrete histogram. Herein, the maximum of the sum of the distributions is searched and it is examined if it is close to 0. If this is the case, then, the distribution curve Λ is examined with respect to its symmetry to the ordinate axis 13 extending to the maximum of the distribution curve Λ. The symmetry can simply be examined by subtracting an integral from the part of the distribution curve Λ disposed to the left of the ordinate axis 13 from an integral of a part of the distribution curve Λ disposed to the right of the ordinate axis 13. If the difference is less than a predetermined difference limit value, it can be assumed that the distribution curve Λ is symmetrical and thus corresponds a sum of Gaussian distribution.

Fig. 4 shows one of the images 9 of the image sequence 10. The images 9 in particular occur with the number A in the image sequence 10. In the images 9, the image areas p_a are determined.

Fig. 5 shows the examined image areas p_a of the images 9, which have been determined based on the image shown in Fig. 4. Thus, the hatched image areas 14 satisfy both the first selection criterion 1 1 and the second selection criterion 12, while the non-hatched image areas 15 do not satisfy at least one of the selection criteria 1 1 , 12.

Fig. 6 shows a first image 16 of the images 9, a second image 17 of the images 9 and a third image 18 of the images 9. The images 16, 17, 18 are captured consecutively in time In the images 16, 17, 18, the locally corresponding image area p_a is drawn. In the image area p_a, first locally corresponding intensity values v_{1 ;} second locally corresponding intensity values v₂ and third locally corresponding intensity values v₃ are drawn. Therein, the locally corresponding intensity values v_{1 ;} v₂, v₃ are in particular preset as a vector.

Fig. 7 shows a first distribution curve Λ_{1 ;} a second distribution curve Λ₂ and a third distribution curve Λ₃, which are formed as follows. = {v_l - mean{v_l )) (12) Λ₂ = (v₂ - mean(v₂)) (13) K = (v - mean(v )) (14)

A fourth distribution curve 19 is described by equation (15).

Thus, on the ordinate of the diagram of Fig. 7, a standard deviation of the respective locally corresponding intensity values Vj, is in particular indicated.

Fig. 8 now shows the images 9 of the image sequence 10, based on which the noise model NLF is generated. In particular, the respective noise model NLF is determined for each color channel of the images 9 of the image sequence 10. Therein, the color channels can for example be present in an RGB color space or else in a YUV color space. For simplicity, the creation of the noise model NLF is explained in the following based on a color channel of the RGB color space, wherein the method is also applicable to the YUV color space. The noise model NLF is mathematically described as follows.

NLF(I(x, y))→a (16)

Wherein a noise characteristic σ or a standard deviation of the noise of a determined intensity value l(x,y) of an image of the images 9 of the image sequence 10 describes as an output. In particular, the noise model NLF is determined for the various color channels, thus for example a red color channel, a blue color channel and a green color channel. It is shown, how the intensity values of the pixels from the images 9 over the time can be combined to a vector v. Fig. 9 shows a diagram, in which an intensity value 21 of a red color channel of the images 9 of the image sequence 10 is plotted on an abscissa 20. The noise characteristic σ or a standard deviation of the noise is registered on an ordinate 22. Thus, the noise model NLF is described for values i with the range of values of i e {0,1,...,255}. Thereby, thus, for each intensity value l(x,y) contained in the images 9, a noise characteristic σ in the form of the standard deviation is provided by the noise model NLF.

This is shown based on Fig. 10. Thus, based on the noise model NLF, a first noise characteristic σι , a second noise characteristic σ₂ and a third noise characteristic σ₃ are here derived from the noise model NLF. The first noise characteristic σι corresponds to a high standard deviation, the second noise characteristic σ₂ corresponds to a medium standard deviation and the third noise characteristic σ₃ corresponds to a low standard deviation.

Now, the noise model NLF is determined as follows. Each discrete value i of NLF (i) is calculated from the selected intensity values over the time. Herein, i=l(x,y) and ie {0,1,...,255}. The noise model NLF is calculated over all of the pixels, which are within the previously selected image areas p_a and have the same average intensity value over the time t. A standard deviation over all of these pixels provides the noise model NLF in the position i. This is schematically explained based on Fig. 8. The vector v also has an average value m_v. It is mathematically described as follows, wherein n is a number of the images 9 of the image sequence 10.

It is assumed that after the image areas p_a satisfy the selection criteria 1 1 , 12, thus do not move, and lighting variations are not to be expected, they are suitable for modeling the noise. Thus, the variations of the intensity values in the vector v are substantially only caused by the noise. For the estimation of the noise model NLF, the noise model NLF is formed over all of the pixels of the images 9, in which the average value m_v of the vector v is the same over the time and thus i=m_v is true. Thus, the noise model NLF has for example been modeled for an intensity value of i=100 for all of the vectors v, which have the average value of m_v=100. Pixels of the image 9, which are encompassed by the noise model NLF for the intensity value of i=100, are marked in Fig. 1 1 . Thus, all of the pixels are characterized there by an average intensity value of 100. Analogously to Fig. 1 1 , Fig. 12 shows the noise model NLF for an intensity value of i=200.

Due to the noise, it can occur that some values of the range of values of the intensity values of the images 9 do not occur in the images 9 and thus the noise model NLF either cannot be generated for these values. Thus, this is for example the case in Fig. 9 for intensity values below i=35. Then, the noise model NLF can for example be approximately estimated for these values or else simply be left empty.

Assuming k is the number of pixels with the same average intensity value (they are the emphasized pixels of Fig. 1 1 or Fig. 12), then, all of the k vectors can be combined in an overall vector γ,. r_i = (v₁ ,v₂ ,...,v_k ) (18)

A vector of the noise differences η can then be calculated as follows.

The noise model NLF is then determined as follows.

Herein, the noise model NLF is a normalized standard deviation of the noise differences η. Furthermore, a normalization constant m is indicated, which can be described as follows.

Additionally, it can be provided that the noise model NLF is again determined or updated depending on an update criterion. Herein, the noise model NLF is then determined as follows.

_{NLF(i) =} _^^~ _i (22) Wherein, three scalars T₀, T^ and T₂ are introduced here, which together are described as T_Q. If new noise differences r, are then to be added, it is only required to form the sums of all of the scalars described by T_Q and to newly determine the noise model NLF. The scalar Tc can be mathematically described as follows.

Γ_β =∑ r₍ Or, Vare {0,1,2} (23)

Claims

Method for generating a noise-reduced image (l_D), in which a noisy image (l_N) is captured by a camera (3) of a motor vehicle (1 ) and the following steps are performed:

a) capturing an image sequence (10) of images (9) by the camera (3);

characterized by

b) determining a noise model (NLF) based on multiple images (9) of the image sequence (10);

c) determining a weighting value (w_xy) for weighting the noisy image (l_N) as well as at least one image (l_B) depending on the noisy image (l_N) based on the noise model (NLF);

d) generating a blurred image (l_B) based on the noisy image (l_N);

e) generating a weighted blurred image by weighting the blurred image (l_B) depending on the weighting value (w_xy);

f) generating a weighted noisy image by weighting the noisy image (l_N) depending on the weighting value (w_xy); and

g) generating the noise-reduced image (l_D) based on the weighted blurred image and the weighted noisy image.

Method according to claim 1 ,

characterized in that

the noise-reduced image (l_D) is generated by a linear combination, in particular by summation, of the weighted blurred image and the weighted noisy image.

Method according to claim 1 or 2,

characterized in that

before step b) at least each one image area (p_a) of the images (9) of the image sequence (10) is selected based on at least one predetermined selection criterion (1 1 , 12) and the noise model (NLF) is then determined in step b) only from the respective image areas (p_a), in particular the steps c) to g) are only performed for the respective image areas (p_a).

4. Method according to claim 3,

characterized in that

a variation rate of an average intensity value ( δ_α' ) of locally corresponding intensity values of the images (9) of the image sequence (10) over the temporal progress of the image sequence (10) is determined as the selection criterion (1 1 ) of the image areas (p_a) and the image areas (p_a) are selected if the variation rate is less than a predetermined variation limit value ( ξ ) .

5. Method according to claim 3 or 4,

characterized in that

a distribution curve (Λ) of differences (λ ,) of respective locally corresponding intensity values (v_j,) of the images (9) of the image sequence (10) with an average intensity value of the locally corresponding intensity values (v_j,) of the images (9) is determined as the selection criterion (12) of the image areas (p_a) and the image areas (p_a) are selected if the distribution curve (Λ) of the differences ( A_j,) has an average value of zero and/or the distribution curve (Λ) of the differences ( A_j,) is symmetrical to the ordinate axis (13) extending to the maximum of the distribution curve (Λ).

6. Method according to any one of the preceding claims,

characterized in that

the steps b) to g) are each performed for at least one intensity value of a color channel of the noisy image (l_N), in particular an RGB color channel of the noisy image (l_N) or a YUV color channel of the noisy image (l_N).

7. Method according to any one of the preceding claims,

characterized in that

the noise model (NLF) is determined in step b) based on intensity values (I) of locally corresponding pixels (v) of the images (9) of the image sequence (10).

8. Method according to claim 7,

characterized in that

the locally corresponding pixels (v) of the images (9) of the image sequence (10) are combined with respect to their respective average intensity values in determining the noise model (NLF) in step b), and by the noise model (NLF) a noise characteristic (σ), in particular a standard deviation, is respectively provided for at least one value (i) from a range of values of the intensity values ( I) of the images (9) of the image sequence (10).

9. Method according to claim 8,

characterized in that

an average value (m_v) is respectively formed from the locally corresponding pixels (v) of the images (9) of the image sequence (10), and the noise model (NLF) is determined depending on all of the corresponding pixels (v) with the same average value (m_v).

10. Method according to any one of the preceding claims,

characterized in that

after step a) a further image is captured by the camera (3) depending on an update criterion and the noise model (NLF) is again determined additionally considering the further image.

1 1 . Method according to any one of the preceding claims,

characterized in that

the weighting value (w_xy) is determined depending on a magnitude of an edge of the noisy image (l_N), and the blurred image is the less severely taken into account in weighting in step e) the higher the magnitude of the edge is and/or the noisy image (IN) is the more severely weighted in weighting in step f) the higher the magnitude of the edge is.

12. Method according to any one of the preceding claims,

characterized in that

step a) and/or step b) are performed at a time, at which an automatic exposure setting of the camera (3) is performed and/or at which the motor vehicle (3) is statically operated.

13. Method according to any one of the preceding claims,

characterized in that

the blurred image is determined in step d) by a convolution of the noisy image (l_N) with a Gaussian kernel, in particular with an average value of zero.

14. Camera system (2) for a motor vehicle (1 ), with a camera (3) and an evaluation unit (4), which is formed to execute a method according to any one of the preceding claims.

15. Motor vehicle (1 ) with a camera system (2) according to claim 14.