Camera and method for detecting image data

Abstract

The present application relates to a camera and a method for detecting image data. A camera (10) is provided having an image sensor (18) and a control and evaluation unit (24) for processing the image data. The image sensor has a first recording channel with a first sensitivity for detecting first image data having a first pixel and a second recording channel with a second sensitivity lower than the first sensitivity for detecting second image data having a second pixel, the first pixel and the second pixel being associated with each other by recording the same object area, the control and evaluation unit being designed to suppress the influence of noise in the second image data using a noise reduction filter, the noise reduction filter assigning a new value to the respective observed second pixel based on the second pixels in the neighborhood of the observed second pixel. The noise reduction filter takes the second pixels in the neighborhood into account with a weighted value, the weighting being dependent on the degree of similarity between the first pixel associated with the second pixel and the first pixel (44) associated with the respective observed second pixel.

Description

[0001] The present invention relates to a camera and a method for detecting image data.

[0002] Image data from cameras is used in many industrial and logistics applications. This allows for the automation of various processes. In addition to various measurement, manipulation, and inspection tasks, it is also known to automatically classify objects based on codes read using image data. This involves reading barcodes and various 2D codes, such as Maxicode or Aztec codes, or reading labels decoded using optical character recognition (OCR).

[0003] In typical application scenarios, such as on production conveyors, in airport baggage handling, or in the automated sorting of parcels in logistics centers, objects are conveyed past a camera, and their image data is recorded. Line scan cameras are particularly well-suited for this relative motion between object and camera, as they achieve exceptionally high resolution and speed. Consecutively detected individual image lines are combined based on the known or measured conveyor speed. This is particularly advantageous in fast conveyor applications, where high frame rates are required, and where simple concatenation of image lines is feasible without requiring extensive computational effort, compared to stitching individual images from a matrix camera.

[0004] Traditional line scan cameras typically record only monochrome images, also known as grayscale or black and white. This achieves the best photon yield and, therefore, the best signal-to-noise ratio. Matrix cameras are more commonly used to record color images. The most common way to achieve color is by applying two green filters, one red filter, and one blue filter per pixel in a Bayer pattern. However, alternative color modes exist, such as adding a white channel (RGBW) or using subtractive primaries such as red, yellow, and blue (RYBY).

[0005] Color line scan cameras are also known. For example, they feature three rows of red, green, and blue (RGB) primary colors, alternating across a row, or mimicking a Bayer pattern, with one row consisting of alternating red and blue pixels and the second row consisting of pure green. The disadvantage of all these color image sensors is that the received light is lost due to color filters, making black and white images more suitable for applications such as code reading, which require higher resolution. Therefore, color information can only be obtained at the expense of poor performance in black and white evaluation.

[0006] The signal-to-noise ratio (SNR) is a key variable for image data quality and user visualization. Color photographs affect the SNR because color filters reduce the number of detectable photons. In principle, the reduced SNR can be countered by increasing the illumination, extending the exposure time, increasing the aperture, or using a more sensitive image sensor. However, these measures have already been optimized, and the illumination encounters hard limits, for example due to eye safety. This does not address the specific shortcomings of color photographs compared to black and white photographs. Another possibility is image post-processing, for example by means of noise filters or neural networks. However, this is also often used and therefore does not reduce the gap between the quality of monochrome and color image data. In addition, edges and other desired structures in color images are also blurred by traditional noise filters.

[0007] DE 20 2019 106 363 U1 discloses a code reader for reading optical codes. The device uses at least one dual row as an image sensor, wherein the receiving pixels of at least one row are sensitive to white light, while the receiving pixels of the remaining rows are sensitive to only one color. This allows for recording both grayscale and color images. In some embodiments, the primary colors are reconstructed from the grayscale image and two other primary colors. However, this document does not address the described issue of the poor signal-to-noise ratio of color images.

[0008] US2010 / 0316291 A1 and US2012 / 0002066 A1 each disclose a code reading device that uses a pixel matrix to record images. Most of the pixels are monochrome, while colored pixels are arranged in a grid. These colored pixels form a Bayer pattern. US2012 / 0002066 A1 also discusses filters for image post-processing. However, this document does not consider the differences between color and black-and-white images, meaning that the signal-to-noise ratio of color images is significantly worse.

[0009] The object of the present invention is therefore to improve image recording with recording channels of different qualities, in particular recording channels for black-and-white image data and color image data.

[0010] This object is achieved by a camera and a method for detecting image data. The image sensor of the camera uses two recording channels with different sensitivities to record first image data with first pixels and second image data with second pixels. The image data overlap and therefore correspond at least partially and preferably completely to the same object area. The first pixel and the second pixel are associated with each other because their visible ranges correspond to the same part of the camera's field of view. Therefore, the associated first and second pixels record the same object structure. In this case, the two recording channels, and therefore the associated first and second pixels, differ from each other in one case, in particular spectrally. The different sensitivities or sensitivity of the two recording channels and the resulting different signal-to-noise ratios of the first and second image data are consequences to be expected rather than desired effects, the reasons for the different sensitivities not being decisive for the present invention, but existing.

[0011] The control and evaluation unit processes the image data using an image processing method. The camera is preferably designed as a code reader for reading optical codes, and its control and evaluation unit is designed to locate the code to be read in the image data and read out the content encoded therein. As a preprocessing step, noise influences on the second image data are suppressed using a noise reduction filter. To this end, the second pixels are preferably processed individually, either sequentially or at least partially simultaneously in a parallelized architecture. Each observed second pixel receives a new adjusted value based on the second pixels in its neighborhood, preferably including its original value. The observed second pixel is then modified, as is customary in image processing using filters, until all second pixels, or a desired subset, or a desired image portion, have been processed.

[0012] The present invention is based on the basic idea of ​​linking the denoising of the second image data with the similarity of the first image data. The neighborhood of the second pixel of each observed second pixel is reweighted accordingly. To this end, the first pixel associated with the observed second pixel and the neighborhood of the first pixel are used. In other words, with respect to the object area, this is the same image portion of the first image data that the denoising filter considers in the second image data. It is checked which first pixels of the neighborhood in the first image data are similar to the first pixel associated with the observed second pixel. During processing with the denoising filter, second pixels associated with similar first pixels have a greater influence on the new value of the respective observed second pixel. Accordingly, second pixels in the neighborhood of the observed second pixel whose associated first pixel is dissimilar or less similar to the first pixel associated with the observed second pixel have a smaller contribution to the denoising filter. The underlying idea is to use the similarity in the more accurate (because more sensitive) first recording channel to find those second pixels in the neighborhood in the less sensitive second recording channel that belong to the same object structure as the observed second pixel.

[0013] The advantage of the present invention is that the higher sensitivity and better signal-to-noise ratio of the first recording channel are transferred to the second recording channel in a certain manner. In the second image data, noise is reduced, thereby improving both quality and signal-to-noise ratio, with little or no loss in spatial resolution. On the other hand, conventional noise reduction filters operating only on the second image data would inevitably blur edges. The method according to the present invention remains simple and can also be executed in real time using manageable computing and storage resources.

[0014] Preferably, the control and evaluation unit is designed to take into account for the denoising filter only those second pixels in the neighborhood whose associated first pixel meets the similarity criterion with the first pixel associated with the respective observed second pixel. This is, so to speak, a digitization or binarization of the similarity: second pixels whose associated first pixel does not meet the similarity criterion are not taken into account at all. Preferably, those second pixels whose associated first pixel meets the similarity criterion are weighted equally with one another. Alternatively to the numerical similarity criterion, a weighting that depends on the similarity can be chosen. A weighting function can then be defined for the weighting of second pixels from the neighborhood of the observed second pixel, the independent variable of which is the difference between the associated first pixel and the first pixel associated with the observed second pixel. In all cases, an additional position-dependent weighting can be superimposed, for example a weighting that decreases with the distance to the observed second pixel.

[0015] Preferably, the similarity criterion has at least one threshold value, with which the difference between a first pixel in the neighborhood and a first pixel associated with the corresponding observed second pixel is evaluated. The threshold value can be set symmetrically with respect to the first pixel associated with the corresponding observed second pixel. Alternatively, different upper and lower threshold values ​​can be envisaged. The first pixel in the neighborhood that is located in the thus formed regional band (Korridor) meets the similarity criterion, while the first pixel in the neighborhood that is outside the regional band does not meet the similarity criterion.

[0016] Preferably, the noise reduction filter is an averaging filter. In this way, the neighborhood of the observed second pixel is smoothed or averaged. However, according to the present invention, the second pixel of the neighborhood depends on the similarity of the associated first pixel and the first pixel associated with the observed second pixel. As mentioned above, the weighting can be binary, that is, the second pixel contributes when there is sufficient similarity in the associated first pixel, and not otherwise, or the weighting can be numerically dependent on the similarity. In both cases, additional weighting can be added, for example a weighting that decreases with the distance to the observed second pixel.

[0017] Preferably, the denoising filter has a filter kernel that defines a neighborhood around the respective observed pixel, and the value of the filter kernel is set in particular according to the similarity of the first pixels in the neighborhood to the central first pixel. Therefore, it is common practice to use a filter kernel for convolution. According to the present invention, there is a special feature of defining a filter kernel for each second pixel observed. The filter kernel depends on the neighborhood of the specific first pixel associated with the respective observed second pixel. Therefore, this is a locally rather than globally defined filter kernel, wherein the rules for creating the filter kernel are of course preferably global. The individual values ​​or weights of the filter kernel determine the proportion in which the second pixels from the neighborhood should be merged into the new pixel value of the observed second pixel. These weights are set based on the similarity in the associated pixels of the first image data.

[0018] Preferably, the filter kernel is defined as a matrix with n pixels in all directions around the observed second pixel. Therefore, the observed second pixel can be called the central second pixel of the corresponding filtering step. The matrix centered in this way is placed in the neighborhood of the second pixel and the convolution is calculated, in particular as the sum of the point-by-point products of the filter kernel of the neighborhood and the second pixels of the neighborhood. Preferably, the observed second pixel is only located in the center, but a decentralized arrangement can also be achieved by zeroing the filter kernel. For the same reason, the matrix is ​​not a practical limitation, but is simply implemented for any neighborhood or vicinity, since zeros can be set at the edges for arbitrary geometric shapes.

[0019] Preferably, the control and evaluation unit is designed to fill the filter kernel with zeros if the similarity criterion is not met, and / or to fill the filter kernel with non-zero values ​​of equal magnitude if the similarity criterion is met. This is a digital or binary embodiment of the described filter kernel, which only considers second pixels whose associated first pixel and the first pixel associated with the observed second pixel are sufficiently similar. Obviously, the filter kernel is positioned around the observed second pixel, and a corresponding portion is placed around the associated first pixel in the first image data. Where the first pixel is not sufficiently similar to the central first pixel, zeros are set in the filter kernel to ignore this portion of the neighborhood in the denoising filter. In the remaining portion of the filter kernel, the neighborhood of the second pixel is considered, preferably with the same weighting. Alternatively, the binary entries of the filter kernel can be set according to a weighting function that depends on the weighting, as described above. Similarly, modulated weighting can be envisioned, for example, with the weighting decreasing as the distance from the observed second pixel increases.

[0020] Preferably, the control and evaluation unit is designed to count the number of times the filter kernel has a non-zero value and, in particular, to normalize the filter kernel using this number. The filter kernel has m non-zero entries that satisfy the similarity criterion. This m is determined so as to correlate with the number of correct input values ​​during averaging or other calculations. If the non-zero entries in the filter kernel have the same size and are, in particular, set to 1, then m is equal to the sum, and if necessary, a scaling factor must still be achieved. The filter kernel can now be normalized by dividing all entries by m.

[0021] Preferably, the control and evaluation unit is designed to adapt the resolutions of the first image data and the second image data to one another. Depending on the image sensor, the first image data and the second image data may initially have different resolutions. This can be compensated, for example, by interpolating the image data with a poorer resolution or by binning the image data with a better resolution. In the overlapping area of ​​the first image data and the second image data, each first pixel is then associated with exactly one second pixel. To repeat, the overlapping area corresponds to the object area that is observed jointly by the two recording channels and is preferably complete. Therefore, the 1:1 relationship between the first and the second pixel applies everywhere. Without resolution adjustment, multiple associations would be possible. Conditional denoising according to the invention is also feasible, but more special and edge cases must be considered in the implementation.

[0022] Preferably, the first recording channel is designed as a single channel sensitive to white light, for recording image data for grayscale images, also known as black and white or monochrome images. Therefore, the first recording channel has light-receiving pixels that are sensitive to white light, meaning they perceive the entire light spectrum and, for example, lack color filters. However, the unavoidable hardware limitations of the camera pixels used impose restrictions on the amount of light that can be received.

[0023] Preferably, the second recording channel is designed to include at least one color channel sensitive to a specific color, for recording image data of that color. Thus, the second recording channel includes light-receiving pixels that are sensitive only to light of the color of the color channel (e.g., due to corresponding color filters). The distribution of the light-receiving pixels of the color channels can form different patterns, depending on the embodiment.

[0024] Preferably, the second recording channel has multiple color channels of different colors. Therefore, the second image data is generated multiple times, preferably in the primary colors RGB (red, green, blue) or CMY (cyan, magenta, yellow). The corresponding monochrome images, for example, red, green, and blue images, can be filtered individually or together using the method according to the present invention. Preferably, the second recording channel has only two color channels for two of the three primary colors. Therefore, there are no color channels or light-receiving pixels sensitive to the third primary color. Particularly preferably, the control and evaluation unit is designed to reconstruct the third primary color from two primary colors using the first image data (i.e., white image data). White is a superposition of all primary colors, so the third primary color can be separated when the other two primary colors are recorded. Preferably, the two primary colors are red and blue. Generally, additive primary colors produce better results. In this preferred embodiment, the green color, which is present twice in the Bayer pattern, is not recorded, so no light-receiving pixels or color channels are required for this purpose. If necessary, the green color is generated from the white, red, and blue image information. Clearly, the green color is therefore reconstructed from G = 3*αW-βR-γB, where α, β, γ are normalization factors, and preferably also color correction is performed.

[0025] Preferably, the image sensor is designed as a line array sensor with at least two rows of light-receiving elements, wherein in particular each row is completely associated with the first recording channel or the second recording channel, respectively. The camera is therefore a line array camera. Preferably, the recording channels consist of entire rows, for example the first row forms the first recording channel and the second row forms the second recording channel. This results in the highest possible resolution in the row direction. In particular in the color channels, the following patterns can also be present instead of continuous monochrome rows, such as alternating light-receiving pixels of different colors, or a combination of uniform color rows and mixed color rows. In principle, the light-receiving elements sensitive to white can also be distributed in one row of a color channel, while the other row of the single channel is responsible for this type of image information.

[0026] Preferably, two, three, or four rows are provided. The numbers given here are exact figures, not minimums. Using several rows allows for a particularly compact design of the image sensor. Two rows result in a double row, while one or two additional rows provide a certain degree of redundancy, particularly for color modes. Furthermore, a row preferably generally belongs to a specific recording channel.

[0027] The camera is preferably mounted stationary above the stream of objects to be recorded. The image lines are preferably recorded sequentially and concatenated to form an image. Alternatively, in the case of barcodes, it is conceivable to read the code from a single linear recording, but preferably, the barcode can also be read from such a combined, planar, complete image. Due to the time delay between reading the image information of different lines, it is particularly possible to superimpose the image information so that the first and second image data correspond to the same portion of the object.

[0028] Preferably, the control and evaluation unit is designed to generate a grayscale image from the first recording channel and a color image from the second recording channel. The grayscale or black-and-white image is detected at full resolution with high contrast and an optimal signal-to-noise ratio. Simultaneously, a color image is obtained that can be used for various additional or alternative evaluations. The additional color detection does not compromise the resolution or signal-to-noise ratio of the grayscale image. On the contrary, the quality of the grayscale image can be at least partially transferred to the color image due to the noise reduction filter according to the present invention.

[0029] Grayscale images are preferably used for code reading, i.e., for reading the content encoded in the code. This allows for code reading with the same quality as conventional monochrome cameras (particularly line scan cameras). The additional color information does not affect the decoding result. Additionally or alternatively, grayscale images can be used for purposes other than code reading.

[0030] A color image is generally an image with colors that are recognizable by the human eye (e.g. RGB) and can be distinguished from a monochrome image (e.g. containing only color information for the color red). Particularly preferably, the color image is used in conjunction with and supports code reading in order to identify, classify and / or distinguish code-carrying objects and / or code areas from the image background. Typically, the background of the code is different in color from the neighborhood, or the color information can be used to identify the code-carrying object and separate it from the background. Alternatively, the color image is used for any other function, in particular as an output and subsequently applied, whether for visualization functions and diagnostic functions or completely different additional tasks. The color image may have a lower resolution than the grayscale image, which can then be adjusted as described above by upsampling / interpolation or downsampling / binning.

[0031] The method according to the invention can be further developed in a similar manner and exhibits similar advantages. Advantageous features are described exemplarily but not exhaustively in the dependent claims which are dependent on the independent claim. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Other features and advantages of the present invention will be described in more detail below based on exemplary embodiments and with reference to the accompanying drawings. In the accompanying drawings:

[0033] Figure 1 shows a schematic cross-sectional view of a line scan camera;

[0034] Figure 2 A three-dimensional view showing an application in which a line scan camera is fixedly mounted above a conveyor belt with objects, in particular for code reading;

[0035] Figure 3 shows a schematic diagram of a linear image sensor having one red row, one blue row, and one white row;

[0036] Figure 4 shows a schematic diagram of a linear image sensor having one red row, one blue row, and two white rows;

[0037] Figure 5 shows a schematic diagram of a linear image sensor having one red-blue alternating row and one white row;

[0038] Figure 6 shows a schematic diagram of a linear image sensor having two alternating red-blue rows and two white rows;

[0039] Figure 7 An example image is shown with pixels assigned new values ​​for noise reduction and the neighborhood that contributes to the new values;

[0040] Figure 8 Shown according to Figure 7 A diagram of the neighborhood of in a pixel grid;

[0041] Figure 9 Shown according to Figure 7 Example image of where the neighborhood contributing to the new value is masked out by the similarity criterion in another recording channel;

[0042] Figure 10 Shown is the corresponding Figure 9 Graph of the filter kernel for the masked neighborhood of ;

[0043] Figure 11 shows an example image before noise reduction; and

[0044] Figure 12The noise reduction according to the present invention is shown. Figure 11 .

[0045] Figure 1 A very simplified block diagram of a camera 10 in an embodiment as a line scan camera is shown. The camera 10 detects received light 12 from a detection area 14 via a photographic objective 16, which is represented here by a simple lens. A line image sensor 18 generates image data of the detection area 14 and any objects and code areas present therein. The image sensor 18 has at least two rows 20 a - 20 b of light-sensitive receiving pixels 22, with several hundred, several thousand, or even more receiving pixels 22 arranged in the row direction.

[0046] The image data of the image sensor 18 are read out by a control and evaluation unit 24. The control and evaluation unit 24 is implemented on one or more digital modules (e.g., microprocessors, ASICs, FPGAs, etc.), which can also be completely or partially arranged outside the camera 10. A preferred part of the evaluation is to string the detected image lines together to form a complete image. In addition, the image data can also be filtered, smoothed, cut into specific areas, or binarized beforehand during the evaluation. This will be referred to later. Figures 7 to 12 In the preferred embodiment of the camera 10 as a code reader, segmentation is generally performed to find individual objects and code areas. The codes in these code areas are then decoded, i.e., the information contained in the code is read out.

[0047] To illuminate the inspection area 14 sufficiently brightly with the emitted light 26, an illumination device 28 with emission optics 30 is provided. This illumination device can also be externally arranged, unlike the illustration. Data, particularly read code information and other data in various processing stages, such as raw image data, pre-processed image data, detected objects, or undecoded code image data, can be output via an interface 32 of the camera 10. Conversely, the camera 10 can be parameterized via interface 32 or another interface.

[0048] Figure 2A possible application of a camera 10 is shown, mounted on a conveyor belt 34. The conveyor belt 34 transports objects 36 through the detection area 14 of the camera 10 in a conveying direction 38, indicated, for example, by an arrow. The objects 36 may carry code areas 40 on their outer surfaces. The camera 10 has the task of identifying the code areas 40, reading the codes attached thereto, decoding these codes, and associating each with the relevant object 36. To also identify code areas 42 attached to the sides, multiple cameras 10 are preferably installed from different angles. Additional sensors, such as an upstream laser scanner for detecting the geometry of the objects 36 or an incremental encoder for detecting the speed of the conveyor belt 34, may also be provided.

[0049] The detection area 14 of the camera 10 is a flat surface with a linear reading field, corresponding to the line image sensor 18. By recording the object 36 line by line in the transport direction 38, a complete image of the passing object 36, including the code area 40, is gradually formed. Here, the lines 20a-20b are so close to each other that they actually detect the same object section. Alternatively, the offset can be compensated computationally or by reading out the lines with a small time offset.

[0050] On the one hand, the camera 10 uses the image sensor 18 to detect a grayscale image or a black and white image for code reading. In addition, color information or a color image is also obtained. The color information can be used for a number of additional functions. One example is the classification of an object 36, for example to find out whether the object is a parcel, an envelope or a pouch. It can be determined whether a conveyor belt container (for example, a pallet or a box of a pallet conveyor) is empty. The image data can be segmented into objects 36 or code areas 40 based on the color information, or such a segmentation can be supported by the color information. Additional image recognition tasks can be solved, such as recognizing specific imprints or labels, for example for hazardous materials recognition, or fonts can be read (OCR, Optical Character Recognition).

[0051] Figures 3 to 6Several examples of embodiments of image sensors 18 for detecting both black and white images and color information are shown. These embodiments share the common feature that at least one of rows 20a-20d is a monochrome or white row, whose receiving pixels 22 detect light across the entire spectral range within hardware limitations. At least another row of rows 20a-20d is a color row, whose receiving pixels 22 are sensitive only to specific colors, typically achieved through the use of corresponding color filters. The distribution of colors across the corresponding receiving pixels 22 in the color rows varies depending on the embodiment. Preferably, at least one completely white row is provided, as this allows the grayscale image to be recorded at full resolution. This also provides a clearer separation between white and color rows. However, in principle, different patterns of mixed white and color receiving pixels 22 within rows 20a-20d are conceivable. Receiving pixels 22 with identical spectral sensitivity are combined in a single channel of a grayscale image or in a single color channel of a monochrome image. For example, receiving pixels 22 sensitive to red are combined in the red color channel of a red image, while receiving pixels 22 sensitive to blue are combined in the blue color channel of a blue image.

[0052] Figure 3 An embodiment is shown having one red row 20a, one blue row 20b, and one white row 20c. Thus, the rows 20a-20c themselves are uniform, and the receiving pixels 22 within the rows 20a-20c are sensitive to the same spectrum. Figure 4 A variant is shown with one additional white row 20d.

[0053] In accordance with Figure 5 In the embodiment of FIG. 2 , receiving pixels 22 sensitive to red and blue are alternately mixed within one color row 20 a. Thus, the combination with the white row 20 b can result in a total of only two rows. Figure 6 A variation is shown in which there are two rows each of colored rows 20a-20b and white rows 20c-20d.

[0054] In accordance with Figure 5 and Figure 6 In an embodiment, the resolution of the individual color channels is different from the resolution of the single channel. It is conceivable to adjust the resolution by interpolation or binning, etc. The examples shown are only based on the selection of the primary colors red and blue and white (RBW). Other embodiments use other color filters and colors. Therefore, it is also conceivable to use green and red or green and blue (RGW, BGW) or all three primary colors (RGBW). In addition, similar combinations of the subtractive primary colors cyan (cyan), magenta (magenta) and yellow (CMW, CYW, MYW or CMYW) can also be considered. Missing primary colors can be reconstructed from white. In order to obtain a color-true image, brightness balance and / or color balance are advantageous.

[0055] The above-described line scan camera with a single channel and a color channel is a preferred embodiment. However, the invention is not limited thereto. The image sensor 18 can also have other forms, in particular, pixels arranged in a matrix in a matrix camera. Instead of a single channel and one or more color channels, it is sufficient to provide any two recording channels with different sensitivities or sensitivity to generate image data with different signal-to-noise ratios. In this case, both recording channels record the same object area and therefore record it twice. The recording areas of at least two recording channels should overlap, and the noise reduction that follows now refers to this overlapping area. In non-overlapping areas, no noise reduction can be used or a different noise reduction method can be used, in particular by setting the filter kernel values ​​for pixels from the non-overlapping area to fixed default values. Preferably, the noise reduction is implemented in the FPGA of the control and evaluation unit 24.

[0056] Figure 7 An example plot of image data from a single channel is shown. As mentioned in reference Figures 1 to 6 As described, there are two images in the two recording channels: an image recorded with a higher sensitivity in the first recording channel, as shown in the figure, and an image (not shown) of the same object area recorded with a lower sensitivity in the second recording channel. Preferably, the image in the first recording channel is a black and white image, while the image in the second recording channel is a color image. A color image can be composed of multiple (e.g., primary color) images, so the following discussion of these two images is not limited in generality.

[0057] Noise reduction should also be applied to color images in order to improve their signal-to-noise ratio. However, the corresponding noise reduction filters are generated or parameterized based on black and white images. For this purpose, pixels 44 located in the vicinity or the center of the neighborhood 46 are always considered. Off-center pixels are also conceivable. This is not discussed separately, since this is also achieved by zeros at the edges of the central noise reduction filter. In the same way, effective deviations from the rectangular shape of the neighborhood 46 can also be achieved, so this deviation is assumed without restriction. During noise reduction, each pixel is alternately the pixel 44 under observation, or at least this is the case for all pixels in the image portion of interest.

[0058] Figure 8A neighborhood 46 around an observed pixel 44 on a formalized pixel grid is shown. Based on the neighborhood 46, a filter kernel is found for the observed pixel 44. The observed pixel 44 is located at position (i, j). A portion with n×n pixels is assumed to be the neighborhood 46, where n=5 is purely exemplary. This results in the position of the neighborhood 46 shown being (i-2:i+2, j-2:j+2). As already mentioned, any deviation from the center matrix is ​​possible and can be achieved particularly easily by zeros in the filter kernel sought. The filter kernel is local: if the observed pixel 44 changes during further processing, in particular in the iterations for i and j, the filter kernel also changes. Incidentally, iteration can mean sequential processing, but parallelization is also possible for processing multiple observed pixels 44 simultaneously.

[0059] The grayscale value G(i,j) of the observed pixel 44 is known. Now, a search is conducted for pixels in the neighborhood 46 that are similar to the observed pixel 44. For this purpose, a symmetric threshold value e is determined. For each pixel in the neighborhood 46 at position (i-2:i+2,j-2:j+2), its grayscale value G is checked to see if it agrees with the tolerance given by the threshold e: G(i,j)-e≤G≤G(i,j)+e. If this similarity criterion is met, the corresponding entry in the filter kernel is set to 1, otherwise to 0.

[0060] Figure 9 Shown again Figure 7 An example image of , where a similar part 48 of the neighborhood 46 is now highlighted, where the pixels meet the similarity criterion and therefore the entries of the filter kernel are set to 1. Due to the similarity in the grayscale values, there is a high probability that this is a common object structure.

[0061] Figure 10 In accordance with Figure 8 The neighborhood 46 around the observed pixel 44 and Figure 9 48 of the similar parts. The pixels of the similar part 48 are displayed in black, while the remaining non-similar pixels are displayed in white. In the filter kernel, the pixels of the similar part 48 are initially assigned the value 1, while the remaining non-similar pixels are assigned the value 0. In an advantageous normalization step, the number m of pixels of the similar part 48 is also counted, for which a simple sum is formed over the filter kernel determined so far. The 1 in the filter kernel is then replaced by the inverse of the number m, 1 / m. In this normalized form, the filter kernel can be applied directly to the color image. Otherwise, the number m should be taken into account during filtering.

[0062] The filter kernel is created from the black and white image, but is now applied to the color image. To this end, for each observed pixel (i, j) of the color image, each pixel in the color image neighborhood is now point-wise multiplied by the relevant entry of the filter kernel, and the sum of these contributions is assigned to the observed pixel (i, j). It is thus an average over the neighborhood of the color image, based on a conventional convolution with a smoothing or averaging kernel. However, the key difference is that the filter kernel is not given globally, but is defined locally for each observed pixel (i, j) by the similarity at the same position in the black and white image. Only that part of the neighborhood of the color image that is sufficiently similar to the observed pixel 44 in the black and white image is included. In this case, the original value of the pixel (i, j) is either included in the number m and weighted equally, or it is specifically weighted more or less.

[0063] Convolution with a filter kernel is a particularly simple way to implement it, but the invention is not limited to a specific way in which similarities in the black and white image contribute to the influence of the neighborhood of the color image in the noise reduction. In addition, using a numerical similarity criterion of whether a pixel from the neighborhood contributes is simple and produces good results. Alternatively, however, a quantitative weighting can also be envisaged, which depends in particular on the degree of similarity or dissimilarity, in particular by means of a weighting function according to the gray value difference |GG(i,j)| of the individual pixels of the neighborhood 46 having the gray value G.

[0064] Figure 11 and Figure 12 Example images before and after applying the noise reduction according to the present invention are shown. For display reasons, the images are only in black and white, so it is somewhat confusing because these example images are color images from the second recording channel. In contrast, Figure 7 and Figure 9 The example image in is not only for display reasons, but is actually a black and white image of the first recorded channel, from which the corresponding local filter kernel is found depending on the similarity of the pixels in the neighborhood 46 of the observed pixel 44.

[0065] exist Figure 11 In the image, scattered dark spots and blurred edges caused by noise can be seen. After noise reduction according to the present invention, these noise effects are significantly and significantly reduced without losing spatial resolution.

Claims

1. A camera (10) comprising an image sensor (18) and a control and evaluation unit (24) for processing image data, the image sensor (18) having a first recording channel with a first sensitivity for detecting first image data having a first pixel and a second recording channel with a second sensitivity lower than the first sensitivity for detecting second image data having a second pixel, wherein the first pixel and the second pixel are associated with each other by recording the same object area, the control and evaluation unit (24) being designed to suppress noise influences in the second image data using a noise reduction filter, the noise reduction filter assigning a new value to the respective observed second pixel based on second pixels in the neighborhood of the observed second pixel, It is characterized by: The denoising filter considers second pixels in the neighborhood with a weighting that depends on the degree of similarity between a first pixel associated with the second pixel and a first pixel (44) associated with the corresponding observed second pixel, wherein during processing using the denoising filter, second pixels associated with similar first pixels have a greater influence on the new value of the corresponding observed second pixel, while second pixels associated with dissimilar or less similar first pixels have a smaller influence on the new value of the corresponding observed second pixel.

2. The camera (10) according to claim 1, in, The control and evaluation unit (24) is designed to take into account for the denoising filter only second pixels in the neighborhood whose associated first pixel meets a similarity criterion with the first pixel (44) associated with the respective observed second pixel.

3. The camera (10) according to claim 2, in, The similarity criterion has at least one threshold value, with which a difference between a first pixel in the neighborhood and an associated first pixel (44) of the corresponding observed second pixel is evaluated.

4. A camera (10) according to any one of the preceding claims, in, The noise reduction filter is an averaging filter.

5. A camera (10) according to any one of the preceding claims, in, The denoising filter has a filter kernel that defines a neighborhood around the respective observed pixel, and the value of the filter kernel is set according to the similarity of a first pixel in the neighborhood (46) to a central first pixel (44).

6. The camera (10) according to claim 5, in, The control and evaluation unit (24) is designed to fill the filter kernel with zeros if the similarity criterion is not met and / or to fill the filter kernel with non-zero values ​​of equal size if the similarity criterion is met.

7. A camera (10) according to claim 5 or 6, in, The control and evaluation unit (24) is designed to calculate a value for the filter kernel having a non-zero value.

8. A camera (10) according to any one of the preceding claims, in, The control and evaluation unit (24) is designed to adapt the resolutions of the first image data and the second image data to one another.

9. A camera (10) according to any one of the preceding claims, in, The first recording channel is designed as a single channel sensitive to white light and is used to record image data of a grayscale image.

10. The camera (10) according to any one of the preceding claims, in, The second recording channel is designed to be at least one color channel sensitive to light of a specific color, and is used to record image data of the color.

11. The camera (10) according to claim 10, in, The second recording channel has a plurality of color channels of different colors.

12. The camera (10) according to any one of the preceding claims, wherein The image sensor (18) is designed as a line array sensor having at least two rows (20a-20b) of light receiving elements (22).

13. The camera (10) according to any one of the preceding claims, The camera is fixedly mounted above the flow of the object (36) to be recorded.

14. The camera (10) according to any one of the preceding claims, in, The control and evaluation unit (24) is designed to generate a grayscale image from the first recording channel and a color image from the second recording channel.

15. The camera (10) according to claim 1, in, The camera (10) is a code reader for reading an optical code (40).

16. The camera (10) according to claim 7, in, The control and evaluation unit (24) is designed to normalize the filter kernel using the value.

17. The camera (10) according to claim 12, in, Each row (20a-20b) is completely associated with the first recording channel or the second recording channel, respectively.

18. A method for detecting image data, for detecting first image data having a first pixel in a first recording channel of a first sensitivity of an image sensor (18) and for detecting second image data having a second pixel in a second recording channel of the image sensor (18) having a second sensitivity lower than the first sensitivity, wherein the first pixel and the second pixel are associated with each other by recording the same object area, wherein the influence of noise in the second image data is suppressed using a noise reduction filter, which assigns a new value to the respective observed second pixel based on second pixels in the neighborhood of the observed second pixel, It is characterized by: The denoising filter considers second pixels in the neighborhood with a weighting that depends on the degree of similarity between a first pixel associated with the second pixel and a first pixel (44) associated with the corresponding observed second pixel, wherein during processing using the denoising filter, second pixels associated with similar first pixels have a greater influence on the new value of the corresponding observed second pixel, while second pixels associated with dissimilar or less similar first pixels have a smaller influence on the new value of the corresponding observed second pixel.

Images