Device and method for automatic image improvement in a vehicle

Abstract

An apparatus for image improvement for a vehicle has a camera and an image processing module. The camera is arranged to detect an initial image, and the image processing module is arranged to determine a result image from the initial image. The image processing module has: a plurality of image processing filters, each of which is arranged to transform the initial image into an intermediate image respectively; an evaluation module, which outputs a quality metric number for each of the intermediate images transformed by means of the image processing filters; a selection module, which selects the intermediate image having the highest quality metric number and outputs the intermediate image having the highest quality metric number as the result image; a learning neural network, which is arranged to, in a learning phase, learn, for each initial image, the image processing filter having the highest quality metric number of the intermediate images among the plurality of image processing filters, and after the learning phase, select, for each initial image, the image processing filter having the highest quality metric number of the intermediate images among the plurality of image processing filters.

Description

Field of the Invention

[0001] The present invention relates to a device and a method for automatic image improvement in a vehicle, in particular a land vehicle. Automatic image improvement in a vehicle is used, for example, to support a series of assistance systems, or automatic image improvement is also used in vehicles with at least partially automated driving. Background Art

[0002] In many cases, vehicle applications of camera-based sensing devices rely on high-quality camera images. In some scenarios (such as when driving through a tunnel or when the weather suddenly changes), the quality of the camera image may suddenly change. In some camera-based systems (such as in a driving assistance system), this may lead to incorrect decisions. Summary of the Invention

[0003] The task of the present invention is to improve the quality of at least some camera images.

[0004] One aspect of the present invention relates to a method for automatic image improvement in a vehicle, the method having the following steps:

[0005] - Detecting an initial image by means of a camera;

[0006] - Transforming the initial image into a plurality of intermediate images by means of a plurality of image processing filters;

[0007] - For each of the intermediate images transformed by the image processing filter, determining a quality metric

[0008] - Selecting, by means of a selection module, the intermediate image having the highest quality metric, and outputting the intermediate image having the highest quality metric as a result image, wherein, in a learning phase, a learning neural network learns, for each initial image, the image processing filter having the highest quality metric of the intermediate images among the plurality of image processing filters, and after the learning phase, the learning neural network selects, for each initial image, the image processing filter having the highest quality metric of the intermediate images from among the plurality of image processing filters.

[0009] The vehicle on which the method is performed has a camera and an image processing module, the camera being used to detect the initial image, and the image processing module being arranged to determine the result image from the initial image. The method or the result image can be used, for example, to control an actuator of the vehicle.

[0010] The vehicle can be a land vehicle (in particular a passenger motor vehicle, a transport vehicle, a heavy goods vehicle, a land-based special vehicle), an amphibious vehicle, a ship or an aircraft. The camera can be a single camera, a stereo camera or multiple cameras (e.g., in the form of a "surround view camera"). The camera can be set up to detect a single image or a sequence of images. An image detected by the camera without further processing is called an "initial image", and sometimes also called a "raw image". An image that is output to the camera-based system after being processed by the image processing module is called a "result image". In many systems, the result image must meet certain quality requirements (e.g., in terms of brightness and / or contrast). Systems that use such result images as a basis include, for example, assistance systems, such as systems for lane recognition, systems for recognizing stationary objects (such as buildings or landmarks), or systems for recognizing moving objects (such as other vehicles or pedestrians). These result images can be used, for example, by at least partially automated driving vehicles.

[0011] The image processing module has a plurality of image processing filters, and each of these image processing filters transforms the initial image into an intermediate image in another step. The image processing filter can, for example, perform a color transformation or a contrast transformation (e.g., changing the gamma value), but can also perform much more complex transformations. The transformation can be supported by an algorithm for image recognition. Here, different settings of one filter are suitable for use as different image processing filters.

[0012] In another step, for each of the intermediate images transformed by the image processing filter, a quality metric number is output by means of an evaluation module. In some embodiments, the evaluation module can be configured very simply - for example, a simple recognition of the average brightness of the image and / or a metric of the contrast. The evaluation module can also be configured complexly. For example, the evaluation module can compare the results of different image recognition algorithms and thereby derive a quality metric number. The quality metric number is, for example, a scale (Skalar) by means of which the quality of the intermediate images can be quickly compared. In particular, by means of the quality metric number, it is possible to compare different types of images with different types of quality problems. Quality problems can include: low contrast, white saturation or color saturation ("saturated"), or distorted images. Each of the mentioned quality problems may require a dedicated filter for image improvement. An image may have more than one quality problem. In particular, if no filter results in an improvement of the initial image but rather in a deterioration, the quality metric number may also be negative.

[0013] In another step, the intermediate image with the highest quality metric is selected by means of a selection module, and the intermediate image with the highest quality metric is output as the result image. Under certain determined preconditions, for example in the following cases, the result image can be the initial image: no filter leads to an improvement of the initial image, or the initial image already has good quality.

[0014] The image processing module further includes a learning neural network which is configured to, during a learning phase, learn for each initial image the image processing filter having the highest quality metric for the intermediate image among a plurality of image processing filters, and which is configured to, after the learning phase, select from the plurality of image processing filters for each initial image the image processing filter having the highest quality metric for the intermediate image. Thus, the image processing module has different operating modes or operating phases: during the learning phase and after the learning phase. Here, during the learning phase, the output values of the neural network are not used, or the neural network does not output any values. The values include, for example, the type and parameters of the image processing filter. After the learning phase, the values output by the neural network are used to select the image processing filter. After the first learning phase, or after the neural network has established a certain "basic knowledge", automatic image improvement can be carried out by means of the image processing module. In one embodiment, the learning phase can be executed, for example, on a server or on a processor in a vehicle. In another embodiment, the neural network can be run on a processor in a vehicle.

[0015] In one embodiment, for an unknown initial image, the learning phase can be temporarily resumed, that is to say, even if the neural network has already been used, it can still "learn for this purpose". Thereby, the image processing module obtains an adaptive characteristic.

[0016] In one embodiment, the image types are classified by means of the learning neural network, for example in order to accelerate the output of the neural network.

[0017] In one embodiment, the learning neural network also uses a classification module which determines an illumination class for each initial image. In one embodiment, the classification module can determine other classes - for example distortion.

[0018] In one embodiment, at least one of the plurality of image processing filters is implemented to use a so-called bilateral grid as an image processing filter. The principle of the bilateral grid is described, for example, in "Real-Time Edge-Aware Image Processing Using the Bilateral Grid" by Chen, J. et al. (Massachusetts Institute of Technology, 2007). In the bilateral grid, the x and y values represent pixel positions and the z value is the intensity interval —— for example, the brightness in a black-and-white image. Using a bilateral grid also has the advantage of keeping the image edges smooth ("edge-aware brush"). Moreover, when using a bilateral grid, different filter settings are understood as different image processing filters.

[0019] In one embodiment, the color parameters of the initial image and the parameters of the bilateral grid (especially the parameters of the so-called "guidance map") are used separately. Here, in the bilateral grid, each of the elements is a color transformation matrix (CTM) in the form of x-y-z coordinates. Here, the x-y coordinates can have a reduced resolution compared to the initial image. In this case, the guidance map defines the correspondence between the pixels of the initial image and the CTM. The position of each pixel defines which CTM to use, and the value of the guidance map determines the associated z coordinate. In the case of gray values, the z coordinate can be, for example, an indication of the brightness value. In the present invention, the z coordinate is not necessarily a brightness value; in addition, the z coordinate can be learned. Thus, the parameters of the guidance map represent the transformation of the color initial image into the guidance map. This can be used to separate the color information and the edge information of the initial image. Thus, for example, the learning phase of a learning neural network can be configured more efficiently.

[0020] In one embodiment, the plurality of image processing filters include a filter for changing brightness, a filter for changing contrast, a filter for changing color, a distortion filter, a sharpness filter, and / or another filter. This another filter can be, for example, a high-pass filter or a low-pass filter. The distortion filter can be used, for example, to compensate for abnormal lines or camera artifacts.

[0021] In one embodiment, the device also has a classification module that is configured to determine an illumination class for each initial image, wherein the learning neural network is configured to, during the learning phase, learn the image processing filter with the highest quality metric for the intermediate image for each initial image and / or for each illumination classification, and after the learning phase, select the image processing filter with the highest quality metric for the intermediate image from the plurality of image processing filters for each initial image and / or for each illumination classification. The classification module can achieve an improvement and / or acceleration of image analysis through the neural network.

[0022] In one embodiment, the classification module is used for at least one of the following procedures: rough lighting estimation, description of weather data, description of the real environment, time information, or another information. This results in a matching of at least one parameter of the image processing filter, and this matching can be used, for example, for changing lighting conditions. Changing lighting conditions occur, for example, when a vehicle drives into a tunnel, or when dark clouds suddenly appear or it rains heavily.

[0023] Here, the rough lighting estimation, for example, takes into account the main lighting situation. The description of weather data is realized, for example, by a network (for example, by a predefined provider) or by other (for example, cooperative) vehicles. The description of the real environment can be realized by an on-board map, a GPS system, or an enhanced GPS system with current traffic data. The time information can include the time of day and the date, and the time information provides, for example, a first indicator for day or night. Other information can include, for example, astronomical data - such as sunset.

[0024] In one embodiment, the camera is set up to detect a sequence of initial images. This can be used, for example, to draw additional conclusions based on the past (i.e., previous initial images). Thus, for example, the conclusion "driving through a tunnel" can be drawn from a rapid change in lighting conditions (compared to the previous initial image); the conclusion "dusk" can be drawn, for example, in the case of a slow change in lighting conditions.

[0025] In one embodiment, when a predefined quality criterion is met, the resulting image is the same as the initial image. For example, this quality criterion may be met when the quality of the initial image is high enough (i.e., for example, when the brightness, contrast, or sharpness of the initial image is sufficient), and / or when the distortion of the initial image is very low. This quality criterion can also be met if no image processing filter can achieve a higher quality for any of the intermediate images.

[0026] One aspect of the present invention relates to a device for image improvement in a vehicle. The device has a camera and an image processing module. The camera is arranged to detect an initial image, and the image processing module is arranged to determine a result image from the initial image. Here, the image processing module has a plurality of image processing filters, and each of the plurality of image processing filters is arranged to transform the initial image into an intermediate image respectively. In addition, the device has an evaluation module and a selection module. The evaluation module outputs a quality metric number for each of the intermediate images transformed by means of the image processing filters. The selection module selects the intermediate image with the highest quality metric number and outputs the intermediate image with the highest quality metric number as the result image. The device also has a learning neural network, which is arranged to, during the learning phase, learn, for each initial image, the image processing filter having the highest quality metric number of the intermediate image among the plurality of image processing filters, and after the learning phase, select, for each initial image, the image processing filter having the highest quality metric number of the intermediate image from the plurality of image processing filters.

[0027] Another aspect of the present invention relates to an image processing module, which is arranged to determine a result image for a vehicle from an initial image. The image processing module has a plurality of image processing filters, and each of the plurality of image processing filters is arranged to transform the initial image into an intermediate image respectively. The image processing filters can, for example, perform color transformation or contrast transformation (such as changing the gamma value), but can also perform much more complex transformations. The image processing module also has an evaluation module, which outputs a quality metric number for each of the intermediate images transformed by means of the image processing filters. The quality metric number is, for example, a scale by which the quality of the intermediate images can be quickly compared. The image processing module also has a selection module, which is arranged to select the intermediate image with the highest quality metric number and output the intermediate image with the highest quality metric number as the result image.

[0028] In addition, the image processing module further includes a learning neural network, which is arranged to, during the learning phase, learn, for each initial image, the image processing filter having the highest quality metric number of the intermediate image among the plurality of image processing filters, and after the learning phase, select, for each initial image, the image processing filter having the highest quality metric number of the intermediate image from the plurality of image processing filters. During the learning phase, the output values of the neural network are not used, or the neural network does not output any values. These values include, for example, the type and parameters of the image processing filters. After the learning phase, the values output by the neural network are used to select the image processing filters.

[0029] Another aspect of the present invention relates to the application of the above-described device or method for automatic image improvement in a vehicle.

[0030] Another aspect of the present invention relates to a program unit which is arranged to execute the mentioned method when implemented on a processor unit. Here, the processor unit can have dedicated hardware for graphics acceleration, a so-called graphics card, and / or dedicated hardware with the function of a neural network - for example, an NNP unit (NNP: Neural Network Processing).

[0031] Another aspect of the present invention relates to a computer-readable medium on which the mentioned program unit is stored.

[0032] In the following, other measures for improving the present invention are shown in more detail with reference to the description of the preferred embodiments of the present invention in conjunction with the accompanying drawings. Description of the Drawings

[0033] The accompanying drawings schematically show:

[0034] Figure 1 An exemplary image sequence which has been detected by a camera according to one aspect of the present invention;

[0035] Figure 2 A vehicle having an embodiment of the above-described device;

[0036] Figure 3 An embodiment of an image processing module according to one aspect of the present invention;

[0037] Figure 4 Another embodiment of an image processing module according to another aspect of the present invention;

[0038] Figure 5 An example of an image processing filter according to another aspect of the present invention;

[0039] Figure 6 An example of a classification module according to another aspect of the present invention;

[0040] Figure 7 An embodiment for image improvement in a vehicle according to one aspect of the present invention. Detailed Description of the Invention

[0041] According to Figure 1, for example, a camera 120 installed on a vehicle detects an exemplary sequence of initial images 150 according to one aspect of the present invention. The sequence of the initial images 150 is shown on a timeline t. The portion of the initial image 150 marked by 155 has quality problems here, such that this portion of the initial image 150 cannot - or cannot without further processing - be used, for example, for an assistance system of the vehicle. These initial images 155 can be transmitted to a device 100 for image improvement (especially for automatic image improvement), such that after image improvement, at least a part of these images can be used as the resultant images 190 of the assistance system. Depending on the implementation, the remaining initial images 150 can be directly transmitted to the assistance system, or the remaining initial images can also be processed by means of a determined filter.

[0042] Figure 2 A vehicle 110 with an embodiment of the above-mentioned device 100 is schematically shown. The device 100 has, for example, a camera 120 arranged at the front of the vehicle 110. The device 100 can also have a plurality of cameras 120. The one or more cameras 120 can also be arranged at other positions of the vehicle 110 - for example, in the side mirrors. Each of the cameras 120 detects an initial image 150 or a sequence of initial images 150. The initial image 150 is transmitted to the input terminal 201 of the image processing module 200. After being processed in the image processing module 200, available resultant images 190 are present at the output terminal 202 of the image processing module 200, and these resultant images can be used, for example, by an assistance system (not shown here). The image processing module 200 can be a part of a processor unit (not shown here). Here, the processor unit can have dedicated hardware for graphics acceleration, a so-called graphics card, and / or dedicated hardware with neural network functions - for example, an NNP unit.

[0043] Figure 3FIG. 0 schematically shows an embodiment of an image processing module 200 according to the present invention. For example, an initial image 150 is transmitted from a camera 120 to an input 201 of the image processing module 200. The initial image 150 is transmitted from the input 201 to a filter bank 220 having a plurality of image processing filters 221, 222, 223, 224. These image processing filters may include, for example: a so-called bilateral grid, a filter for changing brightness, a filter for changing contrast, a filter for changing color, a distortion filter, a sharpness filter, and / or another filter. The another filter may be, for example, a high-pass filter or a low-pass filter. The distortion filter may be used, for example, to compensate for abnormal lines or camera artifacts. Some of these different filters may also be of the same filter type, but use different parameters respectively. Each of the plurality of image processing filters 221, 222, 223, 224 is arranged to transform the initial image 150 into intermediate images 231, 232, 233, 234 respectively. Each of the intermediate images 231, 232, 233, 234 is evaluated by an evaluation module 240, which outputs a quality metric number for each of the intermediate images 231, 232, 233, 234 transformed by the corresponding image processing filters 221, 222, 223, 224. As shown, the evaluation module 240 is composed of sub-modules 241, 242, 243, 244, for example, and available quality metric numbers exist at the outputs of these sub-modules, and these quality metric numbers are transmitted to a selection module 260. The quality metric number may be scaled and may be greater than zero ("improved"), equal to zero ("no quality improvement"), or less than zero ("worse"). One of these image processing filters (e.g., 224) may also be a "direct filter", that is, the initial image 150 is transmitted to the intermediate image 234 without change and is evaluated by means of a quality metric number (e.g., zero). The selection module 260 uses the intermediate image having the highest quality metric number and outputs the selected intermediate image through an output 202 of the image processing module 200. Then, this intermediate image is available as a result image 190 for a module connected later.

[0044] In addition, the initial image 150 is transmitted to the input 301 of the neural network 300. During the learning phase, the neural network 300 learns from the initial image 150, from the image processing filters 221, 222, 223, 224 used (via the interface 320), and from the selection module 260 (via the interface 360): which image processing filter is most suitable for which initial image. After the learning phase, the neural network 300 selects (via the interface 320) the best image processing filter for the initial image 150 and outputs, for example directly via the output 302 of the neural network 300, the resulting image 190 to the output 202 of the image processing module 200. Thus, after the learning phase, the full computational power of the image processing module 200 is only required when the initial image 150 is unknown to the neural network 300; that is, in most cases, only the selected image processing filter is switched on. Therefore, the learning phase is significantly more computationally intensive than the time after the learning phase. Thus, in some embodiments, the learning phase is carried out on a server and then the "learned" neural network 300 is transmitted to the processor unit of the vehicle.

[0045] Figure 4 Shows another embodiment of the image processing module 200 according to another aspect of the invention. Here, most of the components and functions are the same as Figure 3 However, this embodiment additionally has a classification module 400. Here, the initial image 150 is transmitted to the input 401 of the classification module 400. The classification of the initial image 150 takes place in the classification module 400. Here, an illumination class can be determined for each initial image 150. In another embodiment, the classification module 400 can also carry out other classifications - for example the determination of image distortion.

[0046] The classification module 400 uses at least one of the following programs: rough illumination estimation, description of weather data, description of the real environment, time information or another information. After processing, for example, the available illumination class is present at the output 402 of the classification module 400. This illumination class is provided (via the interface 340) to the neural network 300. Additionally or alternatively, the neural network 300 can use the illumination class in addition to the initial image 150. Thus, the classification module 400 can enable an improvement and / or acceleration of the image analysis by the neural network 300.

[0047] Figure 5 Shows an example of the image processing filter 221 according to one aspect of the invention; this image processing filter 221 uses a so-called bilateral grid. For this purpose - in Figure 5In the upper branch thereof - an image 160 with reduced information is used (for example, a black and white image 160). This image with reduced information is represented as a bilateral grid 167 by means of a transformation 165. Using a bilateral grid also has the advantage of keeping the image edges smooth. In Figure 5 In the lower branch thereof, the color parameters 158 of the initial image 150 are transmitted. In unit 159, the bilateral grid 167 is combined with the color parameters 158, thereby generating an intermediate image 231, which, in certain cases, has higher quality for subsequent programs. This separation of the color information and the edge information of the initial image 150 can be used to more efficiently configure the learning phase of a learning neural network.

[0048] Figure 6 An example of a classification module 400 according to another aspect of the present invention is shown. Here, the initial image 150 is transmitted from the camera 120 to the input 401 of the classification module 400. Subsequently, a rough illumination estimation 460 is performed by means of an estimation module 410. Then, this can - depending on the implementation - be passed from the combination module 470 to the output 402 as a classification result 490, and (via the interface 340) to the neural network 300 (see Figure 4 ). In some implementations, the combination module 470 uses, in addition to the rough illumination estimation 410, a description 420 of weather data, a description 430 of the real environment, time information 440, or another piece of information.

[0049] Figure 7A method 500 for image improvement in a vehicle 110 according to an aspect of the present invention is shown. In step 501, an initial image 150 is detected by means of a camera 120. In step 502, the initial image 150 is transformed into a plurality of intermediate images 231, 232, 233, 234 by means of a plurality of image processing filters 221, 222, 223, 224. In step 503, a quality metric number is determined for each of the intermediate images 231, 232, 233, 234. Finally, in step 504, the intermediate image with the highest quality metric number is selected by means of a selection module 260, and the intermediate image with the highest quality metric number is output as a result image 190. Steps 502 to 504 depend on whether the neural network 300 is operating during the learning phase or after the learning phase. During the learning phase, the neural network 300 learns from the initial image 150, from the image processing filters 221, 222, 223, 224 used (via an interface 320), and from the selection module 260 (via an interface 360): which image processing filter is most suitable for which initial image. After the learning phase, the neural network 300 (via the interface 320) selects the best image processing filter for the initial image 150 and outputs the result image 190, for example directly via an output 302 of the neural network 300 to an output 202 of an image processing module 200.

Claims

1. A method for automatic image improvement in a vehicle (110), the method having the following steps: Detecting an initial image (150) by means of a camera (120); Transforming the initial image (150) into a plurality of intermediate images (231, 232, 233, 234) by means of a plurality of image processing filters (221, 222, 223, 224); For each of the intermediate images (231, 232, 233, 234) transformed by means of the image processing filters (221, 222, 223, 224), determining a quality metric number (241, 242, 243, 244) by means of an evaluation module (240); Selecting, by means of a selection module (260), the intermediate image (231, 232, 233, 234) having the highest quality metric number (241, 242, 243, 244) and outputting the intermediate image having the highest quality metric number as a result image (190), Among them, In a learning phase, a learning neural network (300) learns, for each initial image (150), the following image processing filters (221, 222, 223, 224) among the plurality of image processing filters (221, 222, 223, 224): the image processing filter having the highest quality metric number (241, 242, 243, 244) of the intermediate image (231, 232, 233, 234), After the learning phase, the learning neural network (300) selects, for each initial image (150), the following image processing filters (221, 222, 223, 224) from the plurality of image processing filters (221, 222, 223, 224): the image processing filter having the highest quality metric number (241, 242, 243, 244) of the intermediate image (231, 232, 233, 234), wherein, during the learning phase, the output values of the learning neural network are not used, wherein, after the learning phase, the output values of the learning neural network are used to select the image processing filter, wherein the output values include the type and parameters of the image processing filter, wherein, for an unknown initial image, the learning phase is temporarily restored, otherwise only the selected image processing filter is switched on.

2. The method according to claim 1, wherein The learning neural network (300) also uses a classification module (400), the classification module being arranged to determine an illumination class (490) for each initial image (150).

3. The method according to claim 1 or 2, wherein At least one of the plurality of image processing filters (221, 222, 223, 224) is implemented to use a bilateral grid (165).

4. The method according to claim 3, wherein The color parameters (158) of the initial image are used separately from the parameters of the bilateral grid (165), wherein the separation of the color information and edge information of the initial image is used to configure the learning phase of the learning neural network.

5. The method according to claim 1 or 2, wherein The plurality of image processing filters (221, 222, 223, 224) includes: a filter for changing brightness, a filter for changing contrast, a filter for changing color, a distortion filter, a sharpness filter, and / or another filter, wherein the another filter is a high-pass filter or a low-pass filter.

6. The method according to claim 1 or 2, the method having further steps: Determining, by means of a classification module (400), an illumination class for each initial image (150), Among them, During the learning phase, the learning neural network (300) learns, for each initial image (150) and / or for each illumination class, the following image processing filters (221, 222, 223, 224) among the plurality of image processing filters (221, 222, 223, 224): the image processing filter having the highest quality metric number (241, 242, 243, 244) of the intermediate images (231, 232, 233, 234), After the learning phase, the learning neural network selects, for each initial image (150) and / or for each illumination class (490), from the plurality of image processing filters (221, 222, 223, 224) the following image processing filters (221, 222, 223, 224): the image processing filter having the highest quality metric number (241, 242, 243, 244) of the intermediate images (231, 232, 233, 234).

7. The method according to claim 6, wherein The classification module (400) uses at least one of an illumination rough estimate (410), a description of weather data (420), a description of the real environment (430), or time information (440).

8. The method according to claim 1 or 2, wherein The camera (120) is arranged to detect a sequence of initial images (150).

9. The method according to claim 1 or 2, wherein If a predefined quality criterion is met, the result image (190) is the same as the initial image (150).

10. An apparatus (100) for image improvement for a vehicle (110), the apparatus (100) having: A camera (120) arranged to detect an initial image (150), An image processing module (200) arranged to determine a result image (190) from the initial image (150), Among them, The image processing module (200) has: A plurality of image processing filters (221, 222, 223, 224), each of the plurality of image processing filters being arranged to transform the initial image (150) into an intermediate image (231, 232, 233, 234) respectively, An evaluation module (240) arranged to output a quality metric number (241, 242, 243, 244) for each of the intermediate images (231, 232, 233, 234) transformed by means of the image processing filters (221, 222, 223, 224), A selection module (260) configured to select an intermediate image (231, 232, 233, 234) having the highest quality metric numbers (241, 242, 243, 244) and output the intermediate image having the highest quality metric number as a result image (190). A learning neural network (300) configured to in a learning phase, learn, for each initial image (150), an image processing filter (221, 222, 223, 224) among the plurality of image processing filters (221, 222, 223, 224) having the highest quality metric number (241, 242, 243, 244) of the intermediate image (231, 232, 233, 234). and after the learning phase, for each initial image (150), select an image processing filter (221, 222, 223, 224) among the plurality of image processing filters (221, 222, 223, 224) having the highest quality metric number (241, 242, 243, 244) of the intermediate image (231, 232, 233, 234). wherein, during the learning phase, the output value of the learning neural network is not used, and wherein, after the learning phase, the output value of the learning neural network is used to select the image processing filter, and wherein the output value includes the type and parameters of the image processing filter. wherein, for an unknown initial image, the learning phase is temporarily restored, otherwise only the selected image processing filter is turned on.

11. An image processing module (200) configured to determine a result image (190) from an initial image (150). Among them, The image processing module (200) comprises:[[]] a plurality of image processing filters (221, 222, 223, 224), each of the plurality of image processing filters being configured to transform the initial image (150) into an intermediate image (231, 232, 233, 234) respectively. an evaluation module (240) configured to output a quality metric number (241, 242, 243, 244) for each of the intermediate images (231, 232, 233, 234) transformed by the image processing filters (221, 222, 223, 224). a selection module (260) configured to select an intermediate image (231, 232, 233, 234) having the highest quality metric number (241, 242, 243, 244) and output the intermediate image having the highest quality metric number as a result image (190). a learning neural network (300) configured to In the learning phase, for each initial image (150), the image processing filter (221, 222, 223, 224) among the plurality of image processing filters (221, 222, 223, 224) having the highest quality metric number (241, 242, 243, 244) of the intermediate image (231, 232, 233, 234) is learned, and after the learning phase, for each initial image (150), the image processing filter (221, 222, 223, 224) among the plurality of image processing filters (221, 222, 223, 224) having the highest quality metric number (241, 242, 243, 244) of the intermediate image (231, 232, 233, 234) is selected, wherein, during the learning phase, the output value of the learning neural network is not used, and after the learning phase, the output value of the learning neural network is used to select the image processing filter, and the output value includes the type and parameters of the image processing filter, wherein, for an unknown initial image, the learning phase is temporarily restored, otherwise only the selected image processing filter is turned on.

12. An apparatus configured to implement the method according to any one of claims 1 to 9.

13. A computer-readable medium having program units stored thereon, the program units being configured to execute the method according to any one of claims 1 to 9 when implemented on a processor unit.

Images