JP2023009613A - Imaging device and model learning system - Google Patents

Abstract

To provide an imaging device capable of obtaining a corrected captured image with a small degree of blurring.SOLUTION: An imaging device includes an imaging unit 10 for capturing an image in which blurring of an image has symmetry with respect to an optical axis 11a with a light receiving sensor 12 having a plurality of light receiving elements, and outputting a brightness signal indicating the brightness of each light receiving element, and an image processor 20 that receives, as input data, the brightness signal output by the imaging unit 10 and a polar coordinate component indicating the position of a light receiving element that has detected the brightness signal, inputting the input data to a learning network model to generate a corrected captured image in which blur is corrected, and outputting the generated corrected captured image to an arithmetic operation unit 2. The learning network model receives, as input data, a teacher brightness signal for indicating a teacher image having blurring improved for the captured image with a brightness signal for each image position corresponding to the position of the light receiving element, and the polar coordinate component indicating the image position, and is learned by using learning data including the input data and the teacher data.SELECTED DRAWING: Figure 1

Description

The present invention relates to an imaging device having an image processing function and a model learning system.

Non-Patent Document 1 discloses a technique for improving the blur of a captured image by learning in advance the luminance signal of each light receiving element representing the captured image as input data and inputting it to a learning network model. there is

”Learned large field-of-view imaging with thin-plate optics”, ACM Transactions on Graphics, Volume 38, Issue 6, November 2019 Article No.: 219, pp 1-14

The technology disclosed in Non-Patent Literature 1 is still insufficient in improving blurring. A technique capable of outputting an image with less blur is desired.

The present disclosure has been made based on this situation, and aims to provide an imaging device capable of obtaining a corrected captured image with a small degree of blurring, and model learning for learning a model used by the imaging device. It is to provide a system.

The above objects are achieved by the combination of features stated in the independent claims, and the sub-claims define further advantageous embodiments. The symbols in parentheses described in the claims indicate the corresponding relationship with the specific aspects described in the embodiments described later as one aspect, and do not limit the disclosed technical scope.

One disclosure related to an imaging device for achieving the above object is,
An imaging unit that captures a captured image whose blurring is symmetrical with respect to the optical axis (11a) with a light receiving sensor (12) having a plurality of light receiving elements and outputs a luminance signal indicating the luminance of each light receiving element. (10) and
The luminance signal output by the imaging unit and the polar coordinate component indicating the position of the light-receiving element that detected the luminance signal are used as input data, and the input data is input to a learning network model to generate a corrected captured image in which blur is corrected. , an image processing unit (20, 220) that outputs the generated corrected captured image to a predetermined arithmetic device,
The learning-type network model uses a teacher luminance signal that indicates a teacher image in which blurring has been improved with respect to the captured image by a luminance signal for each image position corresponding to the position of the light receiving element, and a polar coordinate component that indicates the image position as teacher data. , and is learned by learning data including input data and teacher data.

An image captured by the imaging unit has symmetrical blur of the image with respect to the optical axis. The image processing unit corrects blurring of the captured image using a learning network model. The learning network model is a model that takes as input data not only luminance signals but also polar coordinate components indicating the positions of the light receiving elements that detected the luminance signals.

This learning-type network model is learned based on a teacher image in which the blur is improved with respect to the captured image. Specifically, in the teacher image, the teacher data is a luminance signal for each image position corresponding to the light-receiving element (that is, teacher luminance signal) and a polar coordinate component indicating the image position. The learning network model is learned using this teacher data and input data as learning data.

If the learning network model is trained using this learning data, the correction accuracy is improved over the model trained using only the luminance signal as input data. The reason why the correction accuracy is improved is as follows. A picked-up image whose blur is symmetrical with respect to the optical axis has the same blur at mutually symmetrical coordinates when the coordinates are in a polar coordinate system. Therefore, by using the learning data described above, the parameters of the learning network model are learned using the symmetry of the blur during learning, so that the learning efficiency is improved. Improving the learning efficiency means that it becomes a learning network model that can improve blurring. Since the picked-up image is corrected using the learned network model learned in this way, the corrected picked-up image with less degree of blur can be obtained.

Another disclosure for achieving the above object is a model learning system that learns a learning network model provided in an imaging device. This model learning system
An imaging unit that captures a captured image whose blurring is symmetrical with respect to the optical axis (11a) with a light receiving sensor (12) having a plurality of light receiving elements and outputs a luminance signal indicating the luminance of each light receiving element. (10) and
Teacher data for generating teacher data including a teacher luminance signal indicating a teacher image in which blurring is improved with respect to a captured image by luminance signals for each image position corresponding to a plurality of light receiving elements, and a polar coordinate component indicating the image position. a generator (123, 323);
A learning network is generated by learning data including the input data and the teacher data generated by the teacher data generator using the luminance signal output by the imaging unit and the polar coordinate component indicating the position of the light-receiving element that detected the luminance signal as input data. and a model learning unit (124, 324) for learning model parameters.

1 is a diagram showing the configuration of an imaging device 1 according to a first embodiment; FIG. FIG. 4 is a diagram for explaining processing executed by an input data generation unit 22 and an image correction unit 23; 1 is a configuration diagram of a model learning system 100; FIG. The figure explaining the effect of 1st Embodiment. The figure which shows the structure of the imaging device 200 of 2nd Embodiment. FIG. 4 is a diagram for explaining processing executed by a coordinate conversion unit 221, an input data generation unit 222, and an image correction unit 223; 1 is a configuration diagram of a model learning system 300; FIG. The figure explaining the effect of 2nd Embodiment.

<First embodiment>
Hereinafter, embodiments will be described based on the drawings. FIG. 1 is a diagram showing the configuration of an imaging device 1 according to the first embodiment. The imaging device 1 of this embodiment is mounted on a vehicle and captures an image around the vehicle.

The imaging device 1 includes an imaging section 10 and an image processing section 20 . The imaging unit 10 captures an image of the surroundings of the vehicle and outputs digital data representing the captured image (hereinafter referred to as captured image) to the image processing unit 20 . The imaging unit 10 includes a lens 11 , a light receiving sensor 12 and an imaging control unit 13 .

The lens 11 collects light incident on the lens 11 from the outside of the vehicle onto the light receiving surface 12 a of the light receiving sensor 12 . The lens 11 is an aspherical lens having a concave surface on the first surface and a convex surface on the second surface, and is designed to provide excellent contrast in the central portion. Further, the lens 11 is preferably made of glass in consideration of weather resistance when used in a vehicle.

The light-receiving sensor 12 has a configuration in which a large number of light-receiving elements are arranged vertically and horizontally on a light-receiving surface 12a. A light-receiving element may also be called a light-receiving pixel or simply a pixel. A light receiving element is an element that converts the intensity of light into an electrical signal. A CMOS sensor, a CCD sensor, or the like can be used as the light receiving sensor 12 . The light receiving sensor 12 detects the intensity of red light, the intensity of green light, and the intensity of blue light by arranging red, green, and blue color filters in front of the optical path.

The light-receiving surface 12a has a curved shape that approaches the lens 11 side as the distance from the optical axis 11a of the lens 11 increases, in order to correct astigmatism. The degree of curvature of the light receiving surface 12a is appropriately determined according to the characteristics of the lens 11 so as to correct astigmatism.

As described above, the lens 11 has excellent central contrast. In other words, the image captured using this lens 11 has a greater degree of blurring from the center of the field of view toward the periphery of the field of view. In addition, since the lens 11 has a symmetrical shape with respect to the optical axis 11a, the captured image is an image in which blurring is symmetrical with respect to the optical axis 11a.

The imaging control unit 13 controls the light receiving element included in the light receiving sensor 12 . The imaging control unit 13 can be realized by a configuration including at least one processor. The imaging control unit 13 causes the light receiving sensor 12 to sequentially capture captured images, and sequentially outputs image signals representing the captured images to the image processing unit 20 . The image signal is, in detail, a luminance signal for each light receiving element. Since the light-receiving elements are arranged in a grid on the light-receiving surface 12a of the light-receiving sensor 12, the imaging control unit 13 sequentially transmits the luminance signal of each light-receiving element row by row in a continuous arrangement in the orthogonal coordinate system. Output to the image processing unit 20 . The image processing unit 20 can identify which light-receiving element detected each luminance signal based on the order of the luminance signals sequentially input to the image processing unit 20 . Note that the imaging control unit 13 may assign orthogonal coordinates to the luminance signal of each light-receiving element and output it to the image processing unit 20 .

The image processing unit 20 generates a corrected captured image in which blurring of the captured image is improved, and outputs image data representing the corrected captured image to the arithmetic device 2 . The computing device 2 may be any device that computes image data. The computing device 2 is, for example, a device that processes image data and detects an obstacle appearing in the corrected captured image. Another specific example of the computing device 2 is a display that displays the captured image after correction.

The image processing unit 20 can be realized by a configuration including at least one processor. Note that the imaging control unit 13 and the image processing unit 20 may be realized by the same processor. The image processing unit 20 includes an input data generation unit 22 and an image correction unit 23 as functions executed by the processor.

The input data generator 22 generates input data to be input to the image corrector 23 . One unit of input data is the red luminance value LR, the green luminance value LG, the blue luminance value LB, and the polar coordinates (r, θ) of the pixel position for one pixel position. The luminance values LR, LG, and LB can be values determined based on the luminance signal.

The distance r in the polar coordinates of the pixel position means the distance from the optical axis 11a. The angle θ in polar coordinates is the angle from the starting line. The input data generator 22 outputs the generated input data to the image corrector 23 .

The image correction unit 23 inputs the input data to the learning network model to generate a corrected captured image. Then, the image correction unit 23 outputs the generated captured image after correction to the arithmetic device 2 . The image processing section 20 includes a storage section 24 . The storage unit 24 is non-volatile, and stores a learned learning network model used by the image correction unit 23 .

A learning network model is a model constructed to perform a deconvolution operation using a learning filter. For example, U-net, which is also disclosed in Non-Patent Document 1, can be used. In Non-Patent Document 1, the U-net has six stages. Also in this embodiment, the number of stages can be six. However, the number of stages other than 6, such as 5, may be used. Since one set of input data has five parameters, the learning network model used in this embodiment has five input channels. Also, the conversion at the first stage is assumed to be 128 channels, for example.

FIG. 2 is a diagram for explaining the processing executed by the input data generation unit 22 and the image correction unit 23. As shown in FIG. In FIG. 2 , “Image_RGB” is an image signal output by the imaging control section 13 . The image signal is a signal group in which three luminance values L of red, green, and blue are arranged in an orthogonal coordinate arrangement for each pixel position. “Adress_rθ” is a polar coordinate position determined by the distance r and the angle θ. The input data generator 22 combines these five parameters to form one set of input data "Input_RGBrθ".

The image correction unit 23 sequentially cuts out the input data in a predetermined size. FIG. 2 illustrates a size of 256×256. The cut out input data is input to a learning network model to obtain output data. “Output_RGBrθ” means output data obtained by sequentially inputting input data to a learning network model. The image correction unit 23 obtains "Image_RGB" by removing the data of the polar coordinates (r, θ) from this output data. "Image_RGB" on the image correction unit 23 side is image data representing the captured image after correction.

[Learning processing]
Next, learning processing of the learning network model will be described. FIG. 3 shows the configuration of a model learning system 100 that performs learning processing. A model learning system 100 includes an imaging unit 10 , a teacher image imaging device 110 and a model generation device 120 .

The teacher image capturing device 110 is a device that captures a teacher image. The teacher image is an image in which blur is improved with respect to the captured image captured by the imaging unit 10 . The teacher image capturing device 110 combines a plurality of lenses in order to capture such teacher images. The teacher image capturing device 110 assigns rectangular coordinates indicating the image position of each teacher luminance signal to the teacher luminance signal representing the teacher image by the luminance signal for each image position corresponding to the position of the light receiving element, and creates a model generation device. output to 120.

The imaging unit 10 is the same as that included in the imaging device 1 . During learning, the imaging unit 10 images the same range as that captured by the teacher imaging device 110 . Note that the teacher image may be created by computer graphics. When the teacher image is an image created by computer graphics, the imaging unit 10 captures the teacher image.

The model generation device 120 includes an input data generation section 122 , a teacher data generation section 123 , a model learning section 124 and a storage section 125 . The input data generator 122 generates input data by the same process as the input data generator 22 included in the imaging device 1 .

The teacher data generation unit 123 acquires the teacher luminance signal from the teacher image capturing device 110 and generates teacher data based on this teacher luminance signal. The teacher data is data obtained by replacing the captured image in the input data with a teacher image. Therefore, the teacher data is data obtained by adding the polar coordinates (r, θ) indicating each image position to the luminance signal for each image position.

The model learning unit 124 learns the parameters of the learning network model using the input data generated by the input data generation unit 122 and the teacher data generated by the teacher data generation unit 123 as learning data. Then, the model learning unit 124 stores the learning network model in which the parameters have been learned in the storage unit 125 . The learning network model stored in the storage unit 24 included in the imaging device 1 is a learning network model learned by this model learning system 100 . In order to store the learned learning network model in the storage unit 24, the imaging device 1 and the model learning system 100 may communicate with each other. Alternatively, the operator may copy the learning network model from the storage unit 125 to the storage unit 24 using a portable storage medium.

In this way, the learning network model is learned using the teacher luminance signal and the image position indicated by polar coordinates as teacher data. When the learning network model is learned in this manner, the correction accuracy is improved over the model trained using only the luminance signal as input data for the following reasons.

In captured images, the blur has symmetry with respect to the optical axis, so by learning the polar coordinates as part of the input data, the symmetry of the blur is used to learn the parameters of the learning network model. It is from. By correcting the captured image using the learned network model learned in this way, the imaging apparatus 1 can obtain the corrected captured image with a small degree of blurring.

[Image improvement effect]
FIG. 4 shows the result of confirming the image improvement effect of the imaging device 1. As shown in FIG. In FIG. 4, the horizontal axis is HalfFoV (°). HalfFoV means the field of view on only one side of the optical axis with the optical axis at 0°. The Field of View is twice the HalfFoV. The vertical axis of FIG. 4 is PSNR (Peak Signal Noise Ratio).

The result shown as Non-Patent Document 1 in FIG. 4 is the result of using the device disclosed in Non-Patent Document 1 for only the learning network model in the configuration of the imaging device 1 of the first embodiment. . The model disclosed in Non-Patent Document 1 uses U-net, but the input data is only the luminance signal of each pixel, and polar coordinate information is not added to the luminance signal of each pixel.

As shown in FIG. 4, according to the first embodiment, the overall PSNR is improved compared to using the input data and learning network model disclosed in Non-Patent Document 1.

Also, in Non-Patent Document 1, a single lens using a diffractive DOE is adopted. Non-Patent Document 1 discloses a graph showing that when HalfFoV is around 9° to 13°, the PSNR is better than when an aspherical lens is employed. However, this graph, in other words, means that when HalfFoV is around 0° to 9°, PSNR is better when aspherical lenses are used. Therefore, in this embodiment, an aspherical lens 11 is employed. As a result, the resolution in the central portion of the field of view is improved as compared with the configuration disclosed in Non-Patent Document 1.

However, when the aspherical lens 11 is employed, the PSNR becomes low in the peripheral vision due to astigmatism. Therefore, in order to correct the astigmatism, the imaging device 1 includes the light receiving sensor 12 having a curved light receiving surface 12a. This configuration also improves the resolution in the periphery of the field of view.

<Second embodiment>
Next, a second embodiment will be described. In the following description of the second embodiment, the elements having the same reference numerals as those used so far are the same as the elements having the same reference numerals in the previous embodiments unless otherwise specified. Moreover, when only part of the configuration is described, the previously described embodiments can be applied to the other portions of the configuration.

FIG. 5 shows the configuration of an imaging device 200 according to the second embodiment. The imaging device 200 includes an imaging unit 10 that is the same as that included in the imaging device 1 . The imaging device 200 includes the imaging section 10 and an image processing section 220 .

The image processing section 220 includes a coordinate conversion section 221 , an input data generation section 222 , an image correction section 223 and a storage section 224 . The coordinate conversion unit 221 converts the luminance signal output from the imaging unit 10 into a polar coordinate array.

The input data generator 222 generates input data to be input to the image corrector 223 . One unit of input data in the second embodiment is a red luminance value LR, a green luminance value LG, and a blue luminance value LB for one pixel position, and the distance r of the polar coordinate component of that pixel position. . Unlike the first embodiment, the input data does not include the angle θ of the polar coordinate component of the pixel position. Therefore, the input data is 4-channel data.

The picked-up image is an image in which the degree of blurring is symmetrical with respect to the optical axis 11a. Therefore, even if the angle θ changes, the degree of blur remains the same. Therefore, in the second embodiment in which the array of luminance signals is converted to the polar coordinate system, the angle θ among the polar coordinate components of the pixel position is not included in the input data.

The image correction unit 223 inputs the input data generated by the input data generation unit 222 to the learning network model to generate a corrected captured image. Then, the image correction unit 223 outputs the generated captured image after correction to the arithmetic device 2 . The image processing section 220 includes a storage section 224 . The storage unit 224 stores a learned learning network model used by the image correction unit 223 .

FIG. 6 is a diagram illustrating processing executed by the coordinate transformation unit 221, the input data generation unit 222, and the image correction unit 223. As shown in FIG. In FIG. 6, "Image_RGB" of the coordinate conversion unit 221 is an image signal output by the imaging control unit 13, and "Polar_RGB" is an image signal arranged in the order of polar coordinates. "Adress_r" is the distance r, which is the polar coordinate component of the pixel position, and "Input_RGBr" is the input data.

The image correction unit 223 sequentially cuts out the input data in a predetermined size, inputs the cut out input data to the learning network model, and obtains output data. "Output_RGBr" means output data obtained by sequentially inputting input data to a learning network model. The image correction unit 223 removes the polar coordinate (r) data from this output data to obtain "Polar_RGB". Since this "Polar_RGB" is the corrected captured image in the array of polar coordinates, "Image_RGB" is obtained by converting "Polar_RGB" into the array of orthogonal coordinates. "Image_RGB" on the side of the image correction unit 223 is image data representing the captured image after correction in the orthogonal coordinate system.

[Learning processing]
Next, learning processing for learning a learning network model in the second embodiment will be described. FIG. 7 shows the configuration of a model learning system 300 that performs learning processing. A model learning system 300 includes an imaging unit 10 , a teacher image imaging device 110 and a model generation device 320 . The imaging unit 10 and the teacher image capturing device 110 are the same as those included in the model learning system 100 .

The model generation device 320 includes a coordinate transformation unit 321 , an input data generation unit 322 , a teacher data generation unit 323 , a model learning unit 324 and a storage unit 325 . The coordinate transformation unit 321 and the input data generation unit 322 execute the same processes as the coordinate transformation unit 221 and the input data generation unit 222 provided in the imaging device 200, respectively.

The teacher data generator 323 acquires the teacher luminance signal from the teacher image pickup device 110 . The teacher luminance signals obtained from the teacher image pickup device 110 are arranged in an orthogonal coordinate system. The teacher data generation unit 323 converts the teacher luminance signal acquired from the teacher image capturing device 110 into a polar coordinate array. Further, the teacher data generating unit 323 adds the distance r of the polar coordinates indicating the image position to the teacher luminance signal for each image position after conversion, and generates teacher data.

The model learning unit 324 learns the parameters of the learning network model using the input data generated by the input data generation unit 322 and the teacher data generated by the teacher data generation unit 323 as learning data. Then, the model learning unit 324 stores the learning network model, which has learned the parameters, in the storage unit 325 . The learning network model stored in the storage unit 325 of the imaging device 200 is a learning network model learned by this model learning system 300 .

Even if the learning-type network model is learned in this way, the correction accuracy is improved as compared with the model learned using only the luminance signal as input data.

[Image improvement effect]
FIG. 8 shows the result of confirming the image improvement effect of the imaging device 200. As shown in FIG. In FIG. 8, as in FIG. 4, the horizontal axis is HalfFoV and the vertical axis is PSNR. As shown in FIG. 8, even in the case of the second embodiment, the overall PSNR is improved compared to using the input data and the learning network model disclosed in Non-Patent Document 1. I understand.

Although the embodiments have been described above, the disclosed technology is not limited to the above-described embodiments, and the following modifications are also included in the disclosed scope. Various modifications can be made.

The imaging devices 1 and 200 do not have to be mounted on the vehicle. It may be mounted on a moving object other than a vehicle, or may be set in a place where it does not move.

Also, one system may be configured that includes the imaging devices 1 and 200 and the model learning systems 100 and 300 .

The imaging control unit 13, the image processing units 20 and 220, and the model generation devices 120 and 320 may be realized by dedicated hardware logic circuits, or may be implemented by a processor executing a computer program and one or more hardware logic circuits. It may also be implemented by one or more dedicated computers configured in combination. Hardware logic circuits are, for example, ASICs and FPGAs. The computer program may be stored in a computer-readable non-transitional tangible recording medium. For example, the computer program is stored in flash memory.

1: Imaging device 2: Arithmetic device 10: Imaging unit 11: Lens 11a: Optical axis 12: Light receiving sensor 12a: Light receiving surface 13: Imaging control unit 20: Image processing unit 22: Input data generation unit 23: Image correction unit 24: Storage Unit 100: Model Learning System 110: Teacher Image Imaging Device 120: Model Generating Device 122: Input Data Generating Unit 123: Teacher Data Generating Unit 124: Model Learning Unit 125: Storage Unit 200: Imaging Device 220: Image Processing Unit 221: Coordinate transformation unit 222: Input data generation unit 223: Image correction unit 224: Storage unit 300: Model learning system 320: Model generation device 321: Coordinate transformation unit 322: Input data generation unit 323: Teacher data generation unit 324: Model learning unit 325: storage unit r: distance θ: angle

Claims (7)

A photographed image having blurring symmetry with respect to an optical axis (11a) is imaged by a light receiving sensor (12) having a plurality of light receiving elements, and a luminance signal indicating the luminance of each of the light receiving elements is output. a part (10);
Correction in which the luminance signal output by the imaging unit and a polar coordinate component indicating the position of the light receiving element that detected the luminance signal are used as input data, and the input data is input to a learning network model to correct the blur. An image processing unit (20, 220) that generates a post-captured image and outputs the generated post-correction captured image to a predetermined arithmetic device,
The learning network model indicates a teacher image in which the blur is improved with respect to the captured image by the brightness signal for each image position corresponding to the position of the light receiving element, and the image position. and a polar coordinate component as teacher data, and learning with learning data including the input data and the teacher data. The imaging device according to claim 1,
The imaging unit outputs the luminance signal in an array of orthogonal coordinates,
The image processing unit
an input data generation unit (22) for generating the input data by adding the distance (r) from the optical axis and the angle (θ) from the starting line, which are the polar coordinate components, to each of the luminance signals;
an image correction unit (23) that inputs the input data to the learning network model to generate the corrected captured image;
The learning network model indicates the teacher image by the brightness signal for each image position corresponding to the position of the light receiving element, and the optical axis is attached to each of the teacher brightness signals arranged in an orthogonal coordinate system. and the data obtained by adding the distance from the starting line and the angle from the starting line are used as the teacher data, and learning is performed by learning data including the input data and the teacher data. The imaging device according to claim 1,
The imaging unit outputs the luminance signal in an array of orthogonal coordinates,
The image processing unit
a coordinate conversion unit (221) for converting the luminance signal output by the imaging unit into a polar coordinate array;
an input data generation unit (222) for generating the input data by adding the distance from the optical axis, which is the polar coordinate component, to each of the luminance signals converted by the coordinate conversion unit;
an image correction unit (223) that inputs the input data to the learning network model to generate the corrected captured image;
The learning network model indicates the teacher image by the brightness signal for each image position corresponding to the light receiving element, and stores the distance from the optical axis in each of the teacher brightness signals arranged in a polar coordinate system. is used as the teacher data, and learning is performed by learning data including the input data and the teacher data. The imaging device according to any one of claims 1 to 3,
The imaging device, wherein the imaging unit includes an aspherical lens (11) whose blur increases with increasing distance from the optical axis. The imaging device according to claim 4,
The image capturing unit includes a light receiving sensor (12) having a light receiving surface (12a) that has a curved shape that approaches the lens side as the distance from the optical axis increases. The imaging device according to claim 4 or 5,
The imaging device, wherein the lens is made of glass. A photographed image having blurring symmetry with respect to an optical axis (11a) is imaged by a light receiving sensor (12) having a plurality of light receiving elements, and a luminance signal indicating the luminance of each of the light receiving elements is output. a part (10);
teacher data including a teacher luminance signal indicating a teacher image in which blur is improved with respect to the captured image by the luminance signals for each image position corresponding to the plurality of light receiving elements, and a polar coordinate component indicating the image position; a training data generation unit (123, 323) to generate;
Input data includes the luminance signal output by the imaging unit and a polar coordinate component indicating the position of the light receiving element that detected the luminance signal, and the input data and the teacher data generated by the teacher data generation unit. A model learning system comprising a model learning unit (124, 324) for learning parameters of a learning type network model from learning data.

Images