Information processing method, information processing device and recording medium

By combining the characteristic information of noise and noise-free data in a single machine learning model, various problems of large amount of noise removal calculation and poor performance in the prior art are solved, and the stable noise removal effect on edge devices is achieved.

CN110889803BActive Publication Date: 2024-11-01PANASONIC INTELLECTUAL PROPERTY CORP OF AMERICA
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN201910830394.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2019-03-06
Filing Date
2019-09-04
Publication Date
2024-11-01
Estimated Expiration
2040-01-05

AI Technical Summary

Technical Problem

The prior art requires multiple neural network models when removing multiple types of noise, which leads to large computational volume and is difficult to practical on edge devices, and it is difficult to take into account the removal performance of multiple noises during training.

Method used

By using a single machine learning model, noise sensing data and its characteristic information are trained, and combined with the corresponding noise-free data and its characteristic information as reference data, the conversion processing of noise removal is realized, reducing the calculation amount and improving the removal performance.

Benefits of technology

A stable multiple noise removal on edge devices is achieved, reducing computational load, and a single model with both removal performance and processing speed is obtained.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN110889803B_ABST
    Figure CN110889803B_ABST
Patent Text Reader

Abstract

An information processing method, an information processing device, and a recording medium are provided. A single machine learning model is obtained by training to output sensor data containing multiple types of noises in which the multiple types of noises are reduced as a result of conversion of sensor data containing multiple types of noises. A computer obtains first sensor data containing noises; inputs the first sensor data to a single converter, obtains second sensor data output from the converter as a result of noise removal processing of the first sensor data by the converter; obtains third sensor data that does not contain noise and is different from the second sensor data in a scene that is the same as or corresponding to the first sensor data; obtains feature information of the second sensor data and feature information of the third sensor data; and trains the converter by machine learning using the second sensor data and its feature information as conversion data and the third sensor data and its feature information as reference data corresponding to the conversion data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present disclosure relates to an information processing method, an information processing device, and a recording medium recording a program. Background Art

[0002] As a technique that applies taught machine learning to remove noise from images of various scenes, a technique for removing noise based on image rendering using Monte Carlo path tracing is conventionally known (see Patent Document 1).

[0003] Prior art literature

[0004] Patent Literature

[0005] Patent Document 1: U.S. Patent Application Publication No. 2018 / 0293713 Summary of the invention

[0006] The noise canceller used in the above-mentioned noise removal technique includes a plurality of neural networks for coping with various scenes, that is, various noises, and therefore requires a large amount of calculations, and therefore the environment in which the noise canceller can be practically executed is limited.

[0007] The present disclosure provides an information processing method, etc., which can obtain a single machine learning model, which outputs sensor data containing multiple types of noise and is obtained by converting sensor data in which the multiple types of noise are reduced as a whole.

[0008] According to an information processing method involved in one embodiment of the present invention, a computer obtains first sensor data containing noise; inputs the first sensor data into a single converter, obtains second sensor data output from the converter as a result of the noise removal processing of the first sensor data by the converter; obtains third sensor data that does not contain noise and is different from the second sensor data in a scene that is the same as or corresponding to the first sensor data; obtains feature information of the second sensor data and feature information of the third sensor data based on the second sensor data and the third sensor data; and trains the converter through the following machine learning, which uses the second sensor data and the feature information of the second sensor data as conversion data, and uses the third sensor data and the feature information of the third sensor data as reference data corresponding to the conversion data.

[0009] In addition, an information processing device involved in one embodiment of the present invention comprises: a processor and a memory, the memory storing a single converter, the processor being configured to: obtain first sensor data containing noise generated by a camera device; input the first sensor data to the converter, obtain second sensor data output from the converter as a result of noise removal processing of the first sensor data by the converter; and output the obtained second sensor data, the converter being trained by the following machine learning, the machine learning using the second sensor data and feature information of the second sensor data based on the second sensor data as conversion data, and using third sensor data that does not contain the noise and is different from the second sensor data in a scene that is the same as or corresponding to the first sensor data, and feature information of the third sensor data based on the third sensor data, as reference data corresponding to the conversion data.

[0010] In addition, a recording medium involved in one embodiment of the present disclosure is a non-volatile recording medium that can be read by a computer having a processor and a memory, and records a computer program that is executed by the processor using the memory so that the computer performs the following processing: obtaining first sensor data containing noise generated by a camera device; inputting the first sensor data into a single converter stored in the memory, obtaining second sensor data output from the converter as a result of noise removal processing of the first sensor data by the converter; and outputting the obtained second sensor data, wherein the converter is trained by the following machine learning, which uses the second sensor data and feature information of the second sensor data based on the second sensor data as conversion data, and uses third sensor data that does not contain the noise and is different from the second sensor data in the scene that is the same as or corresponding to the first sensor data, and feature information of the third sensor data based on the third sensor data, as reference data corresponding to the conversion data.

[0011] Furthermore, these general or specific aspects may be implemented by a system, an integrated circuit, or a computer program, or may be implemented by any combination of an apparatus, a system, an integrated circuit, a method, a computer program, and a recording medium.

[0012] Effects of the Invention

[0013] According to the information processing method and the like involved in the present disclosure, it is possible to obtain a single machine learning model that, as a result of converting sensor data containing multiple types of noise, outputs sensor data in which the multiple types of noise are reduced as a whole. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1This is a block diagram showing an example of the functional configuration of an information processing device according to the embodiment.

[0015] Figure 2 This is a diagram showing an overview of input and output of a generator and a discriminator constituting a GAN (Generative Adversarial Network) constructed in the above-mentioned information processing device.

[0016] Figure 3 This is a diagram showing an overview of input and output of a feature extraction unit and a second conversion result comparison unit included in the above-mentioned information processing device.

[0017] Figure 4 This is a flowchart showing an example of steps of an information processing method executed by the above-mentioned information processing device in order to train a conversion processing unit by the above-mentioned information processing device.

[0018] Figure 5 A diagram showing an example of a functional configuration of another information processing device according to an embodiment and an overview of input and output between functional configuration elements.

[0019] Figure 6 This is a flowchart showing an example of the steps of an information processing method executed by the above-mentioned other information processing apparatus in order to remove noise from an image.

[0020] Description of labels:

[0021] 10.10A Information processing device (computer)

[0022] 20, 20A Noisy image group

[0023] 30 noise-free image groups

[0024] 40 converted images

[0025] 110 Image reading unit

[0026] 120 conversion processing unit (converter, generator)

[0027] 130 First conversion result comparison unit (identifier)

[0028] 140 Feature Extraction Unit

[0029] 150 Second conversion result comparison unit

[0030] 160 Conversion parameter update unit

[0031] 190 Conversion result output section. DETAILED DESCRIPTION

[0032] (Knowledge and insights that form the basis of this disclosure)

[0033] The inventors have discovered that the above-mentioned conventional techniques have the following problems.

[0034] An image containing different types of noise can also be considered as an image of different types of scenes in the above-mentioned conventional techniques. As an example of the types of noise mentioned here, Gaussian noise shown in an image photographed at a high sensitivity setting in a dark place can be cited. In addition, as other examples, pixel loss caused by sensor failure and whitening caused by overexposure can be cited. With respect to different types of noise, the characteristics of the image are different from each other, so the image processing required to remove various noises is also different. For example, for the removal of Gaussian noise, a certain smoothing process for pixel values ​​is effective. For whitening, there is a technology that can suppress the brightness of pixels in particularly bright parts of the image by using a tone curve to reduce the brightness of pixels. In addition, these are a very small part of the types of noise that can be contained in the image. In addition, blackening, vignetting, color shift caused by chromatic aberration, fixed pattern noise, random noise, aliasing, jitter, block noise, and mosquito noise shown in the image due to various reasons generated in the photographic environment, the optical system or circuit system of the photographic equipment, or in data processing such as compression can also be included in the objects referred to by the term noise in this disclosure. In addition, in the present disclosure, distortion aberration and reflection of foreign matter in the optical system such as attachments to the lens can also be included in the noise to be removed from the image. Furthermore, artifacts that are noises without special names as described above and complexes of these noises can also be included in the noise to be removed from the image in the present disclosure. In the conventional method, in order to remove such various noises using a neural network, a model of multiple neural networks trained for specific types of noise is used.

[0035] However, the noise removal process of the model using multiple neural networks requires a lot of calculations as a whole. Therefore, if the processing speed or power consumption is taken into consideration, the environment in which the process can be performed practically is limited. For example, it is difficult to practically install such a conventional noise canceller on portable information terminals such as digital cameras (including digital handheld cameras in the present disclosure, hereinafter also referred to as cameras) or smart phones, or edge devices with limited processing power of the processor or power that can be continuously used.

[0036] If a single neural network is used for noise removal, the amount of computation required for processing can be suppressed. However, when a single model corresponding to multiple types of noise with different characteristics is obtained, in the previous machine learning training method, the result may be a model specialized for the removal of a certain type of noise. That is, in a single model trained with the previous training method, the removal performance varies between the types of noise. Therefore, it is difficult to obtain a model that has the performance required for the removal of various noises. It is possible that the time required for training until a usable model is obtained becomes very long, or the training cannot be completed, that is, a model with the required performance cannot be obtained.

[0037] In order to achieve both suppression of computational load and stable noise removal performance for such a model obtained through training based on machine learning for the purpose of image noise removal, in an information processing method involved in one embodiment of the present disclosure, a computer obtains first sensor data containing noise; inputs the first sensor data into a single converter, and obtains second sensor data output from the converter as a result of the noise removal processing of the first sensor data by the converter; obtains third sensor data that does not contain noise and is different from the second sensor data in a scene that is the same as or corresponding to the first sensor data; obtains feature information of the second sensor data and feature information of the third sensor data based on the second sensor data and the third sensor data; and trains the converter through the following machine learning, which uses the second sensor data and the feature information of the second sensor data as conversion data, and uses the third sensor data and the feature information of the third sensor data as reference data corresponding to the conversion data.

[0038] By using this method, the model is trained to convert the input image into an image that better possesses the characteristics of an image without noise. As a result, a single model with stable noise removal performance can be obtained regardless of the type of noise contained in the input image.

[0039] In addition, the characteristic information of the second sensor data may be obtained by inputting the second sensor data into the following model, wherein the model is trained through machine learning to output the characteristic information of the input sensor data if the sensor data is input, and the characteristic information of the third sensor data may be obtained by inputting the third sensor data into the model.

[0040] In addition, the model may be a multi-layer neural network model, and the feature information of the second sensor data and the feature information of the third sensor data may be outputs of an intermediate layer of the model.

[0041] Furthermore, the intermediate layer may be a plurality of layers, and the feature information of the second sensing data and the feature information of the third sensing data may be outputs from the same layer among the plurality of layers.

[0042] In addition, the training may be an update of parameters of the converter, and the update of the parameters of the converter may be based on a difference between the second sensor data and the third sensor data, and a difference between feature information of the second sensor data and feature information of the third sensor data which are outputs from the same layer among the multiple layers.

[0043] In addition, the difference between the second sensor data and the third sensor data can also be output from the following identifier by inputting the second sensor data and the third sensor data into the following identifier, which is trained using machine learning to identify whether the input sensor data is the conversion data of the converter or whether it is reference data.

[0044] In addition, the sensing data may also be an image output from an image sensor.

[0045] In addition, an information processing device involved in one embodiment of the present invention comprises: a processor and a memory, the memory storing a single converter, the processor being configured to: obtain first sensor data containing noise generated by a camera device; input the first sensor data to the converter, obtain second sensor data output from the converter as a result of noise removal processing of the first sensor data by the converter; and output the obtained second sensor data, the converter being trained by the following machine learning, the machine learning using the second sensor data and feature information of the second sensor data based on the second sensor data as conversion data, and using third sensor data that does not contain the noise and is different from the second sensor data in a scene that is the same as or corresponding to the first sensor data, and feature information of the third sensor data based on the third sensor data, as reference data corresponding to the conversion data.

[0046] According to this information processing device, noise included in an image can be removed at a stable level regardless of its type.

[0047] In addition, a recording medium involved in one embodiment of the present invention is a non-volatile recording medium that can be read by a computer having a processor and a memory, and records a program that is executed by the processor using the memory so that the computer performs the following processing: obtaining first sensor data containing noise generated by a camera device; inputting the first sensor data into a single converter stored in the memory, obtaining second sensor data output from the converter as a result of the noise removal processing of the first sensor data by the converter; and outputting the obtained second sensor data, wherein the converter is trained by the following machine learning, which uses the second sensor data and feature information of the second sensor data based on the second sensor data as conversion data, and uses third sensor data that does not contain the noise and is different from the second sensor data in the scene that is the same as or corresponding to the first sensor data, and feature information of the third sensor data based on the third sensor data, as reference data corresponding to the conversion data.

[0048] By using an information processing device that executes this program, noise included in an image can be removed at a stable level regardless of its type.

[0049] Furthermore, these general or specific aspects may be implemented by a system, an integrated circuit, or a computer program, or may be implemented by any combination of an apparatus, a system, an integrated circuit, a method, a computer program, and a recording medium.

[0050] Hereinafter, an embodiment of an information processing method and an information processing device involved in one embodiment of the present disclosure will be described with reference to the accompanying drawings. The embodiment shown here represents a specific example of the present disclosure. Therefore, the numerical values, shapes, constituent elements, configurations and connection forms of constituent elements, and steps (processes) and the order of steps shown in the following embodiments are examples and do not limit the present disclosure. In addition, the constituent elements not recorded in the independent claims among the constituent elements in the following embodiments are constituent elements that can be added arbitrarily. In addition, each figure is a schematic diagram and does not necessarily have to be illustrated strictly.

[0051] (Implementation Method)

[0052] In the present disclosure, the implementation of the technology involved in noise removal is divided into a training stage and an inference stage for explanation. The training stage in the present disclosure is a stage for training a machine learning model that performs noise removal processing for an image containing noise. In addition, the inference stage in the present disclosure is a stage for inferring an image without noise from an image containing noise using the machine learning model obtained as a result of the training.

[0053] [1. Training phase]

[0054] [1-1. Composition]

[0055] Figure 1 1 is a block diagram showing an example of a functional configuration of an information processing device according to an embodiment. The information processing device 10 according to this embodiment is a device that performs training of a machine learning model for removing noise included in an image.

[0056] like Figure 1 As shown, the information processing device 10 includes an image reading unit 110 , a conversion processing unit 120 , a first conversion result comparing unit 130 , a feature extracting unit 140 , a second conversion result comparing unit 150 , and a conversion parameter updating unit 160 .

[0057] For example, the information processing device 10 may be implemented by a computer having a processor and a memory. In this case, the above-mentioned components of the information processing device 10 may be, for example, functional components implemented by the processor executing one or more programs stored in the memory. In addition, among the above-mentioned components, the conversion processing unit 120, the first conversion result comparison unit 130, the feature extraction unit 140 and the second conversion result comparison unit 150 are machine learning models, and among them, the conversion processing unit 120 is the above-mentioned training object for removing noise contained in the image. The training of each machine learning model will be described later.

[0058] The image reading unit 110 receives an input of an image containing noise (also referred to as a noisy image in the present disclosure). For example, a noisy image may be an image input for the information processing device 10 to remove noise, and a plurality of noisy images may be collected in a recording medium. The recording medium may also be one of the components of the information processing device 10. Alternatively, the recording medium may also be a detachable recording medium connected to an input / output interface (not shown) of the information processing device 10. Alternatively, the input / output interface may be a communication interface, and the recording medium may be a recording medium used in one or more other computers that communicate with the information processing device 10 in a wired or wireless manner. Figure 1 In FIG. 1 , the images collected in this way are represented as a noisy image group 20 .

[0059] The images included in the noisy image group 20 all contain at least one type of noise among the various noises exemplified above. The image containing noise received by the image reading unit 110 is also referred to as the first image below. For example, the first image is an image generated by a camera (image sensor) and contains noise due to some reason in the camera or the photographic environment. As another example of the first image, an image containing noise as a result of processing using a program can be cited. The processing of the image here is to process the image so as to change the pixel value or the number of pixels, such as adjusting the hue, brightness or contrast of the image, converting the shading (grayscale), adding, deleting or deforming the subject, expanding, reducing or changing the size, or changing the compression rate, or changing the format. In addition, the image can also be processed for the purpose of adding or pseudo-reproducing one of the above-mentioned noises to the image. The first image is input from the image reading unit 110 to the conversion processing unit 120.

[0060] The conversion processing unit 120 is an example of a converter stored in the memory of the information processing device 10 in this embodiment, and performs a conversion process for noise removal (also referred to as noise removal process in this disclosure) on the first image input from the image reading unit 110. The conversion processing unit 120 is a single machine learning model, such as a convolutional neural network model.

[0061] The conversion processing unit 120, which has performed noise removal processing on the first image, outputs a converted image (hereinafter also referred to as a second image) as a result of the processing. Among them, the second image in the training stage may still contain noise. For example, due to insufficient noise removal performance of the conversion processing unit 120, the noise contained in the first image may remain in the second image, or noise that is not in the first image may be added to the second image as a side effect of the noise removal processing. In particular, at the beginning of the training of the conversion processing unit 120, there is a high possibility that noise is contained in the second image. In addition, in the first image and the second image, except for the part damaged by noise, the scene (subject and composition) reflected is common.

[0062] In addition, in the training of the conversion processing unit 120, the data of the second image output as a result of the conversion (also referred to as conversion data in the present disclosure) is used as a part of training data.

[0063] The first conversion result comparison unit 130 is a machine learning model of a GAN (Generative Adversarial Network) constructed in the information processing device 10 by combining with the conversion processing unit 120. In this GAN, the conversion processing unit 120 as a converter is a generator (Generator), and the first conversion result comparison unit 130 is a discriminator (Discriminator). Figure 2 , an overview of input and output of the conversion processing unit 120 and the first conversion result comparison unit 130 is shown.

[0064] The first conversion result comparison unit 130 receives input of the second image output by the conversion processing unit 120 as described above and a third image which is an image not containing noise (also referred to as a noise-free image in the present disclosure).

[0065] The third image is an image without noise that shows the same or corresponding scene as the first image, and is input to the information processing device 10 as an image for training the conversion processing unit 120. For example, a plurality of third images may be collected in a recording medium. Figure 1 and Figure 2 In the above, the images collected in this way are represented as a noise-free image group 30. In addition, the recording medium collecting the third images may be the same as the recording medium collecting the first images, which is one of the components of the information processing device 10, or may be a detachable recording medium or any other recording medium used in other information processing devices.

[0066] In addition, when a first image is prepared by adding noise to an image that originally does not contain noise, the third image that shows the same scene as the first image is, for example, the image before the processing. If other representations are used, the first image and the third image that show the same scene are images that share common photographic conditions such as the subject and composition. In addition to these, the photographic conditions include, for example, the date and time of the photograph and the location. In addition, the equipment used and its condition and settings are also photographic conditions. In addition, conversely, an image that originally contains multiple types of noise and an image obtained by removing multiple types of noise using multiple image processing programs that can remove specific types of noise from the image can be used as the first image and the third image that show the same scene, respectively.

[0067] In addition, the third image showing the scene corresponding to the first image is an image having the same or similar subject and composition as the first image, and having a difference in at least a part of other photographic conditions and not containing noise. The following (Example 1) to (Example 3) are specific examples of combinations of images that can be used as the first image and the third image showing the corresponding scene. If among the multiple images exemplified in each of (Example 1) to (Example 3), there are images containing noise and images not containing noise as exemplified above, they can be used as the first image and the third image showing the corresponding scene, respectively.

[0068] (Example 1) A plurality of images having the same subject, location, and composition, and photographed continuously with varying ISO sensitivity.

[0069] (Example 2) Multiple images with the same subject, location, and composition, but shot under different lighting conditions. For example, images shot under different illumination or different light exposure methods (e.g., front light and back light). Images shot at different time periods (e.g., just after sunrise and at noon) or different weather conditions (e.g., sunny and cloudy).

[0070] (Example 3) A group of images taken simultaneously by a stereo camera. In this case, consider a case where only the image taken by one lens is reflected by the light source projected onto the glossy surface of the subject, causing whitening, or a case where the camera has a malfunction and only the image taken by the optical system or processing circuit affected by the malfunction contains noise.

[0071] In addition, of the two frames before and after included in a single shot in a motion image, one has noise removed through processing while the other does not, and the images aligned using motion information at the positions of different subjects between frames can also be used as the first image and the third image reflecting the corresponding scene.

[0072] Furthermore, when the subject and composition are common or similar and none of the multiple images with different other photographic conditions contain noise, one of the images may be processed to add noise and used as the first image, and the unprocessed image may be used as the third image.

[0073] Such data of the third image is used as a part of reference data in the training of the conversion processing unit 120 described later in the information processing device 10. In addition, in the present disclosure, the data of the third image is also referred to as correct answer data or label data, which means data corresponding to the training data of the data of the second image obtained by converting the first image showing the same or corresponding scene.

[0074] The first image (noise image), the second image (converted image), and the third image (noise-free image) prepared in this way are in a corresponding relationship in that they show the same or corresponding scenes, although they differ in the amount of noise or whether they are the results of conversion processing for the noisy image. Therefore, the conversion data (training data) as the data of the second image and the reference data (correct answer data or label data) as the data of the third image are also in a corresponding relationship.

[0075] Returning to the description of the first conversion result comparison unit 130, it is identified whether each input image is a converted image output as a result of the noise removal process performed by the conversion processing unit 120. More specifically, the first conversion result comparison unit 130 identifies the image from the noise-free image group 30 as a true class and the converted image output by the conversion processing unit 120 as a false class.

[0076] Then, the first conversion result comparison unit 130 calculates the difference between the converted image and the noise-free image corresponding to the converted image based on the result of the recognition. The less the noise of the second image, the smaller the difference (also referred to as the first error in the present disclosure) calculated by the first conversion result comparison unit 130. In other words, the more the performance of the noise removal processing of the conversion processing unit 120 is improved through training, the smaller the first error becomes. The calculated first error is input from the first conversion result comparison unit 130 to the conversion parameter updating unit 160.

[0077] In addition, the first conversion result comparison unit 130, which is the recognizer of the GAN, is also trained to improve the accuracy of recognition, which will be described later. In the information processing device 10 that constructs the above-mentioned GAN, the training for the conversion processing unit 120 is as follows: outputting a second image that is closer to the noise-free image by converting the first image, which makes it easy for the first conversion result comparison unit 130 to mistakenly recognize it as a noise-free image. In addition, the training for the first conversion result comparison unit 130 is as follows: it is used to more accurately recognize whether the second image output by the conversion processing unit 120 is a noise-free image (i.e., reference data) and whether the noise-free image is a noise-free image (i.e., reference data). The purpose of this GAN is to improve the performance of both the conversion processing unit 120 and the first conversion result comparison unit 130 in this way, while making them compete with each other, so as to obtain a converter with higher noise removal performance.

[0078] The feature extraction unit 140 is a trained machine learning model, and receives the second image output by the conversion processing unit 120 and the third image from the noise-free image group 30. The feature extraction unit 140 is a machine learning model trained for object recognition (identification) using a large number of images that do not contain noise, such as a convolutional neural network model.

[0079] In the feature extraction unit 140, the final output of the input image, that is, the result of object recognition, is not used, but the output from the intermediate layer is used. The output from the intermediate layer of the neural network used for object recognition indicates the possibility that the input image is an image without noise, that is, the feature quantity that expresses the characteristics or tendency of the image without noise. The features mentioned here can be, for example, contrast, edge or contour, but can also be features that are not easy to recognize by higher-order people. In the case of multiple intermediate layers, the output from a part of the multiple intermediate layers can also be used. In addition, the feature quantity output from each intermediate layer is information related to different features. The features shown by the feature quantity output from any intermediate layer indicate the possibility that the input image is an image without noise, so the feature quantities related to more types of features are used for training, so that it is possible to achieve the improvement of the accuracy of the noise removal process earlier or further improve the accuracy. The feature quantity (hereinafter also referred to as feature information) of each image obtained from the feature extraction unit 140 is input to the second conversion result comparison unit 150.

[0080] In addition, the data of such feature information of the second image is used as a part of conversion data (training data) in the training of the conversion processing unit 120 described later. In addition, the data of such feature information of the third image is used as a part of reference data (correct answer data or label data) in the training of the conversion processing unit 120 described later.

[0081] The second conversion result comparison unit 150, which receives the input of the feature information from the feature extraction unit 140, calculates the difference between the feature information of the second image and the feature information of the third image corresponding to the second image. As for the difference (also referred to as the second error in the present disclosure) calculated by the second conversion result comparison unit 150, the less the noise of the second image, the smaller the second error. In other words, the more the performance of the noise removal processing of the conversion processing unit is improved through training, the smaller the second error becomes. The calculated second error is input from the second conversion result comparison unit 150 to the conversion parameter updating unit 160.

[0082] In addition, when using feature information output from multiple intermediate layers, the second conversion result comparison unit 150 calculates the difference between the feature information of the second image output from the same layer and the feature information of the third image as the second error. For example, the difference between the feature quantity of the edge output from a certain intermediate layer of the feature extraction unit 140 that receives the second image as input and the feature quantity of the edge output from the same intermediate layer of the feature extraction unit 140 that receives the third image as input is calculated as the second error. Figure 3 In FIG. 1 , an overview of the input and output of the feature extraction unit 140 and the second conversion result comparison unit 150 is shown. Figure 3 , regarding the feature amount A of the second image associated with the feature A output from one of the intermediate layers of the feature extraction unit 140 as a neural network 2 and the feature value A of the third image 3 , the difference (second error related to feature A) is calculated in the second conversion result comparison unit 150 of the input destination. In addition, regarding the feature quantity B of the second image related to feature B output from another intermediate layer of the feature extraction unit 140 2 and the feature value B of the third image 3 , the difference (the second error related to feature B) is calculated in the second conversion result comparison unit 150 of the input destination.

[0083] The conversion parameter updating unit 160 updates the parameters of the conversion processing unit 120 based on the first error input from the first conversion result comparing unit 130 and the second error input from the second conversion result comparing unit 150. The parameter to be updated is, for example, as described above, a kernel value (weight) or a bias value when the conversion processing unit 120 is a convolutional neural network model.

[0084] The training of the conversion processing unit 120 is performed as follows: the difference between the data of the second image as the conversion data and the data of the third image as the reference data, that is, the first error indicating the difference, and the difference between the data of the feature amount of the second image as the conversion data and the data of the third image as the reference data, that is, the second error, are reduced by adjusting the parameters of the conversion processing unit 120. In other expressions, the first error is feedback information about the difference between the second image and the third image, and the second error is feedback information about the difference between the second image and the third image related to the feature of the noise-free image.

[0085] To adjust the parameters, for example, the conversion parameter updating unit 160 first adds the first error and the second error, or adds them after multiplying them by weight coefficients. Then, the conversion parameter updating unit 160 calculates a parameter that makes the sum of the addition result zero or a value very close to zero, and replaces the old parameter of the conversion processing unit 120 with the newly calculated parameter.

[0086] In addition, the conversion parameter updating unit 160 also performs machine learning training for the first conversion result comparing unit 130 to improve the accuracy of identifying whether each input image is a converted image output by the conversion processing unit 120 or noise-free data. The training is performed as follows: the second image is used as recognition data and the third image is used as reference data, and the third image is used as recognition data and the second image is used as reference data.

[0087] So far, the components of the information processing device 10 and the functions of each component have been described. In the information processing device 10, the components perform their respective functions, and the training based on machine learning of the conversion processing unit 120, which is also a generator of the GAN, which is a transformer that removes noise contained in the image. Next, the steps of the information processing method performed by the information processing device 10 for this training are described.

[0088] [1-2. Steps of information processing method for training transformer]

[0089] Figure 4 1 is a flowchart showing an example of the steps of the information processing method executed by the information processing device 10 having the above-mentioned configuration in order to train the conversion processing unit 120 based on machine learning.

[0090] First, in the information processing device 10, a noisy image (first image) is obtained from the noisy image group 20 (step S10), and is input to the image reading unit 110. In addition, there are a plurality of first images, each of which contains at least one type of noise among various types of noise, and the first image read by the image reading unit 110 as a whole contains a plurality of types of noise.

[0091] In the image reading unit 110, it is determined whether there is a first image that has not been read, that is, a first image that has not been used for training (step S20). If there is no first image that has not been used for training, that is, if all noisy images have been read for training (yes in step S20), the training of the conversion processing unit 120 in the information processing device 10 using the information processing method is completed. In the case of no in step S20, the unread first image is read and input to the conversion processing unit 120.

[0092] The conversion processing unit 120 converts the first image as a noise removal process to generate a converted image (second image) (step S30 ). The second image obtained as a result of the noise removal process is input to the first conversion result comparison unit 130 and the feature extraction unit 140 .

[0093] Furthermore, in the information processing device 10 , a noise-free image (third image) corresponding to the first image is obtained from the noise-free image group 30 , and is input to the first conversion result comparison unit 130 and the feature extraction unit 140 (step S40 ).

[0094] The first conversion result comparison unit 130 calculates the difference between the second image and the third image received as input, and obtains a first error (step S50 ). The first error is input to the conversion parameter update unit 160 .

[0095] In addition, the feature extraction unit 140 obtains the feature of the noise-free image, that is, the feature amount of the second image and the feature amount of the third image, based on each of the second image and the third image received as input. The feature amount of the second feature information and the feature amount of the third feature information obtained are input to the second conversion result comparison unit 150.

[0096] The second conversion result comparison unit 150 calculates the difference between the second feature information and the third feature information received as input, and obtains a second error (step S60 ). The second error is input to the conversion parameter update unit 160 .

[0097] The conversion parameter updating unit 160 calculates new parameters of the conversion processing unit 120 based on the first error and the second error received as input, and updates the parameters of the conversion processing unit 120 using the calculated parameters (step S70). The new parameters calculated in step S70 are parameters that result in smaller first errors and second errors than the conversion processing using the old parameters. Thus, machine learning is performed using the second image and the feature information of the second image as conversion data, and using the third image and the feature information of the third image as reference data corresponding to the conversion data.

[0098] In the above, the steps of the information processing method performed by the information processing device 10 for the training of the machine learning of the conversion processing unit 120 as a single model are described using examples. In addition, the information processing method is not limited to the above-mentioned step example. For example, the order of step S50 and step S60 in the above-mentioned step example can also be exchanged.

[0099] [1-3. Effect]

[0100] The difference between the second image and the third image obtained as a result of the conversion process performed by the conversion processing unit 120 using the above-mentioned new parameters is reduced compared to the difference between the second image and the third image obtained as a result of the conversion process using the old parameters. This effect is obtained as follows: the above-mentioned training of the converter that performs the conversion for noise removal processing of the image, that is, the conversion processing unit 120, is training for reducing the first error, that is, training for generating an image that is closer to an image that does not contain noise by converting an image containing noise.

[0101] In addition, regarding the characteristics of a noise-free image, the difference between the second image and the third image obtained as a result of the conversion process using the new parameters is reduced compared to the difference between the second image and the third image obtained as a result of the conversion process using the old parameters. This effect is achieved as follows: the above-mentioned training of the conversion processing unit 120 is further training for reducing the second error, that is, training to generate an image having characteristics closer to the characteristics of a noise-free image possessed by an image not containing noise by converting an image containing noise. If other expressions are performed, the converted image output by the conversion processing unit 120 after such training has less sense of disobedience as a noise-free image than the converted image output before the training, that is, it can be said to be closer to nature.

[0102] The conversion processing unit 120 obtained through such training is not specialized for a specific type of noise, and can appropriately remove multiple types of noise.

[0103] The conversion processing unit 120 that is the subject of the training is a single machine learning model. Therefore, in the noise removal process performed by the conversion processing unit 120, the amount of calculation is suppressed compared to the conventional noise removal process that uses multiple neural networks to cope with multiple types of noise removal.

[0104] Next, the inference stage from an image containing noise to an image containing no noise performed by the converter 120 obtained in this way will be described using an information processing device including the conversion processing unit 120 as an example.

[0105] [2. Reasoning (noise removal) stage]

[0106] [2-1. Structure]

[0107] Figure 5 The figure shows an example of functional configuration of another information processing device according to the present embodiment and an overview of input and output between functional configurations. The information processing device 10A according to the present embodiment is a device that receives an input of an image containing noise, converts the image, and obtains an image with the noise removed.

[0108] like Figure 5 As shown, the information processing device 10A includes an image reading unit 110, a conversion processing unit 120, and a conversion result output unit 190. Among these components of the information processing device 10A, components common to the components of the information processing device 10 are denoted by the same reference numerals. In addition, the common components may be described in a simplified manner below.

[0109] For example, the information processing device 10A may be implemented by a computer having a processor and a memory. In this case, the above-mentioned components of the information processing device 10A may be, for example, functional components implemented by the processor executing one or more programs stored in the memory. In addition, the information processing device 10A and the information processing device 10 may be separate computers or implemented by a single computer.

[0110] The image reading unit 110 receives input of an image containing noise, that is, a noisy image. For example, a noisy image may be an image input in order to remove noise from the information processing device 10A, and a plurality of noisy images may be collected in a recording medium. The recording medium may be one of the components of the information processing device 10A. Alternatively, it may be a detachable recording medium connected to an input / output interface (not shown) of the information processing device 10A. Alternatively, the input / output interface may be a communication interface, and may be a recording medium used in one or more other devices that communicate with the information processing device 10A in a wired or wireless manner. Figure 5 In FIG. 2 , the images collected in this way are shown as a noisy image group 20A.

[0111] Although the images included in the noisy image group 20A are different from the images included in the noisy image group 20, they both include various noises as exemplified above. The noisy image received as input by the image reading unit 110 of the information processing device 10A may be referred to as the first image in the following as described in the training stage. For example, the first image is an image generated by a camera device (image sensor) provided by a camera or the like and includes noise due to some reason in the camera or the photographic environment. The first image is input from the image reading unit 110 to the conversion processing unit 120.

[0112] The conversion processing unit 120 is an example of a converter stored in the memory of the information processing device 10A in this embodiment. The difference from the converter described in the training stage is that it is an inference model obtained as a result of machine learning training based on the above-mentioned information processing method. That is, the conversion processing unit 120 is a single machine learning model corresponding to multiple types of noise removal, such as a convolutional neural network model. In the case where the information processing device 10A is separate from the information processing device 10, the information processing device 10A is an information processing device in which the trained conversion processing unit 120 is installed in the information processing device 10.

[0113] The conversion processing unit 120 that has performed the noise removal process on the first image outputs a converted image as a result of the process. In the description of the inference stage, the converted image output by the conversion processing unit 120 is also referred to as a second image in the same manner as in the description of the training stage. The second image in the inference stage is obtained by performing a noise removal process that is better than the noise removal process that the conversion processing unit 120 can perform at the beginning of the training stage, and is a more natural image as a noise-free image.

[0114] The conversion result output unit 190 outputs the second image output by the conversion processing unit 120. The output destination of the second image is, for example, a recording medium in which the second images are collected. Figure 5, the images collected in this way are shown as the converted image group 40. The recording medium may be the same as the recording medium collecting the above-mentioned noisy image group 20, and may be one of the components of the information processing device 10A, or may be a detachable recording medium or any of the recording media used in other devices. For example, the information processing device 10A may be implemented by a system LSI and loaded into a photographic device such as a camera or a driving recorder, and the recording medium used in the photographic device is the output destination of the conversion result output unit 190. In this case, even images photographed under conditions that are prone to noise, such as high-sensitivity photography in a dark place, can be recorded in the recording medium with a clear image quality with less noise. In addition, as another example of the output destination of the second image, it can also be output to a display device as an image prompted to the user. For example, the information processing device 10A may be implemented by a system LSI and loaded into a vehicle-mounted rear monitor system or an electronic mirror system, and the second image is output to a monitor provided by the vehicle. In this case, even images photographed under conditions that are prone to noise can be displayed in the monitor with a clear image quality with less noise. In addition, since the amount of calculation is suppressed, it may be helpful to shorten the time delay from photography to display in the use of such an in-vehicle system. Another example of the output destination of the second image may be other hardware or software that performs some kind of image processing. For example, the second image may also be input into a program for object recognition processing, in which case the information processing device 10A helps to achieve object recognition with less misrecognition or missed recognition and high accuracy. In addition, since the amount of calculation is suppressed, it may be helpful to shorten the processing time from photography to completion of object recognition in the use of the in-vehicle system. In this way, the second image is output from the conversion result output unit 190 in a manner corresponding to its use.

[0115] Next, the steps of an information processing method for noise removal executed by the information processing device 10A having such a configuration will be described.

[0116] [2-2. Steps of information processing method for noise removal]

[0117] Figure 6 : is a flowchart showing an example of the procedure of the information processing method for noise removal executed by the information processing device 10A.

[0118] First, in the information processing device 10A, a noisy image (first image) is acquired (step S10A), and is input to the image reading unit 110 .

[0119] In the image reading unit 110, it is determined whether there is a first image that has not been read, that is, a first image that has not been subjected to noise removal processing (step S20A). If there is no first image that has not been subjected to noise removal processing, that is, if all first images have been subjected to noise removal processing (yes in step S20A), the noise removal processing of the first image using the information processing method in the information processing device 10A ends. If no in step S20A, the read first image is input to the conversion processing unit 120.

[0120] The conversion processing unit 120 converts the first image as noise removal processing to generate a converted image (second image) (step S30A).

[0121] The second image obtained as a result of the noise removal process is input from the conversion processing unit 120 to the conversion result output unit 190 , and is output from the conversion result output unit 190 in a format appropriate for its intended use.

[0122] The steps of the information processing method executed by the information processing device 10A equipped with the conversion processing unit 120 which is a model trained by machine learning have been described above using examples. The information processing method is not limited to the above-mentioned step examples.

[0123] In addition, the information processing method can be performed by the information processing device 10A, for example, by a processor included in the information processing device 10A executing a predetermined program using a memory.

[0124] [2-3. Effect]

[0125] The conversion processing unit 120 that performs the inference from the noisy image to the noise-free image in the above-mentioned inference stage is a machine learning model trained using the method described in the description of the above-mentioned training stage. Therefore, regardless of the type of noise contained in the input noisy image, any type of noise can be appropriately removed.

[0126] In addition, the conversion processing unit 120 is a single machine learning model. Therefore, in the noise removal process performed by the conversion processing unit 120, the amount of calculation is suppressed compared to the previous noise removal process corresponding to multiple types of noise removal using multiple neural networks. Therefore, for the conversion processing unit 120, the limitation of the practical execution environment is relaxed compared to the previous noise canceller. The conversion processing unit 120 can improve the usefulness of such edge devices by being installed in, for example, an edge device with limited processing power of a processor or sustainable power.

[0127] (Transformation and other supplementary matters)

[0128] The information processing method, information processing device and program involved in one or more embodiments of the present disclosure are not limited to the description of the above-mentioned embodiments. As long as it does not deviate from the purpose of the present disclosure, various modifications that a person skilled in the art may think of may be applied to the above-mentioned embodiments. The following modifications and other supplementary matters to the description of the embodiments may be cited.

[0129] (1) The types of noise included in the first image and eliminated by the noise removal process are not limited to those listed in the description of the embodiment. For example, fog or dust generated during photography may appear in the image regardless of the performance, settings, or failure of the camera. In this way, an image of an object that is inevitably present in the photographic range and is reflected in the image and may damage the usefulness of the image may be included in the first image of the present disclosure and may be an object to be removed by the noise removal process. As a result, the range of photographic conditions of images that can, for example, perform object recognition with high accuracy is expanded.

[0130] (2) The term "image containing no noise (or noise-free image)" used in the embodiments is not limited to an image containing no noise at all. For example, as illustrated in the description of the inference stage, whether or not it is a noise-free image may be determined based on the feature quantity of each image obtained as a noise-free image using the feature extraction unit 140, or it may be determined using various other existing image analyses. Alternatively, it may be determined based on other information obtained from the image data such as brightness distribution, image meta-information based on Exif (Exchangeable image file format), or a combination thereof. Alternatively, it may be determined based on human eye observation. In addition, the criterion for determining whether or not it is a noise-free image may also differ depending on the noise removal performance required of the trained converter or the purpose of the image after the noise removal process performed by the converter. In addition, the noise contained in the "image containing noise (or noisy image)" is not limited to noise that can be recognized by human eye observation. For example, it may be noise that is difficult for humans to recognize in screen displays or printed materials, but may cause a decrease in recognition accuracy in processing based on programs such as object recognition.

[0131] (3) The machine learning model of the information processing device in the embodiment is not limited to the machine learning model based on the convolutional neural network technique exemplified above. Each model used as the conversion processing unit 120, the first conversion result comparison unit 130, the feature extraction unit 140 or the second conversion result comparison unit 150 may be trained by other machine learning techniques as long as it can provide the desired function.

[0132] (4) In the above embodiment, an example in which the object of processing is an image is described, but the object of processing may also be sensor data other than an image. For example, sensor data that can obtain reference data, such as sound data output from a microphone, point group data output from a radar such as LiDAR, pressure data output from a pressure sensor, temperature data or humidity data output from a temperature sensor or a humidity sensor, and fragrance data output from a fragrance sensor, may also be set as the object of processing. In this sense, the data of the first image, the second image, and the third image in the above embodiment are examples of the first sensor data, the second sensor data, and the third sensor data, respectively, and the first sensor data, the second sensor data, and the third sensor data are sensor data in the same or corresponding scenes.

[0133] (5) Some or all of the components of each of the above-mentioned information processing devices may also be constituted by a system LSI (Large Scale Integration). System LSI is a super-multifunctional LSI manufactured by integrating multiple components into one chip. Specifically, it is a computer system composed of a microprocessor, ROM (Read-Only Memory), RAM (Random Access Memory), etc. Computer programs are stored in ROM. The microprocessor operates according to the computer program, and the system LSI achieves its functions.

[0134] In addition, although it is set as system LSI here, it is sometimes called IC, LSI, super LSI, and ultra LSI depending on the difference in integration. In addition, the method of integrated circuit is not limited to LSI, and it can also be realized by a dedicated circuit or a general-purpose processor. It is also possible to use an FPGA (Field Programmable Gate Array) that can be programmed after LSI manufacturing, or a reconfigurable processor that can reconfigure the connection or setting of the circuit unit inside the LSI.

[0135] Such a system LSI or the like can be incorporated into, for example, a portable information terminal with a camera such as a digital camera or a smartphone, or can be incorporated into various vehicle-mounted systems such as a monitoring system or an advanced driving assistance system including an object detection function.

[0136] Furthermore, if integrated circuit technology that replaces LSI emerges due to the progress of semiconductor technology or other derived technologies, the functional blocks can also be integrated using this technology. Application of biotechnology is also possible.

[0137] (6) In addition, the components of each of the above-mentioned information processing devices may also be implemented as components of an information processing system that provides the same functions as the above-mentioned information processing devices by, for example, coordinating and operating a plurality of computers that each have a processor and a memory and can communicate with each other. In this case, these components are implemented, for example, by executing one or more programs stored in a part or all of the memories of these computers by a part or all of the processors of these computers.

[0138] (7) One aspect of the present disclosure is not only the above-mentioned information processing devices, but also an information processing method using characteristic components included in the information processing device as steps. The information processing method, for example, uses Figure 4 or Figure 6 The information processing method as described above is a flowchart. In addition, one embodiment of the present disclosure may also be a computer program that causes a computer to execute each characteristic step included in such an information processing method. In addition, one embodiment of the present disclosure may also be a computer-readable non-volatile recording medium that records such a computer program.

[0139] Industrial Applicability

[0140] The present disclosure can be widely used in devices and the like that perform processing for removing noise from various sensor data such as image data.

Claims

1. An information processing method, executed by a computer: Acquire first sensor data containing noise; inputting the first sensor data to a single converter, and obtaining second sensor data output from the converter as a result of noise removal processing of the first sensor data by the converter; Acquire third sensor data that does not contain noise and is different from the second sensor data in a scene that is the same as or corresponds to the first sensor data; Based on the second sensor data and the third sensor data, obtaining feature information of the second sensor data and feature information of the third sensor data; calculating a difference between the second sensing data and the third sensing data; calculating a difference between feature information of the second sensing data as conversion data and feature information of the third sensing data as reference data corresponding to the conversion data; The converter is trained by machine learning to update the parameters of the converter, wherein the machine learning uses the calculated difference between the second sensor data and the third sensor data, and the calculated difference between the feature information of the second sensor data as the conversion data and the feature information of the third sensor data as the reference data corresponding to the conversion data.

2. The information processing method according to claim 1, The feature information of the second sensor data is obtained by inputting the second sensor data into a model, the model being trained through machine learning to output feature information of the input sensor data when the sensor data is input. The characteristic information of the third sensing data is acquired by inputting the third sensing data into the model.

3. The information processing method according to claim 2, The model is a multi-layer neural network model, The feature information of the second sensing data and the feature information of the third sensing data are outputs of the middle layer of the model.

4. The information processing method according to claim 3, The intermediate layer includes a plurality of layers, and the feature information of the second sensing data and the feature information of the third sensing data are outputs from the same layer among the plurality of layers.

5. The information processing method according to claim 1, The difference between the second sensed data and the third sensed data is output from the identifier by inputting the second sensed data and the third sensed data to the identifier trained using machine learning to identify whether the input sensed data is conversion data of the converter or reference data.

6. The information processing method according to any one of claims 1 to 5, The first sensing data, the second sensing data, and the third sensing data are images output from an image sensor.

7. An information processing device comprising a processor and a memory, The memory stores a single converter, The processor is composed of: Acquire first sensor data including noise generated by a camera device; inputting the first sensing data into the converter, and obtaining second sensing data output from the converter as a result of noise removal processing of the first sensing data by the converter; outputting the obtained second sensor data; Acquire third sensor data that does not contain the noise and is different from the second sensor data in a scene that is the same as or corresponds to the first sensor data; acquiring feature information of the second sensing data based on the second sensing data, and feature information of the third sensing data based on the third sensing data; calculating a difference between the second sensing data and the third sensing data; calculating a difference between feature information of the second sensing data as conversion data and feature information of the third sensing data as reference data corresponding to the conversion data, The converter is trained by the following machine learning to update the parameters of the converter, and the machine learning uses the calculated difference between the second sensor data and the third sensor data, and the calculated difference between the feature information of the second sensor data as the conversion data and the feature information of the third sensor data as the reference data corresponding to the conversion data.

8. A recording medium, which is a non-volatile recording medium that can be read by a computer having a processor and a memory, and records a computer program that is executed by the processor using the memory to cause the computer to execute the following processing: Acquire first sensor data including noise generated by a camera device; inputting the first sensing data to a single converter stored in the memory, and obtaining second sensing data output from the converter as a result of noise removal processing of the first sensing data by the converter; outputting the obtained second sensor data; Acquire third sensor data that does not contain the noise and is different from the second sensor data in a scene that is the same as or corresponds to the first sensor data; acquiring feature information of the second sensing data based on the second sensing data, and feature information of the third sensing data based on the third sensing data; calculating a difference between the second sensing data and the third sensing data; calculating a difference between feature information of the second sensing data as conversion data and feature information of the third sensing data as reference data corresponding to the conversion data, The converter is trained by the following machine learning to update the parameters of the converter, and the machine learning uses the calculated difference between the second sensor data and the third sensor data, and the calculated difference between the feature information of the second sensor data as the conversion data and the feature information of the third sensor data as the reference data corresponding to the conversion data.

Citation Information

Patent Citations

  • Denoising monte carlo renderings using machine learning with importance sampling

    US20180293713A1

  • Information processing apparatus, information processing method, learning apparatus and learning method

    CN101005569A

  • Neural network training and image processing methods and apparatuses, electronic device and storage medium

    CN108229525A

  • Regional aware image de-noising method based on machine learning

    CN108416756A

  • A method for analysing media content

    GB2555431A