WO2020179200A1 - 情報処理方法及び情報処理システム - Google Patents
情報処理方法及び情報処理システム Download PDFInfo
- Publication number
- WO2020179200A1 WO2020179200A1 PCT/JP2019/050480 JP2019050480W WO2020179200A1 WO 2020179200 A1 WO2020179200 A1 WO 2020179200A1 JP 2019050480 W JP2019050480 W JP 2019050480W WO 2020179200 A1 WO2020179200 A1 WO 2020179200A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- data
- weight
- information processing
- difference
- image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Ceased
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0475—Generative networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/094—Adversarial learning
Definitions
- This disclosure relates to an information processing method executed by a computer and an information processing system that executes the information processing method.
- the present disclosure provides an information processing method or the like that can more stably converge discriminator training in such a case.
- the information processing method is an information processing method executed by a computer, which acquires the first data and the second data which is simulated data based on the first data and uses the classifier.
- the first data is input to acquire the first identification result data
- the first difference between the reference data in the identification process for the first data by the classifier and the first identification result data is calculated, and the first difference is obtained.
- the second data is input to the discriminator to obtain second discrimination result data, and the discrimination is performed.
- a certain second weight is calculated, the discriminator is trained based on the first squared error data and the second squared error data, and the first weight and the second weight, and the first discrimination result data and
- the second identification result data is data of a tensor having a rank of 1 or more.
- the information processing system inputs the first data to an acquisition unit that acquires first data and second data that is simulated data based on the first data, and a discriminator. And second identification result data obtained by calculating a first difference between the first identification result data obtained by the above-described method and reference data in the identification process for the first data, and inputting the second data into the identifier.
- a second difference between the second data and reference data in the identification process is calculated, and based on the first difference, first square error data and a first weight that is a weight of the first square error data are calculated, A weight calculation unit that calculates the second weight, which is the weight of the second square error data and the second square error data, based on the second difference, the first square error data, and the second square error data.
- An error calculation unit that calculates error data used for training the discriminator based on the first weight and the second weight; and a training unit that trains the discriminator using the error data.
- the first identification result data and the second identification result data are tensor data of rank 1 or higher.
- FIG. 1 is a functional block diagram in the learning phase of the information processing system according to the embodiment.
- FIG. 2 is a functional block diagram in the inference phase of the information processing system according to the embodiment.
- FIG. 3 schematically shows a graph of a function indicating the weight set in the learning phase.
- FIG. 4 is a flowchart showing an operation procedure example of an information processing method executed for training a discriminator in the above information processing system.
- normalization of the weight prevents the discriminator output value from becoming an outlier. More specifically, the singular value of the weight matrix having the weight of each layer of the discriminator as an element is calculated, and the weight matrix is normalized using the norm of this singular value. Then, the weight matrix after normalization is updated based on the output error of the classifier.
- Cycle GAN is a GAN of Patch GAN system and LSGAN (Least Squares GAN) system.
- the output from the PatchGAN classifier is not a scalar value that takes a value of 0 or 1, but a simulated image output by the generator for each of the multiple subregions (patches) that divide the entire input image. Is a matrix whose elements are values indicating the determination result of whether the image is an image that is not output from the generator.
- the LSGAN discriminator is trained based on the squared error between this matrix output by the discriminator and the matrix indicating the correct answer in this discrimination (a matrix in which 0s or 1s are arranged).
- the matrix indicating the correct answer in this discrimination a matrix in which 0s or 1s are arranged.
- CYCLE GAN even a generator trained with a data set of an image that has not been registered shows a preferable conversion result (see Non-Patent Document 1). Therefore, for example, it is expected to have high practicability in applications where it is practically difficult to obtain an aligned data set in an amount necessary for training.
- the image data that can be obtained as an unaligned dataset contains a larger amount of noise than the aligned data set.
- the noise referred to here is, for example, due to variations in the image quality, the degree of focusing, or the tint of the image.
- the data set is a person image, the posture of the person in the image, the range of the body appearing in the image that changes depending on occlusion or composition, or changes in objects other than the person, such as the person's possession or background, It can be cited as an example of the cause of noise.
- Such noise affects the training of the above-mentioned PatchGAN and LSGAN discriminators. Specifically, it makes the discriminator training unstable and thus makes it difficult to obtain a generator that produces an image of the desired quality. Further, there is a problem that the above-mentioned conventional method cannot deal with the stabilization of the discriminator training due to such noise.
- the information processing method devised in view of such a problem is an information processing method executed by a computer, and is a first data and simulated data based on the first data.
- a certain second data is acquired, the first data is input to the classifier to acquire the first identification result data, and the reference data and the first identification result data in the identification process for the first data by the classifier are used.
- the first difference is calculated, the first weight which is the weight of the first square error data and the first square error data is calculated based on the first difference, and the second data is input to the classifier.
- the second identification result data is acquired, the second difference between the reference data and the second identification result data in the identification process for the second data by the classifier is calculated, and the second square error is calculated based on the second difference.
- a second weight that is a weight of the data and the second squared error data is calculated, and the discriminator is based on the first squared error data and the second squared error data, and the first weight and the second weight.
- the second weight may be further reduced to reduce the degree of influence of the second squared error data on the training of the discriminator.
- the first weight when the absolute value of the first difference exceeds the threshold value, the first weight may be set to zero, and when the absolute value of the second difference exceeds the threshold value, the second weight may be set to zero.
- the second data is generated and output from the first data by the generator, and is based on the first square error data, the second square error data, and the first weight and the second weight.
- the generator may be trained.
- the first data may be image data.
- the information processing system inputs the first data to an acquisition unit that acquires first data and second data that is simulated data based on the first data, and a discriminator. And second identification result data obtained by calculating a first difference between the first identification result data obtained by the above-described method and reference data in the identification process for the first data, and inputting the second data into the identifier.
- a second difference between the second data and reference data in the identification process is calculated, and based on the first difference, first square error data and a first weight that is a weight of the first square error data are calculated, A weight calculation unit that calculates the second weight, which is the weight of the second square error data and the second square error data, based on the second difference, the first square error data, and the second square error data.
- An error calculation unit that calculates error data used for training the discriminator based on the first weight and the second weight; and a training unit that trains the discriminator using the error data.
- the first identification result data and the second identification result data are tensor data of rank 1 or higher.
- a recording medium such as a device, an integrated circuit, or a computer-readable CD-ROM. It may be realized by any combination of integrated circuits, methods, computer programs and recording media.
- FIG. 1 and 2 are functional block diagrams showing a functional configuration example of the information processing system according to the embodiment.
- Each of these information processing systems is configured by using one or more information processing devices (computers) each having a processor and a memory to execute a program, and implements CycleGAN.
- the functional configuration for the learning phase of this information processing system is shown in FIG. 1, and the functional configuration for the inference phase is shown in FIG.
- the components of these functional configurations indicated by each block are realized, for example, by executing one or more programs in which a part or all of the above-mentioned processor is stored in a part or all of memory.
- an information processing system 10A includes a first conversion unit 11A, a determination unit 12, a weight calculation unit 13, a first error calculation unit 14, a training unit 15, a second conversion unit 16 and The second error calculator 17 is provided.
- the first conversion unit 11A performs a predetermined conversion on the Real image acquired by the information processing system 10A to generate and output a Fake image.
- the predetermined conversion is, for example, changing the image quality or the style (style) of the image.
- Changing the style of an image means, for example, changing an input photographed image into an image that looks as if it is a painting of a predetermined painter or style, or vice versa. It is to make the image by (Graphics) into an image that looks as if it is a live image.
- Another example of the predetermined conversion is to change the color included in the image according to a predetermined policy. For example, the input photographic image of the natural landscape looks like it was taken with the same composition in different seasons. Is to do so.
- the first conversion unit 11A is one of the two generators included in the CYCLE GAN implemented by the information processing system 10A, and is a generation model of the neural network used for the above-described conversion use. Is.
- the data of the Real image is an example of the first data in this embodiment
- the data of the Fake image is an example of the second data in this embodiment.
- the determination unit 12 performs an identification process for determining whether the input image is a Real image or a Fake image generated by the first conversion unit 11A, and outputs the result.
- This identification process is performed by the above-mentioned PatchGAN method, and the result of the identification process is in the form of a matrix whose elements are values indicating the likelihood of whether the image is a Real image or a Fake image for each small area. It is output. For example, an element corresponding to a small area determined to be a Real image is 1, an element corresponding to a small area determined to be a Fake image is 0, and an element corresponding to a small area other than these is each small area.
- a determination unit 12 is a discriminator in the CycleGAN implemented by the information processing system 10A, and is a discriminative model of a neural network used for the above-mentioned discriminative use.
- the identification result data output by receiving the input of the Real image is also referred to as first identification result data.
- the identification result data output after receiving the input of the Fake image is also referred to as the second identification result data.
- the weight calculation unit 13 calculates the difference between the identification result data output by the determination unit 12 by executing the identification process and the data indicating the correct answer in the identification process of the identification process (hereinafter, also referred to as reference data). Further, the weight calculation unit 13 calculates the weight and square error of each element of the matrix, which is the identification result data output by the determination unit 12, based on this difference.
- the reference data has the same size as the identification result data and is a matrix in which all elements are 1 or all elements are 0. According to the example used in the above description of the determination unit 12, it is a matrix having all elements 1 that indicates the correct answer of the matrix output by the determination unit 12 when the Real image is input.
- the weight and square error calculated by the weight calculation unit 13 with respect to the identification result output by the determination unit 12 into which the Real image is input are also referred to as the first weight and the first square error, respectively.
- the weight and the squared error calculated by the weight calculation unit 13 for the identification result output by the determination unit 12 to which the Fake image is input are also referred to as a second weight and a second squared error, respectively.
- the first error calculation unit 14 calculates the error of the determination unit 12 based on the first weight, the first squared error, the second weight, and the second squared error.
- the training unit 15 uses the error calculated by the first error calculation unit 14 to train the determination unit 12.
- the second conversion unit 16 is the other generator that is the generation model of the neural network in CycleGAN implemented by the information processing system 10A.
- the second conversion unit 16 receives an input of the Fake image generated and output by the first conversion unit 11A. Then, the Fake image is subjected to a conversion that restores the Real image before being converted into the Fake image, and the image generated by the conversion is output.
- the difference between the image output by the second conversion unit 16 and the correct image corresponding to this image, that is, the real image before being converted into the Fake image output by the first conversion unit 11A Calculate the error based on. This error is input to the training unit 15 and used for training of the first conversion unit 11A.
- each of these components is realized by one or more information processing devices that configure the information processing system 10A.
- the Real image set is a set of images that have not been subjected to conversion processing for the above-mentioned simulation by the first conversion unit 11A, and is a moving image including a plurality of still images or a plurality of frames. Consists of.
- the information processing system 10A may read and acquire the data of the Real image set recorded on a non-temporary recording medium such as a DVD (Digital Versatile Disc) or a semiconductor memory by using a reading device. Then, the image signal may be input from the camera and acquired.
- the information processing system 10A may further include a communication device and acquire the data of the Real image set via a signal received by the communication device.
- the information processing system 10B includes a conversion unit 11B which is a generation model obtained by machine learning.
- the conversion unit 11B is the first conversion unit 11A in which the above-described training for the predetermined conversion is repeatedly performed in the information processing system 10A, and, for example, such an evaluation regarding the conversion performance is a desired criterion. When it reaches, it can be treated as the conversion unit 11B.
- the conversion unit 11B performs a predetermined conversion on the unconverted image and outputs the converted image.
- the conversion unit 11B when the conversion unit 11B receives an input of a live-action image as an unconverted image, the conversion unit 11B converts the live-action image to generate and output a converted image that looks as if it is a painting in a predetermined style.
- the conversion unit 11B is realized by one or more information processing devices forming the information processing system 10B.
- the information processing apparatus constituting the information processing system 10B may be common to those constituting the information processing system 10A, and the conversion unit 11B may be the first conversion unit 11A itself whose training has converged to some extent or more. Further, the information processing device forming the information processing system 10B may be different from that forming the information processing system 10A.
- the first conversion unit 11A is on a plurality of stationary computers that form the information processing system 10A
- the conversion unit 11B is a microcontroller provided in a mobile body such as an automobile, a portable information terminal, or a household electric device. May be on top.
- the conversion unit 11B in this case may be obtained by reducing the weight (quantization) of the first conversion unit 11A.
- the weight calculation unit 13 calculates the difference between the identification result data output by the determination unit 12 after executing the identification process and the reference data in the identification process. Further, the weight calculation unit 13 calculates the weight (first weight, second weight) and square error (first square error, second square error) of each element of the matrix output by the determination unit 12 based on this difference. Calculate as
- the weight calculation unit 13 sets the magnitude of the first difference d 1 , that is, a threshold (hereinafter, referred to as an error allowable value) indicating a boundary of whether or not the absolute value is allowable, to T, and the first difference d 1 is calculated.
- a threshold hereinafter, referred to as an error allowable value
- the following function t (d 1 ) indicating the first weight corresponding to is calculated.
- FIG. 3 schematically shows a graph of the function t (d 1 ) showing the first weight.
- the first weight decreases from 1 and approaches 0 as the absolute value of the first difference d 1 increases from 0, and approaches the 0, and the absolute value of the first difference d 1 exceeds the error allowable value T. It is set to zero in the range.
- the determination unit 12 outputs the Fake image G(z) output by the first conversion unit 11A, which is a generator that receives the input of the Real image z, when the determination unit 12 inputs the Fake image G(z) as an identification target.
- the function t(d 2 ) indicating the second weight corresponding to the second difference d 2 is the same as the function t(d 1 ) It is calculated as follows.
- the graph of the function t (d 2 ) showing the second weight is also schematically represented as shown in FIG. That is, the second weight decreases from 1 to approach 0 as the absolute value of the second difference d 2 increases from 0, and becomes zero in the range where the absolute value of the second difference d 2 exceeds the error allowable value T. Is set.
- the weight calculation unit 13 further calculates the squared error of the identification result of the determination unit 12. Specifically, the first squared error (D 1 (x)-1) 2 is calculated based on the first difference D 1 (x)-1 obtained above for each element of the matrix that is the identification result data, The second square error (D 2 (G (z)) -0) 2 is calculated based on the second difference D 2 (G (z)) -0.
- the first error calculator 14 calculates the error of the determiner 12 based on the first weight, the first squared error, the second weight, and the second squared error calculated as described above. Specifically, each element of the first square error, first weight according to the value of d 1 (t (d 1) ) is multiplied. Hereinafter, this result is also referred to as a real image error. Further, to each element of the second square error, the second weight according to the value of d 2 (t (d 2) ) is multiplied by. Hereinafter, this result is also referred to as a Fake image error. Then, the result of adding the Real image error and the Fake image error is obtained as the error of the determination unit 12. The error of the determination unit 12 is used by the training unit 15 for training the determination unit 12.
- the meaning of applying the weights set as above to the squared error is as follows.
- the magnitude (absolute value) of the first difference or the second difference indicates the magnitude of the deviation from the correct answer of the determination for each portion (small area in the above example of the image) of the identification target data.
- each weight set as described above becomes smaller as the deviation from the correct answer of the judgment becomes larger.
- the portion including the outlier of the data input to the determination unit 12 for training can be determined to be far from the correct answer. Therefore, it is possible to suppress the influence of the outlier included in the identification target data on the training of the discriminator. In addition, the greater the degree of deviation of outliers, the stronger the suppression. In the above example, since the weight is set to zero in the portion where the deviation from the correct answer exceeds the threshold, the degree of influence on the training of the discriminator is zero.
- Tukey's biweight estimation method was used to suppress the effect of outliers, but the method of suppressing the effect of outliers is not limited to this.
- Another M estimation method that is a robust estimation method capable of setting the square error as described above may be used.
- FIG. 4 is a flow chart showing a procedure example of the operation of the information processing system 10A in which this information processing method is executed.
- Step S10 The determination unit 12, which is a discriminator, receives an input of an image acquired by the information processing system 10A from a set of Real images.
- Step S11 The determination unit 12 performs identification processing for determining whether the image is a Real image or a Fake image for each small area of the input image, and calculates and outputs a matrix based on the result.
- the calculated matrix is referred to as a first output matrix for convenience.
- Step S12 The weight calculator 13 calculates the first weight, which is the weight of each element, using the first output matrix calculated in step S11.
- the first weight which is the weight of each element, using the first output matrix calculated in step S11.
- For the method of calculating the first weight refer to the example given in “2. Suppressing the influence of outliers” above.
- Step S13 The weight calculator 13 calculates a first squared error which is a squared error of each element of the first output matrix calculated in step S11. For the method of calculating the first squared error, see “2. Suppressing the influence of outliers” above.
- Step S20 The determination unit 12 receives an input of the Fake image generated by converting the Real image by the first conversion unit 11A that is the generator.
- Step S21 The determination unit 12 performs identification processing for determining whether the image is a Real image or a Fake image for each small area of the input image, and calculates and outputs a matrix based on the result.
- the calculated matrix is referred to as a second output matrix for convenience.
- Step S22 The weight calculator 13 calculates the second weight, which is the weight of each element, using the second output matrix calculated in step S21.
- the method of calculating the second weight refer to the example given in “2. Suppressing the influence of outliers” above.
- Step S23 The weight calculation unit 13 calculates the second square error, which is the square error of each element of the second output matrix calculated in step S21. For the method of calculating the second squared error, see “2. Suppressing the influence of outliers” above.
- the first error calculation unit 14 calculates the Real image error by multiplying the first weight calculated in step S12 by the first squared error calculated in step S13.
- the first error calculation unit 14 calculates the Fake image error by multiplying the second weight calculated in step S22 by the second square error calculated in step S23.
- Step S32 The first error calculating unit 14 calculates the error of the determining unit 12 by adding the Real image error calculated in Step S30 and the Fake image error calculated in Step S31.
- Step S33 The training unit 15 uses the error calculated in step S32 to perform training of the determination unit 12, which is a discriminator.
- the content of the information processing method performed by the information processing system 10A is not limited to the above.
- training of the first conversion unit 11A which is a generator, is also performed by the training unit 15. This training is performed using, for example, the above-mentioned error calculated by the second error calculation unit 17.
- the first weight, the first square error data, the second weight, and the second square error data may be used for the training of the first conversion unit 11A.
- the information processing system according to the above-described embodiment has been described by using, as an example, the one in which CycleGAN is mounted, but the information processing system is not limited to this.
- the information processing system according to one aspect of the present disclosure is also applicable to other types of GANs of the Patch GAN system and the LS GAN system, such as Combo GAN.
- processing target by the information processing system is not limited to image data.
- processing targets include voice, distance point cloud, sensor data such as pressure, temperature, humidity, odor, and language data.
- the information processing system according to the above embodiment has been described by using the example in which the identification result is in the matrix format, but is not limited to this.
- the information processing system according to the present disclosure can be applied to information processing that handles identification result data that is tensor data of rank 1 or higher.
- a system LSI is an ultra-multifunctional LSI manufactured by integrating a plurality of components on a single chip. Specifically, a microprocessor, a ROM (Read-Only Memory), and a RAM (Random Access Memory) are used. It is a computer system configured to include the above. A computer program is stored in the ROM. When the microprocessor operates according to this computer program, the system LSI achieves the function of each component.
- the system LSI is used here, but it may also be called IC, LSI, super LSI, or ultra LSI depending on the degree of integration. Further, the method of making an integrated circuit is not limited to LSI, and may be realized by a dedicated circuit or a general-purpose processor. A field programmable gate array (FPGA) that can be programmed after the LSI is manufactured, or a reconfigurable processor capable of reconfiguring the connection and setting of circuit cells inside the LSI may be used.
- FPGA field programmable gate array
- One aspect of the present disclosure may be not only each of the above-described information processing systems, but also an information processing method having steps by processing by characteristic components included in the information processing system.
- This information processing method is, for example, the information processing method described with reference to the flowchart of FIG.
- one aspect of the present disclosure may be a computer program that causes a computer to execute the characteristic steps included in the information processing method.
- one aspect of the present disclosure may be a computer-readable non-transitory recording medium in which such a computer program is recorded.
- the present disclosure can be used for discriminator training in GAN.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
Description
発明者らは、上述した提案されている従来の方法に関し、以下の問題が生じることを見出した。
[1.構成]
図1及び図2は、実施の形態に係る情報処理システムの機能構成例を示す機能ブロック図である。これらの情報処理システムは、それぞれプロセッサ及びメモリを備えてプログラムを実行する情報処理装置(コンピュータ)を一台以上用いて構成され、CycleGANを実装する。この情報処理システムの学習フェーズのための機能構成は図1に、推論フェーズのための機能構成は図2に分けて示している。各ブロックが示すこれらの機能構成の構成要素は、例えば、上記のプロセッサの一部又は全部がメモリの一部又は全部に記憶される1個又は複数個のプログラムを実行することで実現される。
図1に示すように、実施の形態に係る情報処理システム10Aは、第1変換部11A、判定部12、重み算出部13、第1誤差算出部14、訓練部15、第2変換部16及び第2誤差算出部17を備える。
図2に示すように、実施の形態に係る情報処理システム10Bは、機械学習によって得られる生成モデルである変換部11Bを備える。具体的には、変換部11Bは、情報処理システム10Aにおいて上述の所定の変換のための訓練が繰返し実施された第1変換部11Aであり、例えばこのような変換性能に関する評価が所望の基準に達すると変換部11Bとして扱われ得る。変換部11Bは、未変換画像に所定の変換を実行して変換済画像を出力する。例えば変換部11Bは、未変換画像として実写画像の入力を受けると、この実写画像を変換して所定の様式の絵画であるかのように見える変換済画像を生成して出力する。
従来のGANの訓練フェーズにおいては、識別器の出力と正解を示すリファレンスデータとの差分に基づいて算出される二乗誤差が識別器の訓練に用いられている。これに対して本実施の形態に係る情報処理システム10Aによる訓練フェーズでは、識別器である判定部12の訓練における識別対象のデータに含まれる外れ値の影響を抑えるための処理がさらに実施される。この処理の具体例を以下に説明する。この例では、重みの設定にロバスト推定法のひとつであるTukeyのbiweight推定法を利用している。
d1=D1(x)-R1
の式で表される演算によって得る。なお、この場合のリファレンスデータはすべての要素の値が1の行列であるため、R1=1である。
d2=D2(G(z))-R2
の式で表される演算によって得る。なお、この場合のリファレンスデータはすべての要素の値が0の行列であるため、R2=0である。
情報処理システム10Aにおいて実行される、識別器である判定部12の訓練のための情報処理方法の動作について、その手順例を用いて説明する。図4は、この情報処理方法が実行される情報処理システム10Aの動作の手順例を示すフロー図である。
本開示の一又は複数の態様に係る情報処理方法及び情報処理システムは、上記の実施の形態の説明に限定されるものではない。本開示の趣旨を逸脱しない限り、当業者が想到する各種の変形を上記の実施の形態に施したものも、本開示の態様に含まれてもよい。下記にそのような変形の例、及び実施の形態の説明へのその他の補足事項を挙げる。
11A 第1変換部
11B 変換部
12 判定部
13 重み算出部
14 第1誤差算出部
15 訓練部
16 第2変換部
17 第2誤差算出部
Claims (6)
- コンピュータにより実行される情報処理方法であって、
第1データ及び前記第1データに基づく模擬的なデータである第2データを取得し、
識別器に前記第1データを入力して第1識別結果データを取得し、
前記識別器による第1データに対する識別処理におけるリファレンスデータと前記第1識別結果データとの第1差分を算出し、
前記第1差分に基づいて、第1二乗誤差データ及び前記第1二乗誤差データの重みである第1重みを算出し、
前記識別器に前記第2データを入力して第2識別結果データを取得し、
前記識別器による第2データに対する識別処理におけるリファレンスデータと前記第2識別結果データとの第2差分を算出し、
前記第2差分に基づいて、第2二乗誤差データ及び前記第2二乗誤差データの重みである第2重みを算出し、
前記第1二乗誤差データ及び前記第2二乗誤差データと、前記第1重み及び前記第2重みとに基づいて前記識別器を訓練し、
前記第1識別結果データ及び前記第2識別結果データは、階数1以上のテンソルのデータである
情報処理方法。 - 前記第1差分の絶対値が大きいほど前記第1重みをより小さくして前記第1二乗誤差データの前記識別器の訓練への影響度を下げ、
前記第2差分の絶対値が大きいほど、前記第2重みをより小さくして、前記第2二乗誤差データの前記識別器の訓練への影響度を下げる
請求項1に記載の情報処理方法。 - 前記第1差分の絶対値が閾値を超える場合、前記第1重みをゼロにし、
前記第2差分の絶対値が閾値を超える場合、前記第2重みをゼロにする
請求項2に記載の情報処理方法。 - 前記第2データは、生成器によって前記第1データから生成されて出力され、
前記第1二乗誤差データ及び前記第2二乗誤差データと、前記第1重み及び前記第2重みとに基づいて前記生成器を訓練する
請求項1~3のいずれか1項に記載の情報処理方法。 - 前記第1データは画像データである
請求項1~4のいずれか1項に記載の情報処理方法。 - 第1データ及び前記第1データに基づく模擬的なデータである第2データを取得する取得部と、
識別器に前記第1データを入力して取得される第1識別結果データと前記第1データに対する識別処理におけるリファレンスデータとの第1差分を算出し、前記識別器に前記第2データを入力して取得される第2識別結果データと前記第2データに対する識別処理におけるリファレンスデータとの第2差分を算出し、前記第1差分に基づいて、第1二乗誤差データ及び前記第1二乗誤差データの重みである第1重みを算出し、前記第2差分に基づいて、第2二乗誤差データ及び前記第2二乗誤差データの重みである第2重みを算出する重み算出部と、
前記第1二乗誤差データ及び前記第2二乗誤差データと、前記第1重み及び前記第2重みとに基づいて前記識別器の訓練に用いられる誤差データを算出する誤差算出部と、
前記誤差データを用いて前記識別器を訓練する訓練部とを備え、
前記第1識別結果データ及び前記第2識別結果データは、階数1以上のテンソルのデータである
情報処理システム。
Priority Applications (4)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| EP19918433.4A EP3937094A4 (en) | 2019-03-04 | 2019-12-24 | Information processing method and information processing system |
| JP2021503427A JPWO2020179200A1 (ja) | 2019-03-04 | 2019-12-24 | |
| CN201980060195.XA CN112703513B (zh) | 2019-03-04 | 2019-12-24 | 信息处理方法及信息处理系统 |
| US17/194,657 US20210192348A1 (en) | 2019-03-04 | 2021-03-08 | Information processing method and information processing system |
Applications Claiming Priority (4)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| US201962813348P | 2019-03-04 | 2019-03-04 | |
| US62/813,348 | 2019-03-04 | ||
| JP2019168679 | 2019-09-17 | ||
| JP2019-168679 | 2019-09-17 |
Related Child Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| US17/194,657 Continuation US20210192348A1 (en) | 2019-03-04 | 2021-03-08 | Information processing method and information processing system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2020179200A1 true WO2020179200A1 (ja) | 2020-09-10 |
Family
ID=72338641
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/JP2019/050480 Ceased WO2020179200A1 (ja) | 2019-03-04 | 2019-12-24 | 情報処理方法及び情報処理システム |
Country Status (5)
| Country | Link |
|---|---|
| US (1) | US20210192348A1 (ja) |
| EP (1) | EP3937094A4 (ja) |
| JP (1) | JPWO2020179200A1 (ja) |
| CN (1) | CN112703513B (ja) |
| WO (1) | WO2020179200A1 (ja) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2023131139A (ja) * | 2022-03-08 | 2023-09-21 | 日本電気株式会社 | データ処理方法及び電子機器 |
| JP2024507766A (ja) * | 2021-02-15 | 2024-02-21 | コーニンクレッカ フィリップス エヌ ヴェ | 機械学習によるコントラスト強調 |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019004350A1 (ja) | 2017-06-29 | 2019-01-03 | 株式会社 Preferred Networks | データ識別器訓練方法、データ識別器訓練装置、プログラム及び訓練方法 |
| JP7784239B2 (ja) * | 2021-04-23 | 2025-12-11 | キヤノン株式会社 | 情報処理装置、情報処理方法、及びプログラム |
-
2019
- 2019-12-24 EP EP19918433.4A patent/EP3937094A4/en not_active Withdrawn
- 2019-12-24 WO PCT/JP2019/050480 patent/WO2020179200A1/ja not_active Ceased
- 2019-12-24 JP JP2021503427A patent/JPWO2020179200A1/ja active Pending
- 2019-12-24 CN CN201980060195.XA patent/CN112703513B/zh active Active
-
2021
- 2021-03-08 US US17/194,657 patent/US20210192348A1/en not_active Abandoned
Non-Patent Citations (4)
| Title |
|---|
| ANONYMOUS: "PyTorch (15) CycleGAN (horse2zebra)-A breakthrough on artificial intelligence", 24 March 2018 (2018-03-24), pages 1 - 23, XP055736644, Retrieved from the Internet <URL:http://aidiary.hatenablog.com/entry/20180324/1521896184> [retrieved on 20201005] * |
| OSAMU HORI: "A video text extraction method for chracter recognition", RESEARCH REPORT OF INFORMATION PROCESSING SOCIETY OF JAPAN, vol. 99, no. 3(CVIM-114), 22 January 1999 (1999-01-22), pages 129 - 136, XP009523392, ISSN: 0919-6072 * |
| See also references of EP3937094A4 * |
| YUKI SAOTOME; SONOGASHIRA MOTOHARU; MASAAKI IIYAMA: "Simultaneous estimation of object region and depth in foggy scenes using ToF camera", IPSJ SIG TECHNICAL REPORTS, vol. 2018, no. 16 (2018-DCC-020), 31 October 2018 (2018-10-31), pages 1 - 8, XP009523851, ISSN: 2188-8868 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2024507766A (ja) * | 2021-02-15 | 2024-02-21 | コーニンクレッカ フィリップス エヌ ヴェ | 機械学習によるコントラスト強調 |
| JP2023131139A (ja) * | 2022-03-08 | 2023-09-21 | 日本電気株式会社 | データ処理方法及び電子機器 |
| JP7501703B2 (ja) | 2022-03-08 | 2024-06-18 | 日本電気株式会社 | データ処理方法及び電子機器 |
Also Published As
| Publication number | Publication date |
|---|---|
| EP3937094A1 (en) | 2022-01-12 |
| EP3937094A4 (en) | 2022-04-27 |
| JPWO2020179200A1 (ja) | 2020-09-10 |
| CN112703513A (zh) | 2021-04-23 |
| US20210192348A1 (en) | 2021-06-24 |
| CN112703513B (zh) | 2025-05-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110574077B (zh) | 利用虚拟三维深层神经网络的影像分析装置及方法 | |
| CN111723813B (zh) | 基于类内判别器的弱监督图像语义分割方法、系统、装置 | |
| CN109493350B (zh) | 人像分割方法及装置 | |
| CN110147721B (zh) | 一种三维人脸识别方法、模型训练方法和装置 | |
| JP6798183B2 (ja) | 画像解析装置、画像解析方法およびプログラム | |
| WO2020179200A1 (ja) | 情報処理方法及び情報処理システム | |
| JP6866889B2 (ja) | 画像処理装置、画像処理方法およびプログラム | |
| US9305359B2 (en) | Image processing method, image processing apparatus, and computer program product | |
| CN114913588B (zh) | 一种应用于复杂场景下的人脸图像修复及识别方法 | |
| GB2606785A (en) | Adaptive convolutions in neural networks | |
| CN110023989B (zh) | 一种素描图像的生成方法及装置 | |
| CN112132167B (zh) | 图像生成和神经网络训练方法、装置、设备和介质 | |
| CN113658324A (zh) | 图像处理方法及相关设备、迁移网络训练方法及相关设备 | |
| KR102306319B1 (ko) | 포즈 추정 매핑 데이터를 생성하는 방법, 프로그램 및 컴퓨팅 장치 | |
| CN110856014A (zh) | 动态图像生成方法、装置、电子设备及存储介质 | |
| Zeng et al. | SAU-Net: Saliency-Based Adaptive Unfolding Network for Interpretable High Quality Image Compressed Sensing in Internet of Things | |
| WO2024145343A1 (en) | Avatar generation according to artistic styles | |
| CN112884648A (zh) | 多类模糊图像超分辨率重建的方法和系统 | |
| CN111368763A (zh) | 基于头像的图像处理方法、装置及计算机可读存储介质 | |
| KR102488858B1 (ko) | 손상된 유형물의 디지털 복원 방법, 장치 및 프로그램 | |
| CN114943878A (zh) | 一种神经网络预训练方法、图像处理方法及装置 | |
| CN111414864B (zh) | 一种人脸活体检测方法和相关装置 | |
| CN112614199A (zh) | 语义分割图像转换方法、装置、计算机设备和存储介质 | |
| CN118570052A (zh) | 图像风格迁移方法、装置、存储介质与电子设备 | |
| Kannan et al. | Quality Assessment of Low-Light Restored Images: A Subjective Study and an Unsupervised Model |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 19918433 Country of ref document: EP Kind code of ref document: A1 |
|
| ENP | Entry into the national phase |
Ref document number: 2021503427 Country of ref document: JP Kind code of ref document: A |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |
|
| ENP | Entry into the national phase |
Ref document number: 2019918433 Country of ref document: EP Effective date: 20211004 |
|
| WWG | Wipo information: grant in national office |
Ref document number: 201980060195.X Country of ref document: CN |