WO2022250071A1 - 学習方法、学習装置及びプログラム - Google Patents
学習方法、学習装置及びプログラム Download PDFInfo
- Publication number
- WO2022250071A1 WO2022250071A1 PCT/JP2022/021329 JP2022021329W WO2022250071A1 WO 2022250071 A1 WO2022250071 A1 WO 2022250071A1 JP 2022021329 W JP2022021329 W JP 2022021329W WO 2022250071 A1 WO2022250071 A1 WO 2022250071A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- image
- label
- synthetic
- learning
- teacher
- Prior art date
Links
- 238000000034 method Methods 0.000 title claims abstract description 61
- 238000010801 machine learning Methods 0.000 claims abstract description 22
- 239000002131 composite material Substances 0.000 claims abstract description 17
- 238000009826 distribution Methods 0.000 claims description 5
- 230000006870 function Effects 0.000 claims description 3
- 238000009827 uniform distribution Methods 0.000 claims description 3
- 238000010586 diagram Methods 0.000 description 14
- 238000012545 processing Methods 0.000 description 7
- 238000004590 computer program Methods 0.000 description 4
- 230000010365 information processing Effects 0.000 description 4
- 238000012549 training Methods 0.000 description 2
- 241000282326 Felis catus Species 0.000 description 1
- 239000000470 constituent Substances 0.000 description 1
- 230000006866 deterioration Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000000873 masking effect Effects 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000013178 mathematical model Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000000877 morphologic effect Effects 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/73—Deblurring; Sharpening
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/50—Image enhancement or restoration using two or more images, e.g. averaging or subtraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20076—Probabilistic image processing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20212—Image combination
Definitions
- the present disclosure relates to a learning method and the like for generating a learning model used for image recognition.
- Patent Document 1 shows that arbitrary noise is added to the image in order to enable generation of a more versatile and robust classifier.
- the present disclosure provides a learning method and the like capable of generating a learning model that is robust against noise.
- a learning method is a learning method for generating a learning model used for image recognition, wherein a first image is generated by adding noise to a first region of an original image, By adding noise to a second region of the original image excluding the first region to generate a second image, and performing weighted addition of the first image and the second image at a first ratio , generating a composite image, performing a weighted addition of a first base label corresponding to the correct label of the original image and a second base label corresponding to the incorrect label of the original image, the size of the first region;
- a first teacher label is generated for the first image by performing at a second ratio, which is a ratio to the size of the second region, and weighted addition of the first base label and the second base label is performed on the second base label.
- a second teacher label is generated for the second image by performing a ratio opposite to the ratio of 2, and a weighted addition of the first teacher label and the second teacher label is performed by the first ratio,
- the learning model is generated by generating a synthetic teacher label for the synthetic image and performing machine learning using the synthetic image and the synthetic teacher label.
- FIG. 1 is a conceptual diagram showing the results of recognition processing in a reference example.
- FIG. 2 is a conceptual diagram showing learning in the reference example.
- FIG. 3 is a conceptual diagram showing a partial noise added image in the reference example.
- FIG. 4 is a block diagram showing the configuration of the learning device according to this embodiment.
- FIG. 5 is a flow chart showing the operation of the learning device according to this embodiment.
- FIG. 6 is a conceptual diagram showing details of generation of a synthesized image in this embodiment.
- FIG. 7 is a conceptual diagram showing the details of synthetic teacher label generation in this embodiment.
- FIG. 8 is a data diagram showing comparison results of recognition accuracies in this embodiment.
- noise-added image by adding noise to an image and performing machine learning using the noise-added image, it is possible to generate a learning model that is robust against noise. Further, by adding noise to a part of an image and performing machine learning using the image to which noise is added to a part, a learning model that is more robust against noise may be generated.
- a learning method is a learning method for generating a learning model used for image recognition, and by adding noise to a first region of the original image, the first image and adding noise to a second region of the original image excluding the first region to generate a second image, and performing weighted addition of the first image and the second image at a first ratio to generate a synthesized image, and weighted addition of a first base label corresponding to the correct label of the original image and a second base label corresponding to the incorrect label of the original image is performed in the first region and the size of the second region to generate a first teacher label for the first image, weighting the first base label and the second base label A second teacher label for the second image is generated by performing addition at a rate opposite to the second rate, and weighted addition of the first teacher label and the second teacher label is performed at the first rate.
- a synthetic teacher label for the synthetic image is generated, and machine learning is performed using the synthetic image and the synthetic teacher label to generate the learning
- the learning method generates the first image, the second image, the synthetic image, the first teacher label, the second teacher label, and the synthetic teacher label for each of a plurality of first regions.
- a plurality of synthetic images and a plurality of synthetic teacher labels are generated, and machine learning is performed using the plurality of synthetic images and the plurality of synthetic teacher labels, thereby generating the learning model.
- the learning method generates a plurality of synthetic images and a plurality of synthetic teacher labels by generating the synthetic images and the synthetic teacher labels for each of the plurality of first ratios, and
- the learning model is generated by performing machine learning using the image and the plurality of synthetic teacher labels.
- the first region is where W indicates the width of the original image, H indicates the height of the original image, r x1 indicates the left edge of the first area, r y1 indicates the top edge of the first area , r x2 indicates the right end of the first region, r y2 indicates the bottom end of the first region, a to U [b, c] are according to a uniform distribution of a from b to c Indicates that it is determined.
- the first ratio is determined according to the beta distribution of B( ⁇ , ⁇ ), where B represents a beta function and ⁇ represents a positive real number.
- a learning device is a learning device that generates a learning model used for image recognition, and includes a processor and a memory, and the processor uses the memory to generate an original image. generating a first image by adding noise to a first region of the original image; generating a second image by adding noise to a second region of the original image excluding the first region; A composite image is generated by performing weighted addition of the first image and the second image at a first ratio, and a first base label corresponding to the correct label of the original image and an incorrect label of the original image are obtained.
- a first teacher label for the first image is generated by performing a weighted addition with the corresponding second base label at a second ratio, which is the ratio between the size of the first region and the size of the second region. and performing weighted addition of the first basic label and the second basic label at a ratio opposite to the second ratio to generate a second teacher label for the second image, and the first teacher label and the second teacher label at the first ratio to generate a synthetic teacher label for the synthetic image, and perform machine learning using the synthetic image and the synthetic teacher label to perform the Generate a learning model.
- the learning device can execute the above learning method.
- the above learning method is implemented by a learning device.
- the program according to one aspect of the present disclosure may be a program for causing a computer to execute the learning method.
- the program can cause the computer to execute the above learning method. Then, the above learning method is implemented by a program.
- FIG. 1 is a conceptual diagram showing the result of recognition processing in the reference example.
- a noise added image is generated by adding noise to the original image.
- the noise added image is generated by adding the original image and the image obtained by multiplying the noise image by the weight of ⁇ .
- image recognition is performed on this noise-added image, a correct recognition result may not be obtained.
- adding two images together means adding together the pixel values of corresponding pixels in the two images.
- FIG. 2 is a conceptual diagram showing learning in the reference example.
- a model used for image recognition is trained using a noise added image. Specifically, the model is updated so that the noise-added image is correctly recognized as representing a "dog".
- the model used for image recognition is a mathematical model, also called a recognition model or a learning model.
- the model used for image recognition may be a neural network model. Also, as described above, learning that is performed by adding noise to the original image is an example of adversarial training.
- FIG. 3 is a conceptual diagram showing a partial noise added image in the reference example.
- the partial noise added image is an image obtained by adding noise to a partial region of the original image rather than the entire original image.
- a mask image is generated in which areas other than the area to which noise is added are masked.
- each pixel value in the area to which noise is added is set to 1
- each pixel value other than the area to which noise is added is set to 0.
- a noise image is generated that is composed of noise over the entire area of the image.
- a noisy image may, for example, consist of uniform noise over the entire area of the image.
- the partial noise added image is generated by adding the partial noise image and the original image pixel by pixel.
- the model may be trained using such a partially noise-added image. This makes it possible to perform learning using more patterns and obtain a model that is more robust against noise.
- noise is added to a part of the area and no noise is added to the remaining area.
- a partially noise-added image in which the noise addition method is greatly different for each region may not be suitable for learning.
- the label corresponding to the original image may not be appropriate as the label corresponding to the partial noise added image.
- FIG. 4 is a block diagram showing the configuration of the learning device according to this embodiment.
- a learning device 100 shown in FIG. 4 includes a processor 101 and a memory 102 .
- Learning device 100 may be a computer.
- the processor 101 is, for example, a dedicated or general-purpose electric circuit that performs information processing, and is a circuit that can access the memory 102 .
- Processor 101 may be a processor such as a CPU. Also, the processor 101 may be an assembly of a plurality of electric circuits. Further, the processor 101 may perform information processing by reading a program from the memory 102 and executing the program. The processor 101 may perform machine learning as information processing, or may perform image recognition.
- the processor 101 generates images for learning and labels corresponding to the images. Specifically, the processor 101 acquires the original image for learning and the original label corresponding to the original image, and from the original image and the original label, the additional image for learning and the additional image. Generate corresponding additional labels.
- the processor 101 learns the model using images for learning and labels corresponding to the images. For example, the processor 101 learns by updating the model so that the label output from the model by inputting an image to the model matches the label corresponding to the image. Also, the processor 101 may perform image recognition using a trained model.
- the memory 102 is, for example, a dedicated or general-purpose electric circuit that stores information for the processor 101 to process information. Memory 102 may be connected to processor 101 or may be included in processor 101 . Also, the memory 102 may be an assembly of a plurality of electric circuits.
- the memory 102 may be a non-volatile memory or a volatile memory. Also, the memory 102 may be a magnetic disk, an optical disk, or the like, or may be expressed as a storage, recording medium, or the like. Also, the memory 102 may be a non-temporary recording medium such as a CD-ROM.
- the memory 102 may store a model used for image recognition, may store a recognition target image, or may store a recognition result. Further, the memory 102 may store a program for the processor 101 to perform information processing.
- FIG. 4 shows an example of the configuration of the learning device 100
- the configuration of the learning device 100 is not limited to the example shown in FIG.
- the learning device 100 may include multiple components that perform multiple processes described below.
- FIG. 5 is a flow chart showing the operation of the learning device 100 shown in FIG. Specifically, in learning device 100, processor 101 uses memory 102 to perform the operation shown in FIG.
- the processor 101 generates a first image by adding noise to a first region of the original image (S101). Also, the processor 101 generates a second image by adding noise to a second area of the original image excluding the first area (S102). Then, the processor 101 generates a composite image by performing weighted addition of the first image and the second image at the first ratio (S103).
- the processor 101 generates a first teacher label for the first image by performing weighted addition of the first basic label and the second basic label at a second ratio (S104). Also, the processor 101 generates a second teacher label for the second image by performing weighted addition of the first basic label and the second basic label at a ratio opposite to the second ratio (S105). Then, a weighted addition of the first teacher label and the second teacher label is performed at a first ratio to generate a synthetic teacher label for the synthetic image (S106).
- the first basic label corresponds to the correct label of the original image
- the second basic label corresponds to the incorrect label of the original image.
- Each label described above is not limited to a label representing one correct class, but may be a so-called soft label, or may represent a plurality of likelihoods for a plurality of classes.
- the second ratio is the ratio between the size of the first area and the size of the second area.
- processor 101 generates a learning model by performing machine learning using the synthesized image and the synthesized teacher label (S107). Specifically, processor 101 generates a learning model such that a synthetic teacher label is output when a synthetic image is input to the learning model.
- the learning device 100 can add noise according to the first ratio to each of the first area of the original image and the second area of the original image excluding the first area. Therefore, learning device 100 can prevent noise from being added in a manner that differs greatly depending on the area. Therefore, learning device 100 can generate an image suitable for learning.
- the learning device 100 can synthesize two teacher labels using the same ratio as that used to synthesize two images. Therefore, learning device 100 can generate an appropriate synthetic teacher label for a synthetic image.
- the learning device 100 can generate a learning model that is robust against noise by using the synthesized image and the synthesized teacher label.
- the learning device 100 may include a plurality of components respectively corresponding to the plurality of processes (S101 to S107) described above.
- the learning device 100 includes a first image generator, a second image generator, a synthetic image generator, a first teacher label generator, a second teacher label generator, a synthetic teacher label generator, and a learning model generator. may be provided.
- the processor 101 may generate a plurality of synthetic images and a plurality of synthetic teacher labels by performing the above processing (S101 to S106) for each of the plurality of first regions. Then, the processor 101 may generate a learning model by performing machine learning using multiple synthetic images and multiple synthetic teacher labels.
- the plurality of first areas are, for example, different areas in the original image. Also, the plurality of first regions may partially overlap.
- learning device 100 can generate various synthetic images and various synthetic teacher labels according to various first regions.
- learning device 100 can generate a learning model that is robust against noise.
- the processor 101 generates a synthetic image (S103) and a synthetic teacher label (S106) for each of the plurality of first ratios, thereby creating a plurality of synthetic images and a plurality of synthetic teacher labels. You may Then, the processor 101 may generate a learning model by performing machine learning using multiple synthetic images and multiple synthetic teacher labels.
- the processor 101 performs the above processes (S101 to S106) for each of the plurality of first regions, and generates a synthetic image and a synthetic teacher label for each of the plurality of first ratios (S103 and S106) may be performed. And thereby, the processor 101 may generate multiple synthetic images and multiple synthetic teacher labels. Then, the processor 101 may generate a learning model by performing machine learning using multiple synthetic images and multiple synthetic teacher labels.
- the learning device 100 can generate various synthetic images and various synthetic teacher labels according to various first regions and various first ratios.
- learning device 100 can generate a learning model that is robust against noise.
- FIG. 6 is a conceptual diagram showing details of generation of a synthesized image in this embodiment. Specifically, first, the processor 101 determines a first area of the original image, and determines a second area of the original image excluding the first area. Processor 101 may determine the first region according to the following formula.
- W indicates the width of the original image
- H indicates the height of the original image.
- rx1 indicates the left edge of the first area
- ry1 indicates the top edge of the first area
- rx2 indicates the right edge of the first area
- ry2 indicates the bottom edge of the first area.
- a ⁇ U[b,c] indicates that a is determined according to a uniform distribution from b to c. Thereby, the first area is appropriately determined according to the size of the original image.
- the processor 101 generates a first mask image by masking areas other than the first area (that is, the second area) out of the entire area of the original image.
- each pixel value in the first region is set to 1
- each pixel value in the second region other than the first region is set to 0.
- the processor 101 also generates a second mask image in which the entire area of the original image is masked except for the second area (that is, the first area).
- each pixel value in the second area is set to 1, and each pixel value in the first area other than the second area is set to 0.
- the processor 101 generates a noise image whose entire area is composed of the same type of noise. Then, by multiplying the first mask image and the noise image pixel by pixel, a first noise image including noise only in the first region is generated. Further, by multiplying the second mask image and the noise image pixel by pixel, a second noise image including noise only in the second region is generated.
- the first noise image and the second noise image can also be expressed as a first partial noise image and a second partial noise image, respectively.
- the processor 101 adds the first noise image and the original image pixel by pixel to generate the first image.
- a first image is generated in which noise is added to the first region of the original image.
- the processor 101 also generates a second image by adding the second noise image and the original image pixel by pixel.
- a second image is generated in which noise is added to the second area of the original image, excluding the first area.
- the first image and the second image can also be expressed as a first partial noise-added image and a second partial noise-added image, respectively.
- the processor 101 performs weighted addition of the first image in which noise is added to the first region and the second image in which noise is added to the second region at a first ratio to generate a composite image. .
- the processor 101 weights the first image by ⁇ 2 and weights the second image by 1 ⁇ 2 , and performs a weighted addition of these pixel by pixel to generate a composite image.
- ⁇ 2 is a value from 0 to 1, specifically, it may be a value of 0 or more and 1 or less, or may be a value greater than 0 and less than 1.
- Processor 101 may determine ⁇ 2 according to the beta distribution of B( ⁇ , ⁇ ).
- B indicates a beta function and ⁇ indicates a positive real number. This makes it possible to generate a synthetic image and a synthetic teacher label using the first ratio corresponding to ⁇ 2 that is appropriately determined according to a symmetrical probability distribution. Then, when a plurality of data sets are generated from the original image and the original teacher label, occurrence of bias in the plurality of data sets is suppressed.
- a composite image is generated appropriately.
- the above process is an example of the process for generating a composite image, and the process for generating a composite image is not limited to the above process.
- the mask image, the noise image, the first noise image, and the second noise image may not be used, and the same type of noise is directly applied to each region of the original image, so that the first image and the second image are may be generated.
- FIG. 7 is a conceptual diagram showing the details of synthetic teacher label generation in this embodiment.
- ⁇ corresponds to the second ratio between the size of the first area and the size of the second area. Specifically, ⁇ indicates the ratio of the size of the first area to the size of the original image, and 1 ⁇ indicates the ratio of the size of the second area to the size of the original image.
- the first basic label corresponds to the correct label of the original image and may be expressed as the correct label.
- a correct label is a label indicating the correct class of an object appearing in the original image. That is, the first base label may correspond to the teacher label for the original image.
- the first base label may have a 100% likelihood for the correct class of objects in the original image and a 0% likelihood for each other class. For example, the first base label may have a 100% likelihood for the dog class and a 0% likelihood for each of the other classes.
- the second basic label corresponds to the incorrect label of the original image and may be expressed as the incorrect label.
- An incorrect label is a label indicating an incorrect class of an object appearing in the original image. That is, the second base label may correspond to the teacher label for the noise image.
- the second base label may have a 0% likelihood for the correct class of objects in the original image and a greater than 0% likelihood for each other class.
- the second base label may have a 0% likelihood for the dog class and a few percent likelihood for each of the other classes. More specifically, the second base label may have a likelihood of 1/total number of classes for each other class.
- the total number of classes may be the total number of other classes.
- y1 corresponds to the first teacher label for the first image.
- y 1 is by performing a weighted addition of the first and second base labels corresponding to the correct and incorrect labels, respectively, according to the ratio of the noisy and non-noisy regions in the first image. can get. Specifically, as shown in FIG. 7, y 1 is obtained by giving a weight of ⁇ to the first basic label, giving a weight of 1 ⁇ to the second basic label, and performing weighted addition of these.
- y2 corresponds to the second teacher label for the second image. y2 by performing a weighted addition of the first and second base labels corresponding to the correct and incorrect labels, respectively, according to the ratio of noisy to non-noisy regions in the second image. can get. Specifically, as shown in FIG. 7, y2 is obtained by weighting the second basic label with ⁇ and weighting the first basic label with 1 ⁇ and performing weighted addition.
- y2 is obtained by performing a weighted addition of the first and second basic labels in a ratio opposite to that of y1 .
- a reverse ratio means that the weight given to the first base label and the second base label are exchanged.
- y corresponds to the synthetic teacher label for the synthetic image.
- y is obtained by giving a weight of ⁇ 2 to the first teacher label (y 1 ), a weight of 1 ⁇ 2 to the second teacher label (y 2 ), and performing weighted addition of these.
- ⁇ 2 corresponds to the first ratio. That is, the ratio for weighted addition of the first teacher label and the second teacher label is the same as the ratio for generating the synthesized image.
- a synthetic teacher label is generated by the above processing.
- the generation of the first teacher label for the first image and the generation of the second teacher label for the second image reflect the ratio of the noisy region.
- the weighted addition of the first teacher label for the first image and the second teacher label for the second image reflects the first ratio for the weighted addition of the first image and the second image. Therefore, an appropriate synthesized teacher label is generated for a synthesized image in which noise is added to each region.
- FIG. 8 is a data diagram showing comparison results of recognition accuracies in this embodiment. Specifically, FIG. 8 shows the recognition accuracy based on the learning method of the reference example described in FIG. 3 and the recognition accuracy of the present embodiment described in FIGS. A comparison with the recognition accuracy based on the morphological learning method is shown.
- FGSM Fast Gradient Sign Method
- PGD Project Gradient Descent
- PDG-20 PDG-20
- CIFAR Canadian Institute For Advanced Research
- the learning method of the present embodiment suppresses deterioration of recognition accuracy against various noises.
- the recognition accuracy of the learning method of the present embodiment is slightly lower than that of the learning method of the reference example, but it is 90% or more, which is a level without problems.
- aspects of the learning method have been described above based on the embodiments, the aspects of the learning method are not limited to the embodiments. Modifications conceived by those skilled in the art may be applied to the embodiments, and multiple constituent elements in the embodiments may be combined arbitrarily. For example, operations performed by a particular component in an embodiment may be performed by another component in place of a particular component. In addition, the order of multiple processes may be changed, and multiple processes may be executed in parallel.
- the learning method may be executed by any device or system. That is, the learning method may be performed by a learning device, or by another device or system.
- the above learning method may be executed by a computer including a processor, memory, input/output circuits, and the like.
- the learning method may be executed by the computer executing a program for causing the computer to execute the learning method.
- the program may be recorded in a non-temporary computer-readable recording medium such as a CD-ROM.
- the above program is a learning method for generating a learning model to be used for image recognition in a computer, wherein the first image is generated by adding noise to the first region of the original image, and the By adding noise to a second region of the original image excluding the first region to generate a second image, and performing weighted addition of the first image and the second image at a first ratio, A synthetic image is generated, and a weighted addition of a first base label corresponding to the correct label of the original image and a second base label corresponding to the incorrect label of the original image is performed with the size of the first region and the A first teacher label is generated for the first image by performing a second ratio, which is a ratio to the size of the second region, and weighted addition of the first base label and the second base label is performed on the second base label.
- a second ratio which is a ratio to the size of the second region
- a second teacher label for the second image is generated by performing a ratio opposite to the ratio, and a weighted addition of the first teacher label and the second teacher label is performed by the first ratio, so that the A learning method for generating the learning model is executed by generating a synthetic teacher label for the synthetic image and performing machine learning using the synthetic image and the synthetic teacher label.
- the plurality of components of the learning device that executes the learning method may be composed of dedicated hardware, may be composed of general-purpose hardware that executes the above program, etc., or may be a combination thereof.
- the general-purpose hardware may be composed of a memory in which a program is stored, a general-purpose processor that reads and executes the program from the memory, and the like.
- the memory may be a semiconductor memory, a hard disk, or the like
- the general-purpose processor may be a CPU or the like.
- dedicated hardware may be configured with a memory, a dedicated processor, and the like.
- a dedicated processor may refer to the memory and execute the learning method described above.
- each component of the learning device that executes the learning method may be an electric circuit.
- These electric circuits may form one electric circuit as a whole, or may be separate electric circuits. Further, these electric circuits may correspond to dedicated hardware, or may correspond to general-purpose hardware for executing the above-described programs and the like.
- a method of generating learning data is a method of generating learning data for generating a learning model used for image recognition by machine learning. generating a first image, adding noise to a second area of the original image excluding the first area to generate a second image, and performing weighted addition of the first image and the second image; A synthetic image is generated by performing at a ratio of 1, and the weighted addition of the first base label corresponding to the correct label of the original image and the second base label corresponding to the incorrect label of the original image is performed by the weighted addition of the first base label corresponding to the incorrect label of the original image.
- a first teacher label for the first image is generated by performing at a second ratio, which is a ratio of the size of the first region and the size of the second region, and the first base label and the second base label are generated. at a ratio opposite to the second ratio to generate a second teacher label for the second image, and weighted addition of the first teacher label and the second teacher label is performed at the first Learning data including the synthetic image and the synthetic teacher label corresponding to the synthetic image is generated by generating the synthetic teacher label for the synthetic image by performing the ratio.
- the composite image may have a third area whose noise is different from that of the first and second areas.
- a synthetic teacher label may then be generated based on the size of the first area, the size of the second area, and the size of the third area.
- the first area in the composite image is a rectangular area, it may be a non-rectangular area.
- the present disclosure is useful, for example, for learning devices that generate learning models used for image recognition, and is applicable to image recognition systems, character recognition systems, biometric authentication systems, and the like.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Databases & Information Systems (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
Description
101 プロセッサ
102 メモリ
Claims (7)
- 画像認識に用いられる学習モデルを生成する学習方法であって、
元画像のうちの第1領域にノイズを付与することにより、第1画像を生成し、
前記元画像のうち前記第1領域を除く第2領域にノイズを付与することにより、第2画像を生成し、
前記第1画像と前記第2画像との重み付け加算を第1比率で行うことにより、合成画像を生成し、
前記元画像の正解ラベルに対応する第1基礎ラベルと、前記元画像の不正解ラベルに対応する第2基礎ラベルとの重み付け加算を、前記第1領域のサイズと、前記第2領域のサイズとの比率である第2比率で行うことにより、前記第1画像に対する第1教師ラベルを生成し、
前記第1基礎ラベルと前記第2基礎ラベルとの重み付け加算を前記第2比率とは逆の比率で行うことにより、前記第2画像に対する第2教師ラベルを生成し、
前記第1教師ラベルと前記第2教師ラベルとの重み付け加算を前記第1比率で行うことにより、前記合成画像に対する合成教師ラベルを生成し、
前記合成画像及び前記合成教師ラベルを用いて機械学習を行うことにより、前記学習モデルを生成する
学習方法。 - 複数の第1領域のそれぞれについて、前記第1画像、前記第2画像、前記合成画像、前記第1教師ラベル、前記第2教師ラベル及び前記合成教師ラベルを生成することにより、複数の合成画像及び複数の合成教師ラベルを生成し、
前記複数の合成画像及び前記複数の合成教師ラベルを用いて機械学習を行うことにより、前記学習モデルを生成する
請求項1に記載の学習方法。 - 複数の第1比率のそれぞれについて、前記合成画像及び前記合成教師ラベルを生成することにより、複数の合成画像及び複数の合成教師ラベルを生成し、
前記複数の合成画像及び前記複数の合成教師ラベルを用いて機械学習を行うことにより、前記学習モデルを生成する
請求項1又は2に記載の学習方法。 - 前記第1比率は、B(α、α)のベータ分布に従って決定され、
Bは、ベータ関数を示し、
αは、正の実数を示す
請求項1又は2に記載の学習方法。 - 画像認識に用いられる学習モデルを生成する学習装置であって、
プロセッサ及びメモリを備え、
前記プロセッサは、前記メモリを用いて、
元画像のうちの第1領域にノイズを付与することにより、第1画像を生成し、
前記元画像のうち前記第1領域を除く第2領域にノイズを付与することにより、第2画像を生成し、
前記第1画像と前記第2画像との重み付け加算を第1比率で行うことにより、合成画像を生成し、
前記元画像の正解ラベルに対応する第1基礎ラベルと、前記元画像の不正解ラベルに対応する第2基礎ラベルとの重み付け加算を、前記第1領域のサイズと、前記第2領域のサイズとの比率である第2比率で行うことにより、前記第1画像に対する第1教師ラベルを生成し、
前記第1基礎ラベルと前記第2基礎ラベルとの重み付け加算を前記第2比率とは逆の比率で行うことにより、前記第2画像に対する第2教師ラベルを生成し、
前記第1教師ラベルと前記第2教師ラベルとの重み付け加算を前記第1比率で行うことにより、前記合成画像に対する合成教師ラベルを生成し、
前記合成画像及び前記合成教師ラベルを用いて機械学習を行うことにより、前記学習モデルを生成する
学習装置。 - 請求項1又は2に記載の学習方法をコンピュータに実行させるためのプログラム。
Priority Applications (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2023523497A JPWO2022250071A1 (ja) | 2021-05-27 | 2022-05-25 | |
EP22811341.1A EP4350612A1 (en) | 2021-05-27 | 2022-05-25 | Learning method, learning device, and program |
US18/512,767 US20240087098A1 (en) | 2021-05-27 | 2023-11-17 | Training method and training device |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US202163193785P | 2021-05-27 | 2021-05-27 | |
US63/193,785 | 2021-05-27 |
Related Child Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US18/512,767 Continuation US20240087098A1 (en) | 2021-05-27 | 2023-11-17 | Training method and training device |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2022250071A1 true WO2022250071A1 (ja) | 2022-12-01 |
Family
ID=84230098
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2022/021329 WO2022250071A1 (ja) | 2021-05-27 | 2022-05-25 | 学習方法、学習装置及びプログラム |
Country Status (4)
Country | Link |
---|---|
US (1) | US20240087098A1 (ja) |
EP (1) | EP4350612A1 (ja) |
JP (1) | JPWO2022250071A1 (ja) |
WO (1) | WO2022250071A1 (ja) |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2019079374A (ja) | 2017-10-26 | 2019-05-23 | 株式会社Preferred Networks | 画像処理システム、画像処理方法および画像処理プログラム |
JP2020038668A (ja) * | 2018-09-05 | 2020-03-12 | 株式会社ストラドビジョン | 自律走行状況で、障害物検出のためのcnn学習用イメージデータセットの生成方法及びコンピューティング装置 |
JP2020173150A (ja) * | 2019-04-10 | 2020-10-22 | 株式会社神戸製鋼所 | 土質判定装置、土質判定用学習モデル生成装置および土質判定方法 |
JP2021013685A (ja) * | 2019-07-16 | 2021-02-12 | 富士フイルム株式会社 | 放射線画像処理装置、方法およびプログラム |
WO2021176605A1 (ja) * | 2020-03-04 | 2021-09-10 | オリンパス株式会社 | 学習データ作成システム及び学習データ作成方法 |
JP2022037804A (ja) * | 2020-08-25 | 2022-03-09 | トヨタ自動車株式会社 | 画像処理装置、画像処理プログラムおよび画像処理方法 |
-
2022
- 2022-05-25 JP JP2023523497A patent/JPWO2022250071A1/ja active Pending
- 2022-05-25 WO PCT/JP2022/021329 patent/WO2022250071A1/ja active Application Filing
- 2022-05-25 EP EP22811341.1A patent/EP4350612A1/en active Pending
-
2023
- 2023-11-17 US US18/512,767 patent/US20240087098A1/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2019079374A (ja) | 2017-10-26 | 2019-05-23 | 株式会社Preferred Networks | 画像処理システム、画像処理方法および画像処理プログラム |
JP2020038668A (ja) * | 2018-09-05 | 2020-03-12 | 株式会社ストラドビジョン | 自律走行状況で、障害物検出のためのcnn学習用イメージデータセットの生成方法及びコンピューティング装置 |
JP2020173150A (ja) * | 2019-04-10 | 2020-10-22 | 株式会社神戸製鋼所 | 土質判定装置、土質判定用学習モデル生成装置および土質判定方法 |
JP2021013685A (ja) * | 2019-07-16 | 2021-02-12 | 富士フイルム株式会社 | 放射線画像処理装置、方法およびプログラム |
WO2021176605A1 (ja) * | 2020-03-04 | 2021-09-10 | オリンパス株式会社 | 学習データ作成システム及び学習データ作成方法 |
JP2022037804A (ja) * | 2020-08-25 | 2022-03-09 | トヨタ自動車株式会社 | 画像処理装置、画像処理プログラムおよび画像処理方法 |
Also Published As
Publication number | Publication date |
---|---|
JPWO2022250071A1 (ja) | 2022-12-01 |
US20240087098A1 (en) | 2024-03-14 |
EP4350612A1 (en) | 2024-04-10 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
JP6441980B2 (ja) | 教師画像を生成する方法、コンピュータおよびプログラム | |
US11244430B2 (en) | Digital image fill | |
Prakash et al. | It GAN do better: GAN-based detection of objects on images with varying quality | |
US20180225817A1 (en) | Optimization of Integrated Circuit Mask Design | |
US11886990B2 (en) | Classification device, classification method, and computer program product | |
JP6612486B1 (ja) | 学習装置、分類装置、学習方法、分類方法、学習プログラム、及び分類プログラム | |
JP2019159836A (ja) | 学習プログラム、学習方法および学習装置 | |
Valle | Hands-On Generative Adversarial Networks with Keras: Your guide to implementing next-generation generative adversarial networks | |
Bach et al. | Analyzing classifiers: Fisher vectors and deep neural networks | |
US10580127B2 (en) | Model generation apparatus, evaluation apparatus, model generation method, evaluation method, and storage medium | |
Lee et al. | TREND: Truncated generalized normal density estimation of Inception embeddings for GAN evaluation | |
CN112750065B (zh) | 载体对象处理和水印嵌入方法、装置及电子设备 | |
WO2022250071A1 (ja) | 学習方法、学習装置及びプログラム | |
US11727109B2 (en) | Identifying adversarial attacks with advanced subset scanning | |
Kareem et al. | Layout pattern synthesis for lithography optimizations | |
JP6622369B1 (ja) | 訓練データを生成する方法、コンピュータおよびプログラム | |
KR102477700B1 (ko) | 대조 학습과 적대적 생성 신경망을 활용하는 이미지 생성 및 편집 방법과 장치 | |
US20210056711A1 (en) | Arithmetic processing apparatus and method | |
JP7264272B2 (ja) | 検出方法、検出プログラム及び検出装置 | |
CN114596209A (zh) | 指纹图像修复方法、***、设备及存储介质 | |
JP6831307B2 (ja) | 解算出装置、解算出方法及び解算出プログラム | |
Vellasques et al. | A dual-purpose memory approach for dynamic particle swarm optimization of recurrent problems | |
Lee et al. | Overfitting control for surface reconstruction | |
CN115083001B (zh) | 基于图像敏感位置定位的对抗补丁生成方法与装置 | |
JP6994572B2 (ja) | データ処理システムおよびデータ処理方法 |
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: 22811341 Country of ref document: EP Kind code of ref document: A1 |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2023523497 Country of ref document: JP |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2022811341 Country of ref document: EP |
|
ENP | Entry into the national phase |
Ref document number: 2022811341 Country of ref document: EP Effective date: 20240102 |