Patent Application
20250182345
The disclosed technology relates to an image generation apparatus, a learning apparatus, an image processing apparatus, an image generation method, a learning method, and an image processing method.
The following techniques are known for machine learning of mathematical models used in image processing. For example, WO2020/017211A describes a medical image learning apparatus including an image acquisition unit that acquires a first image and a second image whose spectral distributions are different from each other, an image processing unit that performs image processing on the first image to generate a third image, and a learning unit that enables a recognizer used for automatic recognition to learn using the first to third images. The image processing unit generates the third image from the first image by performing at least one of image processing for suppressing signals in a band included in the spectral distribution of the first image and having different characteristics from a band corresponding to the second image, or image processing for enhancing signals in a band included in the first image and having characteristics that are the same as or similar to those of the band corresponding to the second image.
WO2020/175446A describes a learning method for performing machine learning of a generative model that estimates, from a first image, a second image including higher resolution image information than the first image. In this method, a first learning image and a second learning image are used as learning data. The first learning image includes first resolution information having lower resolution than the second image. The second learning image includes second resolution information having higher resolution than the first learning image, and serves as a ground truth image corresponding to the first learning image.
In non-destructive inspection of industrial products using radiographic images, attempts have been made to detect defects (flaws) caused inside the products using mathematical models such as a convolutional neural network (CNN) constructed by machine learning. Machine learning typically suffers from insufficient training data relative to the complexity of problems to be solved. So-called “data augmentation” is often performed to secure as many variations as possible from a limited amount of data. For example, when the training data is image data, the accuracy of a mathematical model being trained can be expected to improve by adding, to the training data, processed versions of an original image, such as the result of increasing or decreasing brightness (pixel values), altering contrast, performing scaling processing, changing the blend ratio of color channels in color images, applying rotation, or performing inversion.
In machine learning of mathematical models used for image recognition, in some cases, even a typical data augmentation technique as described above cannot fully compensate for the lack of real data. Particularly in radiographic images used for non-destructive inspection of an industrial product, various types of information, such as undulations due to thickness variations of a target object (product) itself and the shape features of the target object (product) itself, are captured as background along with a defect (flaw) to be detected. As a result, enormous variations of images are to be prepared as training data, but it is not easy to prepare image sets that cover the diversity of background information.
The disclosed technology has been developed in consideration of the above-mentioned points, and an object thereof is to expand the variations of image sets used as training data in machine learning.
An image generation apparatus according to the disclosed technology includes at least one first processor. The at least one first processor is configured to acquire a first image; enhance or extract mutually different frequency components in the first image to generate a plurality of frequency-processed images; perform different computational processing on each of the plurality of frequency-processed images; synthesize respective frequency components of the plurality of frequency-processed images on which the computational processing is performed to generate at least one second image; and output the first image and the at least one second image as training data used for machine learning of a mathematical model that performs predetermined inference on an input image.
The at least one first processor may be configured to perform, as the computational processing, processing for applying mutually different weight coefficients to the plurality of frequency-processed images. The at least one first processor may be configured to generate a first frequency-processed image including relatively low frequency components and a second frequency-processed image including relatively high frequency components; and perform, as the computational processing, processing for applying a relatively small weight coefficient to the first frequency-processed image and applying a relatively large weight coefficient to the second frequency-processed image. The at least one first processor may be configured to perform, as the computational processing, processing on each of the plurality of frequency-processed images to enlarge a difference from an average value of brightness values for each pixel by a different magnification factor.
The at least one first processor may be configured to perform filter processing on the first image to generate a first frequency-processed image; and subtract frequency components of the first frequency-processed image from the first image to generate a second frequency-processed image. The at least one first processor may be configured to perform normalization processing on the first image and the at least one second image to normalize brightness. The first image may be a radiographic image including an image of a specific structural part of a target object in a specific frequency domain, and the mathematical model may be a model that detects the image of the specific structural part included in the radiographic image.
A learning apparatus according to the disclosed technology includes at least one second processor. The at least one second processor is configured to train the mathematical model using, as training data, the first image and the at least one second image provided from the image generation apparatus described above. The at least one second processor may be configured to perform normalization processing on the first image and the at least one second image to normalize brightness.
An image processing apparatus according to the disclosed technology includes at least one third processor. The at least one third processor is configured to output a detection result of an image of a specific structural part of a target object for an input image using the mathematical model trained by the learning apparatus described above. The at least one third processor may be configured to acquire the input image; enhance or extract a specific frequency component in the input image to generate a frequency-processed image; input the input image to the mathematical model to acquire a first inference result; input the frequency-processed image to the mathematical model to acquire a second inference result; and output the detection result by comprehensively evaluating the first inference result and the second inference result.
An image generation method according to the disclosed technology includes processing performed by at least one first processor that an image generation apparatus has. The processing includes acquiring a first image; enhancing or extracting mutually different frequency components in the first image to generate a plurality of frequency-processed images; performing different computational processing on each of the plurality of frequency-processed images; synthesizing respective frequency components of the plurality of frequency-processed images on which the computational processing is performed to generate at least one second image; and outputting the first image and the second image as training data used for machine learning of a mathematical model that performs predetermined inference on an input image.
A learning method according to the disclosed technology includes processing performed by at least one second processor that a learning apparatus has. The processing includes training the mathematical model using, as training data, the first image and the at least one second image provided using the image generation method described above.
An image processing method according to the disclosed technology includes processing performed by at least one third processor that an image processing apparatus has. The processing includes outputting a detection result of an image of a specific structural part of a target object for an input image using the mathematical model trained by the learning method described above.
The disclosed technology makes it possible to expand the variations of image sets used as training data in machine learning.
Exemplary embodiments according to the technique of the present disclosure will be described in detail based on the following figures, wherein:
Hereinafter, an example of an embodiment of the disclosed technology will be described with reference to the drawings. In the drawings, the same or equivalent components and parts are denoted by the same reference numerals, and redundant descriptions will be omitted.
Hereinafter, details of the image generation apparatus 10, the learning apparatus 20, and the image processing apparatus 30 will be sequentially described. In the following description, by way of example, an image handled in the image processing system 1 is a radiographic image acquired in non-destructive inspection of an industrial product. The radiographic image may include an image of a defect (flaw) caused inside the target object. In the following description, by way of example, the mathematical model handled in the image processing system 1 is a defect detection model that detects a defect (flaw) included in the radiographic image, and the image processing apparatus 30 detects the defect (flaw) included in the radiographic image using the defect detection model. In the following description, furthermore, by way of example, the image generation apparatus 10, the learning apparatus 20, and the image processing apparatus 30 are configured as separate computers.
The non-volatile memory 103 is a non-volatile storage medium such as a hard disk or a flash memory. The non-volatile memory 103 stores an image generation program 110. The RAM 102 is a work memory for the CPU 101 to execute processing. The CPU 101 loads the image generation program 110 stored in the non-volatile memory 103 into the RAM 102, and executes processing in accordance with the image generation program 110. The CPU 101 is an example of a “first processor” in the disclosed technology.
The first image acquisition unit 11 acquires a first image 41. The first image 41 is a radiographic image acquired in non-destructive inspection of an industrial product. The first image 41 includes an image of a defect (flaw) caused inside a target object (product). A typical image can be interpreted by decomposing it into spatial frequency components. The image of the defect (flaw) is in a relatively high frequency domain of the first image 41. For example, when the first image 41 has a resolution of 25 to 100 μm/px, the image of the defect (flaw) has 1 to 2 px at its smallest. The first image 41 may be an original image or an image obtained by performing preprocessing other than frequency processing described below, such as upscaling, downscaling, rotation, and tone processing, on the original image. The first image acquisition unit 11 may perform the preprocessing.
The frequency processing unit 12 performs first frequency processing 51 on the first image 41 to generate a first frequency-processed image 42. The first frequency processing 51 is processing for extracting low-frequency components from the first image 41 or processing for enhancing the low-frequency components in the first image 41. The first frequency processing 51 may be, for example, filter processing using a low-pass filter. The first frequency processing 51 is known as a noise removal method. This processing can also be implemented using a median filter, an averaging filter, a Gaussian filter, or a bilateral filter. In the first frequency-processed image 42, most of the image of the defect (flaw) included in the first image 41 are removed, whereas the image of the background such as the undulations due to thickness variations of the target object (product) itself or the shape features of the target object (product) itself remains unremoved.
The frequency processing unit 12 further performs second frequency processing 52 on the first image 41 to generate a second frequency-processed image 43. The second frequency processing 52 is processing for extracting high-frequency components from the first image 41 or processing for enhancing the high-frequency components in the first image 41. The second frequency processing 52 may be, for example, filter processing using a high-pass filter. The frequency processing unit 12 can also generate the second frequency-processed image 43 by subtracting the frequency components of the first frequency-processed image 42 from the first image 41. In the second frequency-processed image 43, most of the image of the background such as undulations due to thickness variations of the target object (product) itself or the shape features of the target object (product) itself are removed, whereas the image of the defect (flaw) included in the first image 41 remains unremoved. That is, the image of the defect (flaw) is more clearly depicted in the second frequency-processed image 43.
The computational processing unit 13 performs first computational processing 53 on the first frequency-processed image 42 and performs second computational processing 54 different from the first computational processing 53 on the second frequency-processed image 43. For example, the computational processing unit 13 may perform, as the first computational processing 53, processing for applying a relatively small weight coefficient w1 (≥0) to the first frequency-processed image 42, and may perform, as the second computational processing 54, processing for applying a relatively large weight coefficient w2 (>w1) to the second frequency-processed image 43. That is, when each pixel of the first frequency-processed image 42 has a brightness value represented by P1x,y, the computational processing unit 13 outputs w1×P1x,y as a first frequency-processed image 42A after the first computational processing. When each pixel of the second frequency-processed image 43 has a brightness value represented by P2x,y, the computational processing unit 13 outputs w2× P2x,y as a second frequency-processed image 43A after the second computational processing. The weight coefficient w1 to be applied to the first frequency-processed image 42, which mainly includes low-frequency components, may be set to zero. The computational processing unit 13 may change the combination of the weight coefficients w1 and w2 to generate a plurality of image pairs, each being a pair of the first frequency-processed image 42A and the second frequency-processed image 43A after the computational processing.
Instead of or in addition to the processing for applying weight coefficients having mutually different magnitudes to the first frequency-processed image 42 and the second frequency-processed image 43, the computational processing unit 13 may perform, as computational processing, processing for performing contrast enhancement on the first frequency-processed image 42 and the second frequency-processed image 43 at different intensities. Specifically, the computational processing unit 13 may perform, as the first computational processing 53, processing on the first frequency-processed image 42 to enlarge the difference from an average value of brightness values for each pixel by a magnification factor m1 (>0), and may perform, as the second computational processing 54, processing on the second frequency-processed image 43 to enlarge the difference from an average value of brightness values for each pixel by a magnification factor m2 (>m1) larger than m1. The computational processing unit 13 may change the combination of the magnification factors m1 and m2 to generate a plurality of image pairs, each being a pair of the first frequency-processed image 42A and the second frequency-processed image 43A after the computational processing.
The synthesis processing unit 14 performs synthesis processing 55 to synthesize the frequency components of the first frequency-processed image 42A on which the first computational processing 53 is performed and the frequency components of the second frequency-processed image 43A on which the second computational processing 54 is performed, and generates at least one second image 44. For example, when the computational processing unit 13 performs processing for applying the weight coefficients w1 and w2 to the first frequency-processed image 42 and the second frequency-processed image 43, respectively, the synthesis processing 55 generates the second image 44 by performing weighted addition processing on the first frequency-processed image 42 and the second frequency-processed image 43. Setting w1<w2 makes it possible to obtain the second image 44 in which the image of the defect (flaw) present in the high-frequency domain is further enhanced. Setting w1 to zero causes the second frequency-processed image 43 to become substantially the second image 44, and the image of the defect (flaw) is depicted more clearly in the second image 44. When a plurality of image pairs, each being a pair of the first frequency-processed image 42A and the second frequency-processed image 43A after the computational processing, are generated, the synthesis processing unit 14 performs the synthesis processing 55 on each of the plurality of image pairs. As a result, a plurality of second images 44 are generated from the single first image 41.
The image output unit 15 outputs the first image 41 acquired by the first image acquisition unit 11 and at least one second image 44 generated based on the first image 41 as training data used for machine learning of a mathematical model that performs predetermined inference on an input image. The image output unit 15 may perform normalization processing on the first image 41 and the second image 44 to normalize the brightness of each pixel before outputting these images. The normalization processing is expressed by, for example, equation (1) below. In equation (1), Xnorm is a brightness value after the normalization processing. X is the brightness value before normalization. Xmax is the maximum value of brightness (for example, 255). Xmin is the minimum value of brightness (for example, 0). The normalization processing makes it possible to suppress a significant divergence in brightness distribution between the first image 41 and the second image 44.
In step S1, the first image acquisition unit 11 acquires the first image 41. The first image 41 may be an original image or an image obtained by performing preprocessing other than frequency processing, such as upscaling, downscaling, rotation, and tone processing, on the original image. The first image acquisition unit 11 may perform the preprocessing.
In step S2, the frequency processing unit 12 enhances or extracts mutually different frequency components in the first image 41 acquired in step S1 to generate the first frequency-processed image 42 and the second frequency-processed image 43. The first frequency-processed image 42 mainly includes low-frequency components.
In step S3, the computational processing unit 13 performs different computational processing on each of the first frequency-processed image 42 and the second frequency-processed image 43 generated in step S2. For example, the computational processing unit 13 performs, as the first computational processing 53, processing for applying a relatively small weight coefficient w1 (≥0) to the first frequency-processed image 42, and performs, as the second computational processing 54, processing for applying a relatively large weight coefficient w2 (>w1) to the second frequency-processed image 43. The computational processing unit 13 may change the combination of the weight coefficients w1 and w2 to generate a plurality of image pairs, each being a pair of the first frequency-processed image 42A and the second frequency-processed image 43A after the computational processing.
In step S4, the synthesis processing unit 14 synthesizes the respective frequency components of the first frequency-processed image 42A and the second frequency-processed image 43A on which the computational processing is performed in step S3 to generate at least one second image 44. When a plurality of image sets, each being a set of the first frequency-processed image 42A and the second frequency-processed image 43A after the computational processing, are generated, the synthesis processing unit 14 performs the synthesis processing 55 on each of the image sets. As a result, a plurality of second images 44 are generated.
In step S5, the image output unit 15 outputs the first image 41 acquired in step S1 and the at least one second image 44 generated in step S4 as training data used for machine learning of a mathematical model that performs predetermined inference on an input image. The image output unit 15 may perform normalization processing on the first image 41 and the second image 44 to normalize the brightness of each pixel before outputting these images.
As described above, the image generation apparatus 10 according to the embodiment of the disclosed technology acquires the first image 41, and enhances or extracts mutually different frequency components in the first image 41 to generate a plurality of frequency-processed images. The image generation apparatus 10 performs different computational processing on each of the plurality of frequency-processed images, and synthesizes the respective frequency components of the plurality of frequency-processed images on which the computational processing is performed to generate at least one second image 44. The image generation apparatus 10 outputs the first image 41 and the second image 44 as training data used for machine learning of a mathematical model that performs predetermined inference on an input image.
In machine learning of mathematical models for image recognition, even when a typical data augmentation technique such as increasing or decreasing brightness (pixel values), altering contrast, performing scaling processing, changing the blend ratio of color channels in color images, applying rotation, or performing inversion is used for an original image, the lack of real data cannot be fully compensated for. Particularly in radiographic images used for non-destructive inspection of an industrial product, various types of information, such as undulations due to thickness variations of a target object (product) itself and the shape features of the target object (product) itself, are captured as background along with a defect (flaw) to be detected. As a result, enormous variations of images are to be prepared as training data, but it is not easy to prepare image sets that cover the diversity of background information.
In the image generation apparatus 10 according to the embodiment of the disclosed technology, the second image 44 is generated based on the first image 41, which is the original image, and both the first image 41 and the second image 44 are provided as training data. Thus, the training data can be augmented.
The second image 44 is an image obtained by performing different computational processing on each of the first frequency-processed image 42 and the second frequency-processed image 43 and synthesizing the frequency components of the first frequency-processed image 42 and the second frequency-processed image 43. For example, the combination of the weight coefficients used in computational processing is changed to make variations to the computational processing. Accordingly, a plurality of second images 44 can be generated. As a result, further data augmentation can be implemented, allowing for diversity in the training data. As described above, the image generation apparatus 10 according to the embodiment of the disclosed technology makes it possible to expand the variations of image sets used as training data in machine learning.
In addition, since the second image 44 is an image obtained by synthesizing the frequency components of the first frequency-processed image 42 and the second frequency-processed image 43, the appearance of the image of the defect (flaw) present in the high-frequency domain and the appearance of the image of the background present in the low-frequency domain can be adjusted independently. For example, computational processing is performed such that the relatively small weight coefficient w1 is applied to the first frequency-processed image 42 mainly including low-frequency components and the relatively large weight coefficient w2 is applied to the second frequency-processed image 43 mainly including high-frequency components, and then the frequency components of these images are synthesized. Accordingly, the second image 44 in which the image of the defect (flaw) present in the high-frequency domain is more enhanced can be obtained. Further, setting w1 to zero causes the second frequency-processed image 43 to become substantially the second image 44, and the image of the defect (flaw) is depicted more clearly in the second image 44. In machine learning of a defect detection model that detects a defect (flaw) in an input image, images in which the defect (flaw) to be detected is enhanced are used as training data, thereby allowing for more clearly instructing the defect detection model on the image characteristics of the defect (flaw). As a result, the defect detection accuracy in the defect detection model can be improved.
While the present embodiment presents an example in which the frequency processing unit 12 generates two frequency-processed images (the first frequency-processed image 42 and the second frequency-processed image 43), the disclosed technology is not limited to this configuration. The frequency processing unit 12 may generate three or more frequency-processed images by enhancing or extracting mutually different frequency components in the first image 41. This configuration allows for further increasing the variations of computational processing performed in the computational processing unit 13, enabling the generation of more diverse second images 44.
The non-volatile memory 203 stores a learning program 210, a defect detection model 220, and training data 60. The defect detection model 220 is a mathematical model based on a CNN constructed to predict class labels pixel by pixel for an input image. The defect detection model 220 may be, for example, a model that applies a known encoder-decoder model (ED-CNN: Encoder-Decoder Convolutional Neural Network). The encoder-decoder model is a model constituted by an encoder that extracts features from an image using convolutional layers and a decoder that outputs a probability map based on the extracted features. The probability map is the result of deriving the probability that each pixel of the input image belongs to a certain class, on a per-pixel basis. The defect detection model 220 is constructed by machine learning using a plurality of radiographic images with ground truth labels as training data. The training data 60 is a dataset that treats the first image 41 and the second image 44 supplied from the image generation apparatus 10 and ground truth labels 45 and 46 assigned to these images as one unit. A plurality of pieces of training data 60 are stored in the non-volatile memory 203. The training data 60 is used for machine learning of the defect detection model 220.
Instead of the image output unit 15 of the image generation apparatus 10, the training data acquisition unit 21 or the learning unit 22 of the learning apparatus 20 may perform the normalization processing on the first image 41 and the second image 44, which are the training data 60, to normalize the brightness of each pixel. The normalization processing makes it possible to prevent a significant divergence in brightness distribution between the first image 41 and the second image 44. If a significant divergence occurs in brightness distribution between the first image 41 and the second image 44, the defect detection model 220 may unintentionally distinguish between the first image 41 and the second image 44, and may fail to extract features of a defect (flaw) present in common in these images. The normalization processing performed on the first image 41 and the second image 44 can make it less likely that the defect detection model 220 unintentionally distinguishes between the first image 41 and the second image 44.
As described above, the learning apparatus 20 according to the embodiment of the disclosed technology trains the defect detection model 220 using the first image 41 and the second image 44 generated by the image generation apparatus 10 as training data. In the learning apparatus 20 according to the present embodiment, the defect detection model 220 is trained using the training data 60 augmented by the second image 44. Thus, the defect detection accuracy in the defect detection model 220 can be improved.
The processes for generating the second image 44 may be performed in the learning apparatus 20 instead of the image generation apparatus 10. Further, the processes for generating the second image 44 may be performed in an input layer of the defect detection model 220.
The input image acquisition unit 31 acquires an input image 71 that is input as a processing target of the image processing apparatus 30. The input image 71 is a radiographic image similar to the first image 41. The input image 71 may be an image obtained by performing preprocessing other than frequency processing, such as upscaling, downscaling, rotation, and tone processing, on an original image. The input image acquisition unit 31 may perform the preprocessing.
The frequency processing unit 32 performs frequency processing 82 on the input image 71 to generate a frequency-processed image 72. The frequency processing 82 may be, for example, processing for extracting high-frequency components from the input image 71 or processing for enhancing the high-frequency components in the input image 71. The frequency processing 82 may be processing for extracting or enhancing frequency components that are the same as or different from the frequency components extracted or enhanced in the second frequency processing 52 performed in the image generation apparatus 10. The frequency processing unit 32 may generate a plurality of frequency-processed images 72 by enhancing or extracting mutually different frequency components in the input image 71.
The inference result acquisition unit 33 inputs the input image 71 to the trained defect detection model 320 to acquire a first inference result 73. The inference result acquisition unit 33 further inputs the frequency-processed image 72 to the trained defect detection model 320 to acquire a second inference result 74. When a plurality of frequency-processed images 72 are generated, the inference result acquisition unit 33 acquires a plurality of second inference results 74 each corresponding to one of the plurality of frequency-processed images 72. For example, the first inference result 73 and the second inference result 74 may be each a probability map that is the result of deriving, for each pixel, the probability that the pixel indicates a defect (flaw) in the corresponding image.
The detection result output unit 34 outputs a detection result 75 of the image of the defect (flaw) for the input image 71 by comprehensively evaluating the first inference result 73 and the second inference result 74. For example, the detection result output unit 34 may calculate an average value of the probability map as the first inference result 73 and the probability map as the second inference result 74 to generate an integrated probability map. Alternatively, the detection result output unit 34 may calculate a weighted average value of the probability map as the first inference result 73 and the probability map as the second inference result 74 to generate an integrated probability map. Alternatively, the detection result output unit 34 may generate an integrated probability map by applying a maximum value to the probability map as the first inference result 73 and the probability map as the second inference result 74. The detection result output unit 34 may derive the detection result 75 by performing, for example, threshold value processing on the integrated probability map.
In step S31, the input image acquisition unit 31 acquires the input image 71 that is input as a processing target of the image processing apparatus 30. The input image 71 is a radiographic image similar to the first image 41. In step S32, the frequency processing unit 32 enhances or extracts specific frequency components in the input image 71 acquired in step S31 to generate the frequency-processed image 72.
In step S33, the inference result acquisition unit 33 inputs the input image 71 acquired in step S31 to the trained defect detection model 320 to acquire the first inference result 73. In step S34, the inference result acquisition unit 33 inputs the frequency-processed image 72 generated in step S32 to the trained defect detection model 320 to acquire the second inference result 74. In step S35, the detection result output unit 34 outputs the detection result 75 of the image of the defect (flaw) for the input image 71 acquired in step S31 by comprehensively evaluating the first inference result 73 acquired in step S33 and the second inference result 74 acquired in step S34.
As described above, the image processing apparatus 30 according to the embodiment of the disclosed technology outputs a detection result of the image of the defect (flaw) for the input image 71 using the defect detection model 320 trained by the learning apparatus 20. In the image processing apparatus 30 according to the present embodiment, the defect (flaw) is detected using the defect detection model 320 trained using the training data 60 augmented by the second image 44. Thus, the defect detection accuracy can be improved.
In addition, since the detection result 75 is output by comprehensively evaluating the first inference result 73 and the second inference result 74, it is possible to appropriately balance the detection accuracy and the detection sensitivity. The detection result 75 of the image of the defect (flaw) may be output based on only one of the first inference result 73 and the second inference result 74. When only the first inference result 73 is used, the frequency processing unit 32 and the frequency-processed image 72 are not necessary.
The foregoing description provides an example in which the image generation apparatus 10, the learning apparatus 20, and the image processing apparatus 30 are configured as separate computers. However, the image generation apparatus 10, the learning apparatus 20, and the image processing apparatus 30 may be configured as one or two computers. For example, the image generation apparatus 10 and the learning apparatus 20 may be configured as the same computer. Alternatively, the learning apparatus 20 and the image processing apparatus 30 may be configured as the same computer. Alternatively, the image generation apparatus 10, the learning apparatus 20, and the image processing apparatus 30 may be configured as the same computer.
Furthermore, the foregoing description provides an example in which the image handled in the image processing system 1 is a radiographic image acquired in non-destructive inspection of an industrial product. However, the disclosed technology is not limited to this configuration. The disclosed technology can be applied to any image (for example, optical images, magnetic resonance imaging (MRI) images, scanning electron microscope (SEM) images, ultrasound images, etc.) that includes an image of a specific structural part of a target object in a specific frequency domain. Furthermore, the foregoing description provides an example in which the mathematical model to be trained is a defect detection model that detects a defect (flaw) included in a radiographic image. However, the disclosed technology is not limited to this configuration. The disclosed technology can be applied to a mathematical model that detects or recognizes a specific structural part of a target object present in a specific frequency domain in an image.
In the embodiment described above, for example, a hardware structure of a processing unit that executes various processes such as the first image acquisition unit 11, the frequency processing unit 12, the computational processing unit 13, the synthesis processing unit 14, the image output unit 15, the training data acquisition unit 21, the learning unit 22, the input image acquisition unit 31, the frequency processing unit 32, the inference result acquisition unit 33, and the detection result output unit 34 can be implemented using the following various processors. As described above, the various processors include, in addition to a CPU and a GPU that are general-purpose processors serving as various processing units by executing software (programs), a programmable logic device (PLD) that is a processor whose circuit configuration can be changed after manufacturing, such as an FPGA, a dedicated electrical circuit that is a processor having a circuit configuration designed exclusively for executing specific processing, such as an application specific integrated circuit (ASIC), and the like.
One processing unit may be constituted by one of these various processors, or may be constituted by a combination of two or more processors of the same type or different types (for example, a combination of a plurality of FPGAs or a combination of a CPU and an FPGA). Additionally, a plurality of processing units may be configured using a single processor.
Examples of configuring a plurality of processing units using a single processor include, first, a form in which, as represented by computers such as a client and a server, one processor is configured by a combination of one or more CPUs and software and this processor functions as a plurality of processing units. The examples include, second, a form in which, as represented by a system on chip (SoC) or similar technologies, a processor is used that implements the functions of the entire system, including a plurality of processing units, on a single integrated circuit (IC) chip. In this way, the various processing units are configured using one or more of the above-described various processors as a hardware structure. More specifically, the hardware structure of these various processors may be an electrical circuit (circuitry) in which circuit elements such as semiconductor elements are combined.
Moreover, the above embodiment has provided a description of a configuration in which various programs are stored (installed) in advance in a non-volatile memory, but is not limited to this configuration. The various programs may be provided in a form recorded on recording media such as a compact disc read only memory (CD-ROM), a digital versatile disc read only memory (DVD-ROM), and a universal serial bus (USB) memory. The various programs may be downloaded from an external apparatus via a network. In other words, the programs (i.e., program products) described in the present embodiment may be provided not only on a recording medium but also in a form distributed from an external computer.
In relation to the embodiment described above, the following appendices are further disclosed.
An image generation apparatus including at least one first processor,
The image generation apparatus according to appendix 1, wherein
The image generation apparatus according to appendix 1 or 2, wherein
The image generation apparatus according to any one of appendices 1 to 3, wherein
The image generation apparatus according to any one of appendices 1 to 4, wherein
The image generation apparatus according to any one of appendices 1 to 5, wherein
The image generation apparatus according to any one of appendices 1 to 6, wherein
A learning apparatus including at least one second processor,
The learning apparatus according to appendix 8, wherein the at least one second processor is configured to
An image processing apparatus including at least one third processor,
The image processing apparatus according to appendix 10, wherein
An image generation method including processing performed by at least one first processor that an image generation apparatus has, the processing including:
A learning method including processing performed by at least one second processor that a learning apparatus has, the processing comprising
An image processing method including processing performed by at least one third processor that an image processing apparatus has, the processing including
The disclosure of JP2022-146397 filed on Sep. 14, 2022 is incorporated herein by reference in its entirety. All the documents, patent applications, and technical standards described in this specification are incorporated herein by reference to the same extent as if each individual document, patent application, and technical standard were specifically and individually indicated to be incorporated by reference.
| Number | Date | Country | Kind |
|---|---|---|---|
| 2022-146397 | Sep 2022 | JP | national |
This application is a continuation application of International Application No. PCT/JP2023/028554, filed Aug. 4, 2023, the disclosure of which is incorporated herein by reference in its entirety. Further, this application claims priority from Japanese Patent Application No. 2022-146397, filed on Sep. 14, 2022, the disclosure of which is incorporated herein by reference in its entirety.
| Number | Date | Country | |
|---|---|---|---|
| Parent | PCT/JP2023/028554 | Aug 2023 | WO |
| Child | 19044626 | US |
