Data processing device, data processing method, and program

Abstract

Provided are a data processing device, a data processing method, and a program that generate, by data expansion, a supervision data set that can be efficiently learned in a learning model. A data processing device (10) includes a processor (14) that generates an inverted image in which pixel values of a single-channel image constituting a first supervision data set are inverted, as learning data of a second supervision data set obtained by expanding the first supervision data set, and as learning data of a third supervision data set obtained by expanding the second supervision data set, based on a correct answer tag of the single-channel image constituting the first supervision data set, an inverted label corresponding to an inverted image in which a pixel value is determined on the basis of a physical quantity sensed by a light-receiving element at the time of imaging or a pixel value is determined by reversibly converting the physical quantity is generated as a correct answer label constituting the second supervision data set.

Description

Technical Field

[0001] The present invention relates to a data processing device, a data processing method and a program, and in particular to a data processing device, a data processing method and a program for generating a supervision data set. Background Art

[0002] Generally speaking, in machine learning, the lack of supervised data sets becomes a problem due to the complexity of the problems to be answered. Therefore, in order to ensure a supervised data set with as many changes as possible from a relatively small number of supervised data sets, a supervised data set is newly generated by expanding the data from the original supervised data set. Here, the supervised data set consists of learning data for the learning model to learn and the correct answer label indicating the correct answer to the learning data.

[0003] For example, Patent Document 1 describes a technique for performing data expansion on transmission X-ray images to generate data sets of hundreds of thousands of images.

[0004] Previous technical literature

[0005] Patent Literature

[0006] Patent Document 1: Japanese Patent Application Publication No. 2021-110753 Summary of the invention

[0007] Technical issues to be solved by the invention

[0008] One embodiment of the technology involved in the present invention provides a data processing device, a data processing method, and a program that generate a supervised data set that can perform effective learning in a learning model through data expansion.

[0009] Means for solving technical problems

[0010] A first embodiment of the present invention is a data processing device, comprising a processor, wherein the processor generates an inverted image by inverting the pixel values ​​of a single-channel image constituting a first supervised data set as learning data of a second supervised data set obtained by expanding the first supervised data set, and generates an inverted label corresponding to the inverted image based on the correct answer label of the single-channel image constituting the first supervised data set as the correct answer label constituting the second supervised data set, wherein the single-channel image determines the pixel value according to the physical quantity sensed by the photosensitive element during photography, or determines the pixel value by performing a reversible transformation on the physical quantity.

[0011] The data processing device of the second mode of the present invention is preferably based on the first mode, wherein a single-channel image is an image captured by a digital detector array (DDA) that receives radiation transmitted through a subject, or an imaging plate (IP) that receives radiation transmitted through a subject is a computed radiography (CR) image that outputs a light-receiving signal in the form of a digital value through a reading device.

[0012] The data processing device of the third aspect of the present invention is preferably based on the first or second aspect, wherein the single-channel image is an image based on lens-free imaging.

[0013] The data processing device of the fourth mode of the present invention is preferably based on any one of the first to third modes, wherein each pixel of the single-channel image has a digital value proportional to the amount of light received in the corresponding photosensitive element, or a digital value correlated with the amount of light received.

[0014] The data processing device of the fifth aspect of the present invention is preferably based on any one of the first to fourth aspects, wherein each pixel of the single-channel image has a digital value correlated with a physical quantity at each point of the corresponding object.

[0015] The data processing device of the sixth aspect of the present invention is preferably based on any one of the first to fifth aspects, wherein, when the first supervisory data set is expanded, the first supervisory data set selects supervisory data that does not produce contradictions in the reversed labels.

[0016] The data processing device of the seventh aspect of the present invention is preferably based on any one of the first to sixth aspects, wherein, when the first supervised data set is expanded, the first supervised data set selects supervised data whose inverted labels are effective for learning.

[0017] The data processing device of the eighth mode of the present invention is preferably based on any one of the first to seventh modes, wherein the processor inverts the pixel values ​​of a portion of the single-channel image to generate an inverted image, and generates a correct answer label constituting a second supervised data set based on the correct answer label of the single-channel image corresponding to the portion of the image.

[0018] The data processing device according to a ninth aspect of the present invention is preferably any one of the first to eighth aspects, wherein the processor normalizes or standardizes the learning data.

[0019] The data processing device of the tenth aspect of the present invention is preferably based on any one of the first to ninth aspects, wherein the processor edits the class design of the correct answer labels constituting the second supervised data set corresponding to setting the inverted image as learning data.

[0020] The data processing device of the eleventh aspect of the present invention is preferably based on any one of the first to tenth aspects, wherein the reversible transformation is at least one of a linear transformation, a logarithmic transformation, and a nonlinear transformation using a pixel value correspondence table.

[0021] In the data processing method of the twelfth embodiment of the present invention, the processor performs the following steps: generating an inverted image that inverts the pixel values ​​of a single-channel image constituting a first supervised data set, as learning data of a second supervised data set obtained by expanding the first supervised data set, wherein the pixel value of the single-channel image is determined according to a physical quantity sensed by a light-receiving element during photography, or the pixel value is determined by performing a reversible transformation on the physical quantity; and generating an inverted label corresponding to the inverted image based on the correct answer label of the single-channel image constituting the first supervised data set, as the correct answer label constituting the second supervised data set.

[0022] The program of the thirteenth mode of the present invention enables the processor to execute a data processing method, which includes the following steps: generating an inverted image that inverts the pixel values ​​of a single-channel image constituting a first supervised data set, as learning data for a second supervised data set obtained by expanding the first supervised data set, wherein the pixel value of the single-channel image is determined based on a physical quantity sensed by a light-receiving element during photography, or by performing a reversible transformation on the physical quantity; and generating an inverted label corresponding to the inverted image based on the correct answer label of the single-channel image constituting the first supervised data set, as the correct answer label constituting the second supervised data set.

[0023] Effects of the Invention

[0024] According to the present invention, a supervised data set that can be used to effectively learn a learning model can be generated based on the original supervised data set through data expansion. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a block diagram showing a hardware configuration example of a data processing device.

[0026] Figure 2 This is a diagram showing a functional block of a function F implemented by a processor executing a data processing program stored in a memory.

[0027] Figure 3 This is a schematic diagram for explaining the photography of a single-channel image.

[0028] Figure 4 is a graph representing a single-channel image.

[0029] Figure 5 This is a diagram for explaining the class rewriting process of the correct answer label.

[0030] Figure 6This is a diagram for explaining the class rewriting process of the correct answer label.

[0031] Figure 7 A diagram illustrating the class design for editing the correct answer labels (reversing labels) of the second supervised dataset.

[0032] Figure 8 The diagram illustrates specific examples of the first supervised data set and the second supervised data set.

[0033] Fig. 9 This is a flowchart that explains the data processing method.

[0034] Fig.10 It is a graph showing the pixel values ​​that make up an image.

[0035] Fig.11 This is a diagram for explaining the first supervised data set and the second supervised data set according to the third embodiment.

[0036] Fig.12 This is a block diagram schematically showing a configuration example of a photography system. DETAILED DESCRIPTION

[0037] Hereinafter, preferred embodiments of the present invention will be described with reference to the accompanying drawings.

[0038] <First embodiment>

[0039] <Data processing device>

[0040] Figure 1 This is a block diagram showing a hardware configuration example of the data processing device of the present invention.

[0041] The data processing device 10 includes a processor 14, a memory 16 formed of a non-transitory tangible object, and an input / output interface 12. The computer functioning as the data processing device 10 may be a workstation, a personal computer, a tablet terminal, or a server.

[0042] The processor 14 is composed of a CPU (Central Processing Unit). Alternatively, the processor 14 may be composed of a GPU (Graphics Processing Unit). The processor 14 is connected to the memory 16 and the input / output interface 12 via the bus 8 .

[0043] Information is input to the data processing device 10 via the input / output interface 12. Information is also output via the input / output interface 12. For example, a supervisory data set DB (database) 22 is connected to the data processing device 10 via the input / output interface 12, and the first supervisory data set and the second supervisory data set are input to the data processing device 10.

[0044] The memory 16 includes a memory as a main storage device and a storage device as an auxiliary storage device. The memory 16 may be, for example, a semiconductor memory, a hard disk drive (HDD) device, a solid state drive (SSD) device, or a combination of these.

[0045] The memory 16 stores a program for controlling the data processing device 10. For example, a data processing program 18 for executing a data processing method is stored. In addition, the memory 16 stores a general program for operating the data processing device 10.

[0046] exist Figure 1 In the illustrated case, an external supervisory data set DB 22 is connected to the data processing device 10. Alternatively, the supervisory data set DB may also be configured in the data processing device 10.

[0047] The supervised dataset DB22 includes a first supervised dataset DB24 and a second supervised dataset DB26. The first supervised dataset and the second supervised dataset are used to learn the model 40 ( Figure 8 ) and the correct answer label of the learning data. A plurality of first supervised data sets are stored in the first supervised data set DB24. For example, the first supervised data set is composed of the learning data of the camera system 500 ( Fig.12 ) and the defects read out using the single-channel image (equivalent to the correct answer label). A plurality of second supervisory data sets are stored in the second supervisory data set DB26. The second supervisory data is composed of an inverted image of the single-channel image (equivalent to the learning data) and an inverted label (equivalent to the correct answer label) obtained by rewriting the correct answer label. The second supervisory data set is a supervisory data set obtained by data expansion of the first supervisory data, and is generated by the data processing device 10.

[0048] In addition, Figure 1 , an example is shown in which the first supervised dataset DB 24 and the second supervised dataset DB 26 are included in the supervised dataset DB 22, but the first supervised dataset DB 24 and the second supervised dataset DB 26 may be separately provided.

[0049] Figure 21 is a diagram showing a functional block of a function F implemented by the processor 14 executing the data processing program 18 stored in the memory 16 .

[0050] Function F implemented by the processor 14 executing the data processing program 18 includes a single-channel image acquisition unit 30, a correct answer label acquisition unit 32, a reverse image generation unit 34, and a reverse label generation unit 36. By implementing the above functions F, the first supervision data set is data expanded to generate a second supervision data set.

[0051] The single-channel image acquisition unit 30 acquires the learning data of the first supervised data set, that is, the single-channel image, from the supervised data set DB22. Here, the single-channel image refers to an image whose pixel value is determined according to the physical quantity sensed by the light-receiving element during photography, and an image whose pixel value is determined by reversibly transforming the physical quantity. In addition, the specific reversible transformation here is at least one of a linear transformation, a logarithmic transformation, and a nonlinear transformation using a pixel value correspondence table. In addition, the single-channel image is an image in which only the transformation pursued by the inverse transformation is implemented on the original physical phenomenon. That is, the single-channel image can also be said to be an image that has not been irreversibly transformed. As a specific example of a single-channel image, a Raw image can be cited. Here, the Raw image is composed of undeveloped image data output from the imaging element as a light-receiving element, and the amount of light received by each light-receiving element and the signal value output by the light-receiving element have a linear relationship. In addition, as another specific example of a single-channel image, for example, an image obtained through a red filter and a microscope, etc., is a black and white image that captures the amount of light in a wavelength of a specific spectrum (color). In addition, a single-channel image is an image captured by a digital detector array (DDA) receiving radiation transmitted through a subject, or a computed radiography (CR) image in which an imaging plate (IP) receiving radiation transmitted through a subject outputs a light-receiving signal as a digital value through a reading device. In addition, a single-channel image is an image based on lensless imaging.

[0052] Figure 3 This is a schematic diagram for explaining the photography of a single-channel image.

[0053] exist Figure 3 , there are shown a radiation source 50, an object OBJ, and a detector 54. The detector 54 is composed of a light receiving panel, which is an example of a light receiving element.

[0054] The radiation emitted from the radiation source 50 is irradiated toward the object OBJ. The radiation having passed through the object OBJ is received by the detector 54. The detector 54 outputs a single-channel image 56 of the object OBJ.

[0055] Figure 454 is a diagram showing a single-channel image 56 of the object OBJ acquired by the detector 54. The pixel Q of the single-channel image 56 has a value determined according to the intensity of the radiation that has passed through the object on the straight line L connecting the detector 54 and the radiation source 50, and corresponds to the area P of the object that is penetrated by the straight line L. Therefore, each pixel of the single-channel image has a digital value proportional to the amount of light received in the corresponding detector 54, or a digital value correlated with the amount of light received. In addition, each pixel of the single-channel image has a digital value correlated with the physical quantity at each point (e.g., point P) of the corresponding object OBJ (object).

[0056] Next, for reference, a specific example of the present invention that is not a single-channel image is described. Since the characteristics of the original physical quantity are unknown, an image composed of pixel values ​​read through an external device whose two-dimensional characteristics are unknown, such as pixel values ​​read through a low-bit monitor, is not a single-channel image. In addition, an image that has undergone a LUT (Look Up Table) transformation, such as producing a white or black color, is not a single-channel image. In addition, an image that has undergone a transformation such as a combination of multiple values ​​into one value, which becomes a "many-to-one" transformation, is also not a single-channel image because the amount of information decreases. An image that has undergone nonlinear processing such as gamma correction is also not a single-channel image because it becomes asymmetric when the inversion processing is performed by the inversion image generation unit 34. However, if the inverse transformation of the correction can be defined theoretically and the information is not degraded by the transformation, it can be set as a single-channel image. However, even if the theoretical function can be calculated, if the correspondence is not achieved due to the balance with the quantization, it cannot be set as a single-channel image.

[0057] The correct answer label acquisition unit 32 ( Figure 2 ) Obtain the correct answer label of the first supervised dataset from the supervised dataset DB22.

[0058] Here, the correct answer label has information related to the correct answer when the learning model 40 is trained using the single-channel image as learning data. For example, the correct answer label has information on a defect read out from the single-channel image.

[0059] The reverse image generating unit 34 ( Figure 2 ) Generate an inverted image by performing data expansion on a single-channel image. The inverted image generating unit 34 inverts the pixel values ​​of the single-channel image constituting the first supervised data set to generate an inverted image.

[0060] As a specific example of the inversion processing performed by the inversion image generating unit 34, when the minimum value of the pixel value of the single-channel image is 0 and the maximum value is 65535, a process of replacing the pixel value "p" in each coordinate with "65535-p" is performed.

[0061] In addition, as another specific example of the inversion process performed by the inversion image generating unit 34, an inversion process is performed with a transformation based on a function in between. When the pixels are configured in proportion to the logarithm of the physical quantity, the inversion image generating unit 34 can perform the following inversion process: after converting the pixel value "p" constituting the image into "exp(p)", a linear inversion (max(exp(p))-exp(p)) is performed after calculating the maximum and minimum values ​​after the transformation, and the logarithm of the value is taken.

[0062] In addition, the inversion processing of the inversion image generating unit 34 is not limited to the above-mentioned inversion processing, but also includes other processing such as transformation of a correspondence table (LUT; look up table) of pixel values ​​and pixel values ​​arbitrarily determined in advance, such as making points with relatively high pixel values ​​and points with relatively low pixel values ​​become low points and high points.

[0063] In addition, the interpretation of the physical quantity represented by the pixel is not limited to brightness, such as the case where the pixel value of a single-channel image is based on a brightness signal, the case where it is based on a pixel value signal representing depth information captured by a LiDAR (Light Detection and Ranging) camera device, the case where it is based on concentration information of the light and dark inspection read by a thin film that allows X-rays to pass through, etc.

[0064] As described above, in the reverse image generation unit 34, when there is a relationship that a physical phenomenon is expressed in a pixel value, a transformation is performed that can assume that the inverted pixel value corresponds to the physical phenomenon. On the other hand, before and after the transformation by the reverse image generation unit 34, a transformation that does not clearly transform the original defect is excluded from the transformation performed by the reverse image generation unit 34.

[0065] The reverse label generating unit 36 ​​( Figure 2 ) Generate reverse labels by performing data expansion on the correct answer labels.

[0066] When the single-channel image is inverted, the inverted label generating unit 36 ​​performs a class rewriting process of the correct answer label of the single-channel image. For example, when the single-channel image is a transmission image, information such as whether the object OBJ is concave or excessively raised is expressed in the brightness value of the single-channel image as a transmission image. In this case, the brightness value sometimes becomes important information in order to distinguish the type of defects. In this case, the class in the correct answer label is rewritten, the data of the correct answer label is expanded, and an inverted label is generated.

[0067] Figure 5 This is a diagram for explaining the class rewriting process of the correct answer label by the inversion label generating unit 36 ​​.

[0068] Figure 5 An image 60 representing a low density defect and an image 62 representing a high density defect are shown.

[0069] Image 60 has a region 66 indicating FMLD (Foreign Material Less Dense). Region 66 is darker than background 64. Here, FMLD is a general term for low-density foreign matter. For example, when bubbles or gas are mixed in when a gap or insufficient inflow of cast metal or when it is damaged and cut off, these regions (regions indicating FMLD (region 66)) appear darker than the surrounding (background 64). In addition, image 60 and the correct answer label 72 constituting the first supervised data set become FMLD.

[0070] Image 62 has an area 70 representing FMMD (Foreign Material More Dense). Area 70 is brighter than background 68. Here, FMMD is a general term for high-density foreign matter. For example, when a part is made of aluminum and it bulges more than necessary during the casting process, or when splashed metal is attached and impurities with a higher density than aluminum are buried inside, these areas (areas representing FMMD (areas 70)) are difficult to penetrate X-rays and therefore appear "white" (bright). That is, in the area representing FMMD (area 70), the pixel value becomes larger than usual. In addition, when image 62 is used as learning data, the correct answer label 74 becomes FMMD.

[0071] As described above, the relationship between the region 66 indicating FMLD and the background 64 is inverted with the relationship between the region 70 indicating FMMD and the background 68. Therefore, when the region where foreign matter or defects exist (the region 66 indicating FMLD and the region 70 indicating FMMD) and the region of the background (the background 64 and the background 68) are "black and white reversed" as a whole, the height of the defect region relative to the background is reversed. In this case, the correct answer label 72 is also reversed and converted into a reversed label (correct answer label 74).

[0072] In addition, Figure 5 In the above description, the case where the image 60 having the area 66 indicating FMLD is reversed from black and white to the image 62 having the area 70 indicating FMMD is described, but the present invention is not limited to this. Figure 6 As shown, the image 62 having the region 70 indicating FMMD can also be similarly inverted from black and white to the image 60 having the region 66 indicating FMMLD.

[0073] Furthermore, by rewriting the correct answer label, it is possible to handle without any problem a situation where defects of both classifications are mixed in the same image.

[0074] In the above description, the case where FMLD and FMMD correspond one to one is described. However, FMLD and FMMD do not necessarily correspond one to one. In such a case where there is no one-to-one correspondence, it is preferred to edit the class design corresponding to the correct answer label constituting the second supervised data set by setting the inverted image as the learning data.

[0075] Figure 7 A diagram illustrating the class design for editing the correct answer labels (reversing labels) of the second supervised dataset.

[0076] Figure 7 (A) is a diagram for explaining the case where the correct answer label FMMD has multiple rewrite addresses. Since the gas defect caused by the mixed gas is reflected relatively darker (small pixel value) compared to the background, it is the same as the above-mentioned FMLD. Therefore, when the correct answer label of FMMD is set as the inverted label, the rewrite address becomes FMLD and gas defect multiple (refer to Figure 7 (A)). Therefore, even if there is room for gas defects and FMLD to be considered as different classes in the correct answer label, gas defects and FMLD are edited and combined as a new class.

[0077] Figure 7 (B) is a diagram for explaining the class design of the correct answer label that edits and collects the rewrite addresses of the correct answer label FMMD. Figure 7 In the case shown in (B), the correct answer label of the rewritten address is edited to a correct answer label such as FMLD or gas defect, and is aggregated as a new class. This prevents the rewritten address from becoming multiple, and enables efficient learning in the detection or classification model learning.

[0078] In addition, the first supervised data set of the present invention preferably selects supervised data that does not produce contradictions in the reversed labels when the first supervised data set is expanded. Specifically, in the case of a defect indicating a correct answer label, it is preferred to select supervised data that has defects when the reversed labels are set as the first supervised data set. In addition, the first supervised data set preferably selects supervised data whose reversed labels are effective for learning when the first supervised data set is expanded.

[0079] In the above data expansion, the correct answer label (eg, classification class) associated with the image is sometimes constrained to remain unchanged before and after the data expansion. This is because the result of the data expansion does not become an erroneous class (non-existent data).

[0080] The present invention also considers selectively inverting only specific classes. For example, porosity (= tiny gas) defects tend to be densely generated over a large area, unlike foreign matter mixing. Images including this porosity defect can also be pixel-inverted in the same way and are equally regarded as a large number of white defects. However, in the case of an image where the inspection object does not actually produce an inverted result, that is, a "high-density and densely generated" defect, or even if there is a possibility of generation, it will not have an adverse effect on the quality of the product, so sometimes learning is not required in the case of being outside the inspection item.

[0081] In such a situation, it is preferable to apply pixel inversion only to images that do not include Porosity defects.

[0082] Thus, the first supervised data set of the present invention preferably selects supervised data that does not produce contradictions in the reversed labels when the first supervised data set is expanded. In addition, the first supervised data set preferably selects supervised data whose reversed labels are effective for learning when the first supervised data set is expanded.

[0083] Figure 8 The diagram illustrates specific examples of the first supervised data set and the second supervised data set.

[0084] The first supervised data set S is composed of Raw images 42 and correct answer labels 44 as learning data.

[0085] The raw image 42 is subjected to an inversion process by the inversion image generating unit 34, and a black-and-white inversion image 46 is generated. In addition, the correct answer label 44 is subjected to a class rewriting process by the inversion label generating unit 36, and an inversion label 48 is generated. Moreover, the black-and-white inversion image 46 and the inversion label 48 constitute a second supervisory data set T. The first supervisory data set S and the second supervisory data set T are used as learning data for machine learning of the learning model 40.

[0086] As described above, the second supervised dataset T is generated by performing data expansion from the first supervised dataset S. Then, the learning model 40 is learned using the first supervised dataset S and the second supervised dataset T. Thus, the learning model 40 can be efficiently learned.

[0087] Fig. 9 is a flowchart for explaining a data processing method executed by the data processing device 10. Figure 8 The flowchart will be described with reference to the specific example described in . In addition, each step is performed by the processor 14 of the data processing device 10 executing the data processing program 18 .

[0088] First, the single-channel image acquisition unit 30 acquires the Raw image 42 from the supervised dataset DB22 (step S10). Thereafter, the correct answer label acquisition unit 32 acquires the correct answer label 44 of the Raw image 42 from the supervised dataset DB22 (step S11). Next, the inverted image generation unit 34 generates a black-and-white inverted image 46 by inverting the Raw image 42 (a process of generating learning data for the second supervised dataset: step S12). Thereafter, the inverted label generation unit 36 ​​rewrites the class of the correct answer label 44 to generate an inverted label 48 (a process of generating the correct answer label constituting the second supervised dataset: step S13). Thereafter, the data processing device 10 outputs the black-and-white inverted image and the inverted label from the input / output interface 12, and stores the black-and-white inverted image and the inverted label in the second supervised dataset DB26 (step S14).

[0089] As described above, by performing data expansion (reversal processing, class rewriting processing) on ​​the Raw image 42 and the correct answer label 44 constituting the first supervised data set S, a black-and-white reversed image 46 and a reversed label 48 constituting the second supervised data set T are generated. Moreover, the generated first supervised data set S and the second supervised data set T are used together for learning the learning model 40. Thus, the learning model 40 can be effectively learned.

[0090] In the above-mentioned embodiment, the hardware structure of the processing unit (single-channel image acquisition unit 30, correct answer label acquisition unit 32, reverse image generation unit 34, and reverse label generation unit 36) that performs various processing is various processors (processors) as shown below. The various processors include general-purpose processors that execute software (programs) and function as various processing units, namely CPUs (Central Processing Units), FPGAs (Field Programmable Gate Arrays) and other processors whose circuit structures can be changed after manufacturing, namely programmable logic devices (PLDs), ASICs (Application Specific Integrated Circuits), and other processors that have circuit structures specially designed to perform specific processing, namely dedicated circuits, etc.

[0091] A processing unit can be composed of one of these various processors, or it can be composed of two or more processors of the same or different types (for example, multiple FPGAs, or a combination of a CPU and an FPGA). In addition, multiple processing units can also be composed of one processor. As an example of a plurality of processing units composed of one processor, first, there is a method such as a computer such as a client or server, which is represented by a combination of one or more CPUs and software to constitute a processor, and the processor functions as multiple processing units. Second, there is a method such as a system on chip (System On Chip: SoC) and the like, which uses a processor that implements the functions of the entire system including multiple processing units by one IC (Integrated Circuit) chip. In this way, various processing units are composed of one or more of the above-mentioned various processors as a hardware structure.

[0092] Furthermore, more specifically, the hardware structure of these various processors is a circuit (circuitry) formed by combining circuit elements such as semiconductor elements.

[0093] The above-mentioned structures and functions can be appropriately implemented by any hardware, software, or a combination of the two. For example, the present invention can also be applied to a program that causes a computer to execute the above-mentioned processing steps (processing sequence), a computer-readable recording medium (non-transitory recording medium) recording such a program, or a computer capable of installing such a program.

[0094] <Second embodiment>

[0095] Next, a second embodiment will be described. In this embodiment, the learning data is normalized or standardized. According to this embodiment, it is possible to prevent the deviation of pixel values ​​due to inversion, and to effectively learn the learning model 40.

[0096] Fig.10 is a graph showing the pixel values ​​that make up the image. Fig.10 (A) shows the pixel values ​​that make up the Raw image a. Fig.10 (B) shows the pixel values ​​constituting the reverse image a obtained by the reverse image generation unit 34 performing the reverse processing on the Raw image a. Fig.10 , the pixel value is shown on the horizontal axis and the number of pixels is shown on the vertical axis.

[0097] exist Fig.10 In the case shown in (A), when the available pixel values ​​of the image are 0 to 65535, the pixel values ​​constituting the target area of ​​the Raw image a are in the range of 10000 to 25000 (see Fig.10In this case, the pixel values ​​constituting the target area in the reverse image a are distributed roughly between 40,000 and 55,000 (see Fig.10 Thus, when the pixel values ​​of the target area are significantly different between the raw image a and the reversed image a, the learning model 40 may not be able to perform learning appropriately even if the first and second supervised datasets using these pixel values ​​as learning data are input to the learning model 40.

[0098] Therefore, in this embodiment, by normalizing or standardizing the Raw image a, we can only focus on the relative highs and lows of the pixel value distribution. Therefore, the influence of normalization or standardization is suppressed, so that the learning model 40 can effectively learn the first supervised data set and the second supervised data set.

[0099] In addition, in the above description, a method for normalizing or standardizing the Raw image a is described, but as a processing method when the pixel values ​​of the focus area are significantly different between the Raw image a and the inverted image a, the input layer of the learning model 40 may also have a mechanism such as Layer Normalization that suppresses the high and low of each patch image.

[0100] <Third Embodiment>

[0101] Next, the third embodiment is described. In this embodiment, the pixel values ​​of a part of the single-channel image are inverted to generate an inverted image, and the correct answer label constituting the second supervised data set is generated based on the correct answer label of the single-channel image corresponding to the part of the image.

[0102] Fig.11 The diagrams are for explaining the first supervised data set and the second supervised data set of the present embodiment.

[0103] The single-channel image 80 includes a component region C1 and a background region D other than the component region C1. Alternatively, a mask image may be superimposed on the single-channel image 80 to separate the component region C1 from the background region D.

[0104] The reverse image generation unit 34 performs the reverse processing on only the part region C1 of the single-channel image 80 to generate the reverse image 82. The reverse image 82 is composed of an image in which only the part region C1 is reversed, and the background region D is still the background region D of the single-channel image 80. In addition, the correct answer label 84 in the part region C1 is FMLD, and the reverse label generation unit 36 ​​rewrites the correct answer label 86 (FMMD).

[0105] Thus, in this embodiment, a reverse image is generated only for a focus area such as a part area, and a reverse label is generated based on the correct answer label for the focus area. Thus, the second supervised data set can be generated focusing on the focus area, so that the learning model 40 can learn more effectively.

[0106] <Configuration example of imaging system>

[0107] Next, an example of the imaging system 500 for capturing an image (single-channel image) of the object OBJ will be described. Fig.12 514 is a block diagram schematically showing a configuration example of an imaging system 500. The imaging system 500 is used to image an object OBJ placed in an imaging room 514, and includes an imaging control unit 502, an imaging operation unit 504, an image recording unit 506, an imaging device 508, and radiation sources 510 and 512.

[0108] The photography control unit 502 includes a CPU that controls the operation of each unit of the photography system 500. The photography control unit 502 receives operation input from an operator (photographer) via the photography operation unit 504, sends a control signal corresponding to the operation input to each unit of the photography system 500, and controls the operation of each unit.

[0109] The imaging operation unit 504 includes an input device for receiving operation input from the operator. The operator can input information related to the object OBJ, input instructions for imaging conditions and imaging execution for the imaging device 508, input instructions for irradiation conditions of radiation sources 510 and 512, and input instructions for recording images obtained by imaging in the image recording unit 506 through the imaging operation unit 504. The imaging conditions include, for example, imaging conditions such as exposure time, focal length, aperture, imaging angle, and imaging part. The irradiation conditions of radiation include irradiation start time, irradiation duration, irradiation angle, and irradiation intensity.

[0110] The image recording unit 506 records the image data (light-receiving image) of the object OBJ captured by the imaging device 508. Information for identifying the object OBJ is recorded in the image recording unit 506 in association with the image data.

[0111] The imaging device 508 and the radiation sources 510 and 512 are arranged inside the imaging room 514. The radiation sources 510 and 512 are, for example, X-ray sources, and the partition wall and the entrance and exit between the imaging room 514 and the outside are protected by X-ray shielding materials (for example, lead or concrete, etc.). In addition, when the object OBJ is irradiated with visible light for imaging, it is not necessary to use the protected imaging room 514.

[0112] The radiation sources 510 and 512 irradiate radiation to the object OBJ placed in the imaging room 514 in accordance with an instruction from the imaging control unit 502 .

[0113] The imaging device 508 receives radiation irradiated from the radiation source 510 to the object OBJ and reflected by the object OBJ, or radiation irradiated from the radiation source 512 to the object OBJ and transmitted through the object OBJ, and images the object OBJ in accordance with the instruction of the imaging execution from the imaging control unit 502. The object OBJ is held in the imaging room 514 by a holding member (for example, a manipulator, a stage, or a movable stage, etc.) not shown in the figure, and the distance and angle of the object OBJ relative to the imaging device 508 and the radiation sources 510 and 512 are adjustable. The operator can control the relative positions of the object OBJ, the imaging device 508, and the radiation sources 510 and 512 through the imaging control unit 502, and can image a desired part of the object OBJ.

[0114] The radiation sources 510 and 512 terminate irradiation of the object OBJ with radiation in synchronization with the termination of the imaging by the imaging device 508 .

[0115] In addition, Fig.12 In the example shown, the imaging device 508 is arranged inside the imaging room 514, but the imaging device 508 may be arranged outside as long as it can capture the object OBJ in the imaging room 514. Fig.12 In the example shown, one imaging device 508 is provided, and two radiation sources 510 and 512 are provided, but the number of imaging devices and radiation sources is not limited thereto. For example, there may be multiple imaging devices and radiation sources, or one each. The imaging control unit 502, the imaging operation unit 504, and the image recording unit 506 may be implemented using a combination of computer hardware and software.

[0116] The data processing device 10 may be connected to the photographing system 500 in a communicable manner, and the data processing device 10 may be configured to function as the photographing control unit 502 , the photographing operation unit 504 , and the image recording unit 506 of the photographing system 500 .

[0117] As mentioned above, although the example of this invention was demonstrated, this invention is not limited to the said embodiment, Of course, various deformation|transformation is possible within the range which does not deviate from the main point of this invention.

[0118] Explanation of symbols

[0119] 8 Bus

[0120] 10 Data processing device

[0121] 12 Input and output interfaces

[0122] 14 Processor

[0123] 16 Memory

[0124] 18 Data Processing Procedures

[0125] 30 Single channel image acquisition unit

[0126] 32 Correct answer label acquisition section

[0127] 34. Reverse image generation unit

[0128] 36 Reverse label generation unit

[0129] 40 Learning Model

[0130] 42 Raw images

[0131] 44 Correct Answer Tags

[0132] 46 Black and white reverse image

[0133] 48 Reverse Label

[0134] 50 Radiation Source

[0135] 54 Detector

[0136] 56 single channel images

Claims

1. A data processing device, comprising a processor, wherein: The processor generating an inverted image in which pixel values ​​of a single-channel image constituting a first supervised data set are inverted, as learning data of a second supervised data set obtained by expanding the first supervised data set, wherein the pixel values ​​of the single-channel image are determined according to a physical quantity sensed by a light-receiving element during photography, or the pixel values ​​are determined by reversibly transforming the physical quantity, Based on the correct answer label of the single-channel image constituting the first supervised data set, a reverse label corresponding to the reversed image is generated as the correct answer label constituting the second supervised data set.

2. The data processing device according to claim 1, wherein: The single-channel image is an image captured by a digital detector array DDA that receives radiation transmitted through a subject, or a computed radiography CR image obtained by outputting a light-receiving signal in the form of a digital value via a reader on an imaging plate IP that receives radiation transmitted through a subject.

3. The data processing device according to claim 1, wherein: The single-channel image is an image based on lens-free imaging.

4. The data processing device according to claim 1, wherein: Each pixel of the single-channel image has a digital value proportional to the amount of light received by the corresponding light-receiving element, or a digital value correlated with the amount of light received.

5. The data processing device according to claim 1, wherein: Each pixel of the single-channel image has a digital value correlated with the physical quantity at each point of the corresponding object.

6. The data processing device according to claim 1, wherein: In the case where the first supervised dataset is extended, the first supervised dataset selects supervised data that does not produce contradictions in the inversion labels.

7. The data processing device according to claim 1, wherein: In the case where the first supervised dataset is extended, the first supervised dataset selects supervised data whose inverted labels are effective for learning.

8. The data processing device according to claim 1, wherein: The processor Inverting the pixel values ​​of a portion of the single-channel image to generate the inverted image, Based on the correct answer label of the single-channel image corresponding to the part of the area, a correct answer label constituting the second supervised data set is generated.

9. The data processing device according to claim 1, wherein: The processor normalizes or standardizes the learning data.

10. The data processing device according to claim 1, wherein: The processor edits a class design of the correct answer labels constituting the second supervised dataset corresponding to setting the inverted image as the learning data.

11. The data processing device according to claim 1, wherein: The reversible transformation is at least one of a linear transformation, a logarithmic transformation, and a nonlinear transformation using a pixel value correspondence table.

12. A data processing method, wherein: The processor performs the following steps: Generate an inverted image by inverting pixel values ​​of a single-channel image constituting a first supervised data set, as learning data of a second supervised data set obtained by expanding the first supervised data set, wherein the pixel value of the single-channel image is determined according to a physical quantity sensed by a light-receiving element during photography, or the pixel value is determined by reversibly transforming the physical quantity; as well as Based on the correct answer label of the single-channel image constituting the first supervised data set, a reverse label corresponding to the reversed image is generated as the correct answer label constituting the second supervised data set.

13. A program, wherein: The processor executes a data processing method, which includes the following steps: Generate an inverted image by inverting pixel values ​​of a single-channel image constituting a first supervised data set, as learning data constituting a second supervised data set obtained by expanding the first supervised data set, wherein the pixel value of the single-channel image is determined according to a physical quantity sensed by a light-receiving element during photography, or the pixel value is determined by reversibly transforming the physical quantity; as well as Based on the correct answer label of the single-channel image constituting the first supervised data set, a reverse label corresponding to the reversed image is generated as the correct answer label constituting the second supervised data set.

14. A recording medium which is non-transitory and computer-readable, wherein: The program according to claim 13 is recorded.

Images