Inference device, inference method, and inference system

Abstract

This approach reduces wasted time and effort on human work while minimizing the impact of unhelpful or misleading training data on the model, thereby reducing false positives and missed detections. [Solution] The inference device provides a processor with an inference function that inputs input data to an AI model and performs inference, and a correction function that corrects the inference results from the inference function based on the quality of the features extracted from the input data.

Description

Technical Field

[0001] The present invention relates to an inference apparatus, an inference method, and an inference system.

Background Art

[0002] Patent Literature 1 describes an inspection apparatus. The inspection apparatus includes: an image storage unit that captures a plurality of images of an inspection object with different input systems under predetermined imaging conditions corresponding to each input system, and stores multiple inspection images each being a set of the plurality of images of the inspection object obtained by the imaging; and a determination unit that processes, for each pixel, the multiple inspection images stored in the image storage unit based on a learning model created in advance through learning using images with the same imaging conditions as the multiple inspection images to obtain a degree of quality defect, and determines the quality state of the inspection object by comparing the degree of quality defect with a preset threshold.

Prior Art Literature

Patent Literature

[0003]

Patent Literature 1

Summary of the Invention

Problem to be Solved by the Invention

[0004] Inference using an AI model may have errors in the inference results. For example, in the case of an AI model that performs an object detection task, false detection of objects may occur. Errors in inference results may be caused when the AI model learns confounding data. Note that confounding data refers to data that hinders model learning when only context-only information is input to the AI model.

[0005] A typical example of disruptive data is data where the location of the data feature that the AI ​​focuses on during inference differs from the location of the data feature itself. For example, suppose an input image containing a truck is input to an AI model, and the model correctly classifies it as a "truck" through inference. However, upon examining the saliency map (described later), it may be found that the AI ​​focused not on the truck itself in the input image, but on a watermark with the word "truck" inserted at the edge of the input image, classifying the object as a truck. In such a case, if an input image without the watermark is input to the same AI model, it may not be able to correctly classify it as a "truck." A saliency map, also known as a heatmap, is an image that shows the areas that the AI ​​model is looking at during prediction or analysis.

[0006] To correct the errors in the inference results described above, it is necessary for humans to identify the interfering data, remove it from the dataset, and then retrain the AI ​​model. However, this manual process is time-consuming and problematic.

[0007] The present invention has been made in view of the above circumstances, and aims to provide an inference device, inference method, and inference system that reduce false positives and false negatives by minimizing the impact of unhelpful or erroneous training data on the model, while reducing wasted time and effort spent by humans. [Means for solving the problem]

[0008] To achieve the aforementioned objectives, the inference device, inference method, and inference system of the present invention are characterized by the following:

[0009] [1] The processor, An inference function that performs inference by inputting data into the AI ​​model, A correction function that corrects the inference results obtained by the inference function based on the quality of the features extracted from the input data, An inference device that makes this possible.

[0010] [2] An inference method using a device equipped with a processor, An inference step in which input data is entered into an AI model and inference is performed, A correction step in which the inference result by the inference function is corrected based on the quality of the features extracted from the input data, An inference method having the following characteristics.

[0011] [3] An inference system comprising at least a device equipped with a processor, The aforementioned processor, An inference function that performs inference by inputting data into the AI ​​model, A correction function that corrects the inference results obtained by the inference function based on the quality of the features extracted from the input data, An inference system that makes this possible. [Effects of the Invention]

[0012] According to the present invention, it is possible to provide an inference device, inference method, and inference system that reduce false positives and false negatives by minimizing the impact of unhelpful or erroneous training data on the model, while reducing wasted time and effort spent by humans. [Brief explanation of the drawing]

[0013] [Figure 1] A block diagram showing an example of the configuration of an inference system corresponding to at least one embodiment of the present invention. [Figure 2] A block diagram showing the configuration of a server corresponding to at least one embodiment of the present invention. [Figure 3] A block diagram showing the configuration of a server corresponding to at least one embodiment of the present invention. [Figure 4] A conceptual diagram showing an example of calculating quality, corresponding to at least one embodiment of the present invention. [Figure 5] A block diagram showing the configuration of a server corresponding to at least one embodiment of the present invention. [Figure 6]A block diagram showing the configuration of a server corresponding to at least one embodiment of the present invention. [Figure 7] A block diagram showing the configuration of a server corresponding to at least one embodiment of the present invention. Mode for Carrying Out the Invention

[0014] Hereinafter, examples of embodiments of the present invention will be described with reference to the drawings. Various constituent elements in the examples of each embodiment described below can be appropriately combined within a range that does not cause contradiction, etc. Further, with respect to the content described as an example of a certain embodiment, the description thereof may be omitted in other embodiments. In addition, the content of operations and processes not related to the characteristic parts of each embodiment may be omitted. Furthermore, the order of various processes constituting the various flows and sequences described below is not limited within a range that does not cause contradiction in the process content.

[0015] Hereinafter, an inference program executed in a server, which is an example of a computer, will be described as an example. However, the computer may be another device such as a user terminal. Further, the inference system as a whole may execute the inference program.

[0016] Figure 1 is a block diagram showing an example configuration of an inference system corresponding to at least one embodiment of the present invention. The inference system 1 includes a server 10 and a user terminal 20 used by a user of the inference system 1. User terminals 20A, 20B, and 20C are each an example of the user terminal 20. The configuration of the inference system 1 is not limited to this. For example, the inference system 1 may be configured such that a single user terminal is used by a plurality of users. The inference system 1 may include a plurality of servers.

[0017] The server 10 is an example of an inference apparatus according to an embodiment of the present disclosure. The inference system 1 may be a foreign object detection system using X-rays, but may also be a system other than that. In this case, the server 10 or the user terminal 20 may be a foreign object detection apparatus, but may also be an apparatus other than that.

[0018] The server 10 and the user terminal 20 are each communicably connected to a communication network 30. The connection between the communication network 30 and the server 10, and the connection between the communication network 30 and the user terminal 20 may be either a wired connection or a wireless connection. The communication network 30 may be a LAN network, a Wifi network, an Internet line, an analog telephone line, or a line combining a plurality of these, for example.

[0019] By including the server 10 and the user terminal 20, the inference system 1 implements various functions for executing various processes in accordance with user operations.

[0020] The server 10 includes a processor 11, a memory 12, and a storage device 13. The processor 11 is, for example, a central processing unit such as a CPU (Central Processing Unit) that performs various calculations and controls. Further, when the server 10 includes a GPU (Graphics Processing Unit), part of various calculations and controls may be performed by the GPU. The server 10 causes the processor 11 to execute various types of information processing using data read into the memory 12, and stores the obtained processing result in the storage device 13 as necessary.

[0021] The storage device 13 functions as a storage medium for storing various types of information. The configuration of the storage device 13 is not particularly limited, but from the viewpoint of reducing the processing load on the user terminal 20, it may be configured to store all the various types of information necessary for the control performed by the inference system 1. Examples of such configurations include HDDs and SSDs. However, the storage device that stores the various types of information only needs to have a storage area that is accessible to the server 10, and for example, it may be configured to have a dedicated storage area outside the server 10.

[0022] The user terminal 20 is managed by the user. The user terminal 20 may be, for example, various measuring devices. The user terminal 20 may also be any device that provides services to the user.

[0023] The user terminal 20 may be equipped with hardware and software for performing various processes by connecting to the communication network 30 and communicating with the server 10. Each of the multiple user terminals 20 may also be configured to communicate directly with one another without going through the server 10.

[0024] The user terminal 20 may include a processor 21, memory 22, and storage device 23. The processor 21 is, for example, a central processing unit such as a CPU (Central Processing Unit) that performs various calculations and controls. If the user terminal 20 is equipped with a GPU (Graphics Processing Unit), some of the various calculations and controls may be performed by the GPU. The user terminal 20 uses the data read into the memory 22 to perform various information processing using the processor 21, and stores the obtained processing results in the storage device 23 as needed. The storage device 23 functions as a storage medium for storing various information.

[0025] Figure 2 is a block diagram showing the configuration of a server corresponding to at least one embodiment of the present invention. Server 10 comprises an inference unit 101, an explanation unit 102, an identification unit 103, a storage unit 104, an analysis unit 105, a correction unit 106, and a preprocessing unit 107. The processor 11 of Server 10 refers to an inference program held in the storage device 13 and executes the program, thereby functionally realizing the inference unit 101, the explanation unit 102, the identification unit 103, the analysis unit 105, the correction unit 106, and the preprocessing unit 107.

[0026] The processor 11, memory 12, and storage device 13 of the server 10 functionally implement the storage unit 104. Alternatively, the storage unit 104 may be functionally implemented in an external storage device that is connected to the server 10 in a manner that allows communication.

[0027] The inference unit 101 has the function of inputting input data into the AI ​​model and performing inference. The explanation unit 102 has the function of generating explanatory information for the AI ​​model's judgment based on the inference results from the inference unit 101. Explanatory information for the judgment is information generated by an explainable AI (XAI) that expresses the basis for the results derived by the AI ​​and the process of the judgment in a way that humans can understand. The identification unit 103 has the function of identifying the quality of the features extracted from the input data based on the annotation of the input data and the explanatory information generated by the explanation unit 102. For example, the identification unit 103 may identify whether the features extracted from the input data are pure data or interfering data as the quality.

[0028] The memory unit 104 stores the features extracted from the input data and the identification results from the identification unit 103. The features extracted from the input data may be, for example, feature vectors.

[0029] The analysis unit 105 has the function of analyzing features extracted from the input data according to a plurality of features stored in the memory bank of the storage unit 104. The correction unit 106 has the function of correcting the inference result by the inference unit 101 based on the quality of the features extracted from the input data. The correction unit 106 may correct the inference result by the inference unit 101 according to the analysis result by the analysis unit 105. The preprocessing unit 107 has the function of performing preprocessing on the features extracted from the input data.

[0030] The input data may be, for example, image data. The AI ​​model in this case may be, for example, an AI model for object detection. The image data may be in formats such as grayscale or RGB. The input data may also be in other formats, such as text data.

[0031] Figure 3 is a block diagram showing the configuration of a server corresponding to at least one embodiment of the present invention.

[0032] The inference unit 101 is equipped with an AI model. The AI ​​model itself can be a conventional one, so no specific limitations are imposed. For the sake of explanation, in this example, we will assume that the AI ​​model is an object detection network, consisting of a neural network that performs object detection. Furthermore, we will assume that the input data is image data.

[0033] Figure 3 shows the state of the preparation phase after training the AI ​​model. The object detection network comprises a feature extraction unit 101A, a classifier 101B, and a localizer 101C. An input image is input to the feature extraction unit 101A. The feature extraction unit 101A extracts features from the input image and outputs a feature map. A feature map is data in which the extracted features are represented, for example, as feature vectors. The output feature map is input to the classifier 101B, the localizer 101C, and the preprocessing unit 107.

[0034] Classifier 101B works in conjunction with localizer 101C to detect anomalies in the input data. Localizer 101C predicts the location of the anomaly, and classifier 101B predicts the type of anomaly and an estimated reliability (probability). The prediction result of localizer 101C may be, for example, the smallest bounding box surrounding the anomaly. In this example of an object detection network, classifier 101B classifies the type of anomaly of the object (e.g., NG-Foreign Object A, NG-Foreign Object B). The output of classifier 101B may be a value indicating the classified class or a continuous value. In this example, classifier 101B outputs a real number ranging from 0 to 1, and in Figure 3, 0.84 (meaning 84%) is output. If multiple anomalies are detected, the bounding box, classification, and reliability are output for each anomaly. The detection result obtained through the cooperation of classifier 101B and localizer 101C corresponds to the inference result of the AI ​​model.

[0035] The explanation unit 102, also known as XAI, is composed of, for example, Eigen-CAM. Based on the inference results of the inference unit 101, the explanation unit 102 generates explanatory information for the judgment made by the AI ​​model. In this example, the explanation unit 102 generates a spleness map as explanatory information for the judgment, which shows which parts of the input image were focused on when the inference unit 101 performed its inference.

[0036] A sampling map, also known as a heatmap, is an image that shows the areas that an AI model is looking at when performing predictions or analyses.

[0037] The identification unit 103 identifies the quality of the features extracted from the input image based on the annotations of the input image and the explanatory information generated by the explanation unit 102. Quality is an indicator of whether or not the features extracted from the input data are suitable for training the AI ​​model.

[0038] The evaluation of goodness can be a binary evaluation. In this example, the feature extraction unit 101A identifies two types of features extracted: disruptive data and pure data. Pure data refers to data that is useful for learning the AI ​​model by inputting meaningful information into the AI ​​model. Disruptive data refers to data that hinders model learning by inputting context-only information into the AI ​​model. The identification unit 103 may also identify the goodness by dividing it into three or more classes.

[0039] The preprocessing unit 107 performs preprocessing on the features of the input image acquired from the AI ​​model, such as compression, PCA, or SVD. Other preprocessing may also be performed, and since the types of preprocessing are conventional, a detailed explanation will be omitted.

[0040] The memory unit 104 is also called a memory bank. The memory unit 104 stores features extracted from the input data obtained from the AI ​​model and the identification results from the identification unit 103. The memory unit 104 may store pre-processed features as shown in Figure 3, or it may store features without preprocessing. When evaluating the quality in a binary manner, the memory unit 104 stores information indicating whether the feature is interfering data or pure data as an identification result from the identification unit 103.

[0041] Figure 4 is a conceptual diagram showing an example of quality calculation corresponding to at least one embodiment of the present invention.

[0042] In this embodiment, IoU is used as the algorithm for calculating the quality score. IoU is an abbreviation for Intersection over Union, and is an index that represents how much two regions overlap. IoU can be expressed as IoU = Area of ​​the overlapping region of the two regions / Total area of ​​the two regions.

[0043] In this embodiment, the identification unit 103 identifies the IoU of the annotation and the highly splendid portion as the degree of quality. More specifically, it is the IoU of the annotated region in the input image and the region determined to be splendid in the splendor map. In this example, the area is the number of pixels.

[0044] In the calculation example shown in Figure 4, there is no overlap between the annotated region A in the input image and region B, which is determined to have splendor in the splendor map, and therefore the area is 0 pixels. On the other hand, the total area of ​​regions A and B combined is 143 pixels. In this case, the goodness score is calculated as 0 / 143 = 0.

[0045] The identification unit 103 can distinguish the quality of data using binary values, for example, by evaluating the IoU value calculated as described above as pure data if it is above a predetermined threshold, and evaluating the IoU value as disruptive data if it is below a predetermined threshold. It is also possible to use two thresholds to distinguish data using three values, such as less than the first threshold, above the first threshold and less than the second threshold, and above the second threshold. Similarly, it is also possible to distinguish data using four or more values.

[0046] Figure 5 is a block diagram showing the configuration of a server corresponding to at least one embodiment of the present invention.

[0047] Figure 5 shows an example configuration corresponding to the inference phase. Detailed explanations of the configurations shown in Figure 5 that are the same as those shown in Figure 3 are omitted.

[0048] The analysis unit 105 analyzes the features extracted from the input data according to the multiple features stored in the memory bank of the storage unit 104. More specifically, the analysis unit 105 analyzes the feature map output from the feature extraction unit 101A, or the data after the feature map has been preprocessed by the preprocessing unit 107.

[0049] The analysis algorithm in the analysis unit 105 may be, for example, a K-NN search. In the case of a K-NN search, when the analysis unit 105 is given a feature to be analyzed, such as a feature vector, it refers to K features in the vicinity of the feature to be analyzed from among the multiple features stored in the memory bank of the storage unit 104. In this embodiment, the features in the vicinity of a given feature may be referred to as neighboring features. K is a natural number greater than or equal to 2.

[0050] The analysis unit 105 generates a coefficient corresponding to the quality of the K neighboring features that were referenced. The coefficient may be calculated, for example, as (number of pure data) / (number of pure data + number of interfering data). For example, six neighboring features in the vicinity of the feature to be analyzed are referenced. In this case, K is 6. When the quality of each of the six neighboring features is obtained from the memory bank of the storage unit 104, there is 1 pure data and 5 interfering data. In this case, the coefficient generated by the analysis unit 105 is 1 / (1+5)=1 / 6.

[0051] The correction unit 106 corrects the inference results from the inference unit 101 based on the quality of the features extracted from the input data. More specifically, the correction unit 106 corrects the inference results from the inference unit 101 based on the analysis results from the analysis unit 105.

[0052] For example, if the inference result of the inference unit 101 is 0.93 (i.e., 93%), the correction unit 106 corrects the inference result by multiplying it by the above coefficient. In the example shown in Figure 5, the corrected inference result is 0.93 × (1 / 6) = 0.155 (i.e., 15.5%).

[0053] Furthermore, the analysis algorithm used by the analysis unit 105 is not limited to K-NN search, and a different analysis algorithm may be used. The correction performed by the correction unit 106 is not limited to coefficient multiplication, and may also be, for example, the addition of a penalty value according to the analysis result by the analysis unit 105.

[0054] (modified version) As a variation, an example configuration is shown where the AI ​​model is a black box model. Figure 6 is a block diagram showing the configuration of a server corresponding to at least one embodiment of the present invention. Figure 7 is a block diagram showing the configuration of a server corresponding to at least one embodiment of the present invention. Figure 6 shows an example configuration corresponding to the preparation phase shown in Figure 3. Figure 7 shows an example configuration corresponding to the inference phase shown in Figure 5.

[0055] If a given AI model is a black-box model, it may not be possible to obtain feature maps, which are intermediate results in the AI ​​model's processing, from partway through the network. In such cases, the inference unit 101 further includes a white-box AI model. The white-box AI model includes at least a feature extraction unit 101A, as shown in Figures 3 and 5. The white-box AI model may further include a classifier 101B and a localizer 101C.

[0056] The same input image is input to both the black-box model and the white-box model. Then, instead of the output from the feature extraction unit 101A described in Figures 3 and 5, the output from the feature extraction unit 101A of the white-box model is used. This makes it possible to handle cases where the given AI model is a black-box model in the same way as the embodiment described with reference to Figures 1 through 5.

[0057] Although various embodiments have been described above with reference to the drawings, it goes without saying that this disclosure is not limited to such examples. It is clear to those skilled in the art that various modifications or alterations can be conceived within the scope of the claims, and these will naturally also fall within the technical scope of this disclosure. For example, the steps in the methods of this disclosure can be performed in any order as long as they do not cause any inconsistencies. Furthermore, the components of the embodiments described above may be combined in any way without departing from the spirit of the disclosure.

[0058] In each of the embodiments described above, the user terminal 20 and the server 10 execute the various processes described above in accordance with various control programs (e.g., inference programs) stored in their own storage devices. Furthermore, other computers, not limited to the user terminal 20 and the server 10, may also execute the various processes described above in accordance with various control programs (e.g., inference programs) stored in their own storage devices.

[0059] Furthermore, the configuration of the inference system 1 is not limited to the configuration described as an example of the embodiment above. For example, the server may perform some or all of the processes described as being performed by the user terminal, or the user terminal may perform some or all of the processes described as being performed by the server. Alternatively, the user terminal may be equipped with some or all of the storage unit (memory device) provided by the server. In other words, the inference system 1 may be configured such that one of the user terminals or the server provides some or all of the functions provided by the other.

[0060] Furthermore, the program may be configured to implement some or all of the functions described above as examples of each embodiment in a standalone device that does not include a communication network.

[0061] [Note] The above-described embodiments are written in such a way that at least the following invention can be put into practice by a person with ordinary skill in the art to which the invention pertains.

[0062] [1] In the processor, An inference function that performs inference by inputting data into the AI ​​model, A correction function that corrects the inference results obtained by the inference function based on the quality of the features extracted from the input data, An inference device that makes this possible.

[0063] [2] The aforementioned processor, An explanatory function that generates explanatory information for the judgment made by the AI ​​model based on the inference results of the aforementioned inference function, An identification function that identifies the quality of features extracted from the input data based on the annotation of the input data and the explanatory information generated by the explanatory function, To make it happen, The system further includes a storage means for storing features extracted from the input data and the identification results obtained by the identification function. [1] The inference device described above.

[0064] [3] The aforementioned identification function identifies whether the features extracted from the input data are pure data or interfering data, as the degree of quality. [2] The inference device described above.

[0065] [4] In the aforementioned correction function, The inference result obtained by the inference function is corrected based on the quality of multiple neighboring features in the vicinity of the features extracted from the input data. The inference apparatus described in [2] or [3].

[0066] [5] The aforementioned input data is image data. The aforementioned AI model is an AI model for object detection. [1] The inference device described above.

[0067] [6] The inference apparatus according to [1], further comprising a preprocessing function that performs preprocessing on features extracted from the input data for the processor.

[0068] [7] An inference method using a device equipped with a processor, An inference step in which input data is entered into an AI model and inference is performed, A correction step that corrects the inference result from the inference step based on the quality of the features extracted from the input data, An inference method having the following characteristics.

[0069] [8] An inference system comprising at least a device equipped with a processor, The aforementioned processor, An inference function that performs inference by inputting data into the AI ​​model, A correction function that corrects the inference results obtained by the inference function based on the quality of the features extracted from the input data, An inference system that makes this possible. [Explanation of symbols]

[0070] 1. Inference System 10 servers 11 processors 12 memory 13 Storage device 20, 20A, 20B User Terminals 21 processors 22 memory 23 Storage device 30 Communication Networks 101 Reasoning part 101A Feature extraction unit 101B Classifier 101C Localizer 102 Explanation Section 103 Identification Unit 104 Storage section 105 Analysis Department 106 Correction Unit 107 Pre-processing section

Claims

1. In the processor, An inference function that performs inference by inputting data into the AI ​​model, A correction function that corrects the inference results obtained by the inference function based on the quality of the features extracted from the input data, An inference device that makes this possible.

2. The aforementioned processor, An explanatory function that generates explanatory information for the judgment made by the AI ​​model based on the inference results of the aforementioned inference function, An identification function that identifies the quality of features extracted from the input data based on the annotation of the input data and the explanatory information generated by the explanatory function, To make it happen, The system further includes a storage means for storing features extracted from the input data and the identification results obtained by the identification function. The inference device according to claim 1.

3. The aforementioned identification function identifies whether the features extracted from the input data are pure data or interfering data, as the degree of quality. The inference device according to claim 2.

4. In the aforementioned correction function, Multiple neighboring features are obtained from the vicinity of the features extracted from the input data, and the inference result from the inference function is corrected based on the quality of the multiple neighboring features. The inference device according to claim 2 or claim 3.

5. The aforementioned input data is image data. The aforementioned AI model is an AI model for object detection. The inference device according to claim 1.

6. The inference apparatus according to claim 1, further comprising a preprocessing function that performs preprocessing on features extracted from the input data in the processor.

7. An inference method using a device equipped with a processor, An inference step in which input data is entered into the AI ​​model and inference is performed, A correction step that corrects the inference result from the inference step based on the quality of the features extracted from the input data, An inference method having the following characteristics.

8. An inference system comprising at least a device equipped with a processor, The aforementioned processor, An inference function that performs inference by inputting data into the AI ​​model, A correction function that corrects the inference results obtained by the inference function based on the quality of the features extracted from the input data, An inference system that makes this possible.

Images