Information processing device, information processing method, and program

Abstract

The present invention provides an information processing device, an information processing method, and a program that are useful for detecting abnormalities in an inspection object. [Solution] An information processing device according to an embodiment includes a first acquisition means, a first extraction means, a selection means, and a learning means. The first acquisition means acquires at least one training image including an inspection object. The first extraction means inputs the training image into a feature extraction model and extracts first feature values ​​of n dimensions (n ​​is an integer equal to or greater than 2) of the training image output from the feature extraction model. The selection means selects k dimensions (k is an integer equal to or greater than 1 and less than n) from the n dimensions. The learning means generates an anomaly detection model used to infer the state of the inspection object by learning the first feature values ​​of the k dimensions selected from the first feature values ​​of the n dimensions of the training image.

Description

[Technical Field]

[0001] An embodiment of the present invention relates to an information processing device, an information processing method, and a program. [Background technology]

[0002] In recent years, there has been a demand for automating, for example, visual inspections of products manufactured in factories and incoming inspections of parts used to manufacture those products using images, thereby resolving the labor shortage required for those inspections and achieving standardization of those inspections.

[0003] In such inspections, it is conceivable to detect abnormalities in the inspection object, such as a product or part, from an image using an anomaly detection model (trained model) generated by performing supervised learning using normal data (e.g., an image containing the inspection object in a normal state) and abnormal data (e.g., an image containing the inspection object in an abnormal state). However, defects rarely occur in the inspection object at the site where such inspection is performed, and it is not efficient to collect sufficient abnormal data before performing supervised learning of the anomaly detection model.

[0004] On the other hand, unsupervised learning (unsupervised anomaly detection technology) differs from the supervised learning described above in that it can be performed using only normal data, the cost of creating the data used for the learning (learning data) is low, and it is easy to introduce into the field.

[0005] However, an anomaly detection model generated by unsupervised learning, which learns only from normal data, may produce erroneous results in detecting anomalies in the test subject (i.e., the accuracy of the anomaly detection model may be low).

[0006] For this reason, there is a need for a mechanism that enables efficient learning of an anomaly detection model based on the above-mentioned perspective. [Prior art documents] [Patent documents]

[0007] [Patent Document 1] Japanese Patent Application Laid-Open No. 2004-351100 Summary of the Invention [Problem to be solved by the invention]

[0008] Therefore, an object of the present invention is to provide an information processing device, an information processing method, and a program that are capable of realizing efficient learning of an anomaly detection model. [Means for solving the problem]

[0009] An information processing device according to an embodiment includes a first acquisition means, a first extraction means, a selection means, and a learning means. The first acquisition means acquires a training image from an image database storing at least one training image including an inspection target. The first extraction means inputs the training image to a feature extraction model and extracts first feature quantities of n dimensions (n ​​is an integer equal to or greater than 2) of the training image output from the feature extraction model. The selection means selects k dimensions (k is an integer equal to or greater than 1 and less than n) from the n dimensions. The learning means generates an anomaly detection model used to infer the state of the inspection target by learning the selected first feature quantities of the k dimensions from the n-dimensional first feature quantities of the training image. [Brief explanation of the drawings]

[0010] [Figure 1] FIG. 1 is a block diagram showing an example of the functional configuration of an information processing apparatus according to a first embodiment. [Figure 2] FIG. 1 is a diagram showing an example of a hardware configuration of an information processing apparatus. [Figure 3] 10 is a flowchart showing an example of a processing procedure of a learning process. [Figure 4] FIG. 10 is a diagram showing an example of a normal image. [Figure 5] FIG. 10 is a diagram showing an example of an abnormal image. [Figure 6] FIG. 10 is a diagram showing an outline of a process for extracting feature amounts from learning images. [Figure 7] FIG. 1 is a diagram conceptually showing dimensional feature amounts. [Figure 8] 1 is a diagram conceptually illustrating dimension selection based on Mahalanobis distance. [Figure 9] 10 is a flowchart showing an example of a processing procedure of an inference process. [Figure 10] FIG. 1 is a diagram showing an overview of inference processing. [Figure 11] FIG. 10 is a block diagram showing an example of the functional configuration of an information processing apparatus according to a second embodiment. [Figure 12] FIG. 10 is a diagram for explaining an attention network used for weighting. [Figure 13] FIG. 1 is a diagram for explaining attention network learning. [Figure 14] FIG. 11 is a block diagram showing an example of the functional configuration of an information processing device according to a third embodiment. [Figure 15] FIG. 10 is a block diagram showing an example of the functional configuration of an information processing apparatus according to a fourth embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0011] Hereinafter, each embodiment will be described with reference to the drawings. (First embodiment) First, a first embodiment will be described. An information processing device according to this embodiment operates as an anomaly detection device for detecting an anomaly in an inspection object (inspecting the state of the inspection object) using, for example, an image including the inspection object. The inspection object in this embodiment is assumed to be, for example, a product manufactured in a factory or a part used to manufacture the product, but the inspection object may be any object or the like in which an anomaly occurs in the appearance represented in the image.

[0012] 1 is a block diagram showing an example of the functional configuration of an information processing device according to this embodiment. As shown in Fig. 1, the information processing device 10 includes an image database (DB) 11, a first model storage unit 12, a dimension selection information storage unit 13, a second model storage unit 14, a learning processing unit 15, and an inference processing unit 16.

[0013] The learning processing unit 15 includes a first acquisition unit 151 , a first extraction unit 152 , a first selection unit 153 , and a learning unit 154 .

[0014] Images (hereinafter referred to as training images) including inspection objects used for training an anomaly detection model (inference model) described below are stored in the image database 11. The training images stored in the image database 11 include at least images including inspection objects in a normal state (hereinafter referred to as normal images), but the training images may also include images including inspection objects in an abnormal state (hereinafter referred to as abnormal images).

[0015] The first acquisition unit 151 acquires (reads) one or more learning images from the image database 11. The first acquisition unit 151 may acquire learning images stored in a location indicated by a path specified by, for example, a user who uses the information processing device 10 (an administrator who manages the information processing device 10).

[0016] A feature extraction model is stored in the first model storage unit 12. The first extraction unit 152 uses the feature extraction model stored in the first model storage unit 12 to extract feature amounts of the learning image acquired by the first acquisition unit 151 from the learning image.

[0017] Here, the feature extraction model is realized by a neural network (NN) model such as a convolutional neural network (CNN) or a vision transformer (ViT) trained using a large number of images, and the first extraction unit 152 extracts the outputs of the intermediate layer and output layer of the neural network model to which the training image is input as the feature of the training image. In this embodiment, the first extraction unit 152 extracts n-dimensional feature values ​​(n is an integer equal to or greater than 2) from the training image. The feature values ​​extracted by the first extraction unit 152 may be n-dimensional feature vectors or H×W×n (H and W are integers equal to or greater than 1) feature maps.

[0018] Here, it has been described that features are extracted using a feature extraction model realized by a neural network model, but the first extraction unit 152 may also extract features such as color histograms or HOG (Histograms of Oriented Gradients) (hereinafter referred to as non-NN features).

[0019] Furthermore, the feature extracted by the first extraction unit 152 may be a combination of feature extracted using a plurality of feature extraction models (neural network models), or may be a combination of feature extracted using the feature extraction models and non-NN feature.

[0020] Hereinafter, the n-dimensional feature amounts extracted from the learning image by the first extraction unit 152 will be referred to as the n-dimensional feature amounts of the learning image.

[0021] The first selection unit 153 selects k dimensions (k is an integer equal to or greater than 1 and less than n) from the above-mentioned n dimensions. Information indicating the k dimensions selected by the first selection unit 153 (hereinafter referred to as dimension selection information) is stored in the dimension selection information storage unit 13.

[0022] The learning unit 154 learns an anomaly detection model using the k-dimensional feature amounts selected by the first selection unit 153 from the n-dimensional feature amounts of the learning image (i.e., feature amounts excluding the feature dimensions other than those selected by the first selection unit 153). In other words, the learning unit 154 can generate an anomaly detection model by performing such learning.

[0023] The anomaly detection model corresponds to a neural network model constructed to infer the state of an inspection target (for example, to detect anomalies) by learning k-dimensional features of a training image or the distribution of those features. Specifically, the anomaly detection model includes, for example, a normalizing flow neural network model that converts features to conform to a normal distribution, and an autoencoder that outputs the same features as input features (i.e., reproduces those features). The anomaly detection model may also be a model that applies a method that can detect anomalies by learning image features, such as a one-class support vector machine (SVM).

[0024] The anomaly detection model learned by the learning unit 154 as described above (that is, the anomaly detection model that has learned the k feature amounts of the learning image) is stored in the second model storage unit 14.

[0025] The inference processing unit 16 includes a second acquisition unit 161 , a second extraction unit 162 , a second selection unit 163 , an inference unit 164 , and an output unit 165 .

[0026] The second acquisition unit 161 acquires an image (hereinafter referred to as an inspection image) including an inspection target that is a target for detecting an abnormality (i.e., that needs to be inspected). The inspection image acquired by the second acquisition unit 161 is specified, for example, by a user who uses the information processing device 10. Specifically, for example, when the user specifies a path indicating a location where the inspection image is stored, the second acquisition unit 161 can acquire (read) the inspection image stored in the location indicated by the path. The inspection image may be, for example, an image (data) captured by a camera (imaging device) or an image (data) captured by the scanner.

[0027] The second extraction unit 162 uses the feature extraction model stored in the first model storage unit 12 to extract feature quantities of the test image acquired by the second acquisition unit 161 from the test image. In this case, the second extraction unit 162 extracts n-dimensional feature quantities from the test image. Hereinafter, the n-dimensional feature quantities extracted from the test image by the second extraction unit 162 will be referred to as the n-dimensional feature quantities of the test image.

[0028] The second selection unit 163 selects k dimensions from the above-mentioned n dimensions based on the dimension selection information stored in the dimension selection information storage unit 13. Note that the k dimensions selected by the second selection unit 163 are the same as the k dimensions selected by the above-mentioned first selection unit 153.

[0029] The inference unit 164 infers the state of the inspection object included in the inspection image (detects an abnormality) using the anomaly detection model stored in the second model storage unit 14. This inference of the state of the inspection object is performed based on the output of the anomaly detection model when n-dimensional feature quantities of the inspection image are input to the anomaly detection model.

[0030] The output unit 165 outputs the result of the inference performed by the inference unit 164 (i.e., the abnormality detection result of the test object). Note that the result of the inference performed by the inference unit 164 (hereinafter referred to as the inference result) includes, for example, whether the test object is normal or abnormal.

[0031] When the inference unit 164 infers the state of the inspection object as described above, the inspection image is stored in the image database 11 together with the inference result. In other words, the inspection image stored in the image database 11 in this manner is used as a learning image for learning the anomaly detection model described above. An inspection image stored in the image database 11 together with the inference result that the inspection object is normal corresponds to a normal image, and an inspection image stored in the image database 11 together with the inference result that the inspection object is abnormal corresponds to an abnormal image.

[0032] Fig. 2 shows an example of the hardware configuration of the information processing device 10 shown in Fig. 1. The information processing device 10 includes a CPU 10a, a nonvolatile memory 10b, a main memory 10c, a communication device 10d, and the like.

[0033] The CPU 10a is a processor for controlling the operation of various components within the information processing device 10. The CPU 10a may be a single processor or may be configured with multiple processors. The CPU 10a executes various programs loaded from the non-volatile memory 10b to the main memory 10c. These programs include, for example, an operating system (OS) and application programs.

[0034] The nonvolatile memory 10b is a storage medium used as an auxiliary storage device. The main memory 10c is a storage medium used as a main storage device. Although only the nonvolatile memory 10b and the main memory 10c are shown in FIG. 2, the information processing device 10 may include other storage devices.

[0035] The communication device 10d is a device configured to communicate with an external device (for example, a server device, etc.).

[0036] In this embodiment, the image database 11, the first model storage unit 12, the dimension selection information storage unit 13, and the second model storage unit 14 included in the information processing device 10 shown in FIG. 1 are realized by, for example, a non-volatile memory 10b or other storage device.

[0037] In this embodiment, part or all of the learning processing unit 15 and the inference processing unit 16 included in the information processing device 10 shown in Fig. 1 are realized by causing the CPU 10a (i.e., the computer of the information processing device 10) to execute a predetermined program, that is, by software. This program may be stored in a computer-readable storage medium and distributed, or may be downloaded to the information processing device 10 via a network. Note that part or all of the learning processing unit 15 and the inference processing unit 16 may be realized by hardware such as an integrated circuit (IC), or by a combination of software and hardware.

[0038] Although not shown in FIG. 2, the information processing device 10 may further include an input device including a mouse, a keyboard, and the like, and a display device including a display, and the like.

[0039] Next, a processing procedure of the information processing device 10 according to this embodiment will be described. Here, a process executed by the learning processing unit 15 included in the information processing device 10 (hereinafter referred to as learning process) and a process executed by the inference processing unit 16 (hereinafter referred to as inference process) will be described.

[0040] First, an example of the processing procedure of the above-mentioned learning processing will be described with reference to the flowchart of FIG.

[0041] Here, assuming that a plurality of learning images are stored in the image database 11, in the learning process, the first acquisition unit 151 acquires the plurality of learning images (a group of learning images) from the image database 11 (step S1).

[0042] It is assumed that the multiple learning images acquired in step S1 include at least one normal image. Fig. 4 shows an example of a normal image. In the example shown in Fig. 4, a normal image including, for example, a part in a normal state as an inspection target is shown.

[0043] Furthermore, the plurality of learning images acquired in step S1 may or may not include an abnormal image. Fig. 5 shows an example of an abnormal image. In the example shown in Fig. 5, an abnormal image including, as an inspection target, a part with a chipped part or a part with a scratch on its surface is shown.

[0044] Each of the multiple learning images acquired in step S1 is assigned information (e.g., a label) indicating whether the learning image is a normal image or an abnormal image. Specifically, the learning image is assigned a label of "normal" indicating that the test object is normal or a label of "abnormal" indicating that the test object is abnormal.

[0045] In step S1, all of the plurality of learning images stored in the image database 11 may be acquired, or only a portion of the plurality of learning images may be acquired.

[0046] Next, the first extractor 152 uses the feature extraction model stored in the first model storage unit 12 to extract feature amounts from each of the plurality of learning images acquired in step S1 (step S2).

[0047] Figure 6 shows an overview of the process of extracting features from training images. In Figure 6, a CNN that has trained on a large number of images is used as a feature extraction model, and a single training image 200 is input to the CNN, and a feature map (H x W x n) output from the intermediate layer of the CNN is extracted as a feature. This feature map corresponds to the n-dimensional feature of the training image 200.

[0048] Here, for simplicity of explanation, the output of one hidden layer of the CNN is used as a feature, but a feature that combines the outputs of multiple hidden layers may also be extracted.

[0049] 3, the first selection unit 153 determines whether or not there is an abnormal image among the plurality of learning images acquired in the above-mentioned step S1 (step S3). Whether or not a learning image is an abnormal image can be determined based on the label attached to the learning image.

[0050] As described above, in a field where inspections are performed on products manufactured in a factory or parts used to manufacture the products, defects rarely occur in the inspection objects, and it may be difficult to collect abnormal images for training the anomaly detection model before the anomaly detection model is put into operation. In this case, for example, in a situation where the anomaly detection model is trained before being put into operation, there may be cases where no abnormal images are stored in image database 11 (i.e., no abnormal images have been collected as training images). In this case, in step S3, it is determined that there are no abnormal images among the multiple training images (NO in step S3).

[0051] As described above, if there are no abnormal images among the multiple learning images, the first selection unit 153 randomly selects k dimensions from the n dimensions for which feature amounts were extracted in step S2 (step S4). When the processing of step S4 is executed, dimension selection information indicating the k dimensions selected in step S4 is stored in the dimension selection information storage unit 13.

[0052] Although it has been described here that the k dimensions are selected randomly, the k dimensions may be selected based on (statistics of) the n-dimensional features of each of the plurality of learning images extracted in step S2. Specifically, first selection unit 153 may select k dimensions from the n dimensions that have small variations in feature values ​​among the plurality of learning images.

[0053] When the process of step S4 is executed, the learning unit 154 performs learning of the anomaly detection model using the k feature quantities selected in step S4 from among the n-dimensional feature quantities of the learning image (normal image) (step S5). Note that, assuming that the feature map (H×W×n) is extracted in step S2 as described above, the k-dimensional feature quantities correspond to a feature map (H×W×k) obtained by excluding feature quantities of dimensions other than the k dimensions selected in step S4 from the n-dimensional feature quantities of each learning image (i.e., by dimension reduction).

[0054] In this case, the learning unit 154 calculates the feature distribution of the training image for each spatial element of the dimension-reduced feature map. That is, the learning unit 154 calculates the distribution of k-dimensional features (H×W feature distribution) for each vertical and horizontal element (H×W elements) of the feature map. The learning unit 154 generates an anomaly detection model by learning the feature distribution calculated for each training image.

[0055] In this embodiment, the processing of steps S1 to S5 described above is performed to generate an anomaly detection model that has learned the learning images (features). However, in order to improve the accuracy (anomaly detection accuracy) of the anomaly detection model generated in this way after starting operation of the anomaly detection model, it is preferable to repeatedly perform the processing (learning processing) shown in FIG. 3 even after starting operation of the anomaly detection model.

[0056] Here, we have explained the case where it is determined in step S3 that there are no abnormal images among the multiple learning images. However, when the inference process described below is executed, for example, an inspection image in which it is inferred that the inspection object is abnormal (an abnormal image including an inspection object in which an abnormality has been detected) is stored in the image database 11. Therefore, for example, when the learning process is executed repeatedly, there is a possibility that an abnormal image is stored in the image database 11.

[0057] When the learning process is executed with abnormal images stored in the image database 11 in this way, it is determined in step S3 that there is an abnormal image among the multiple learning images (YES in step S3). In this case, the first selection unit 153 selects k dimensions from the n dimensions based on the feature amounts of the n dimensions of each of the multiple learning images (normal images and abnormal images) extracted in step S2 (step S6). In step S6, for example, the difference in feature amount between a normal image and an abnormal image is calculated for each of the n dimensions, and the k dimensions with the largest calculated difference are selected as feature dimensions to be used for anomaly detection.

[0058] The processing of step S6 will be specifically described below. Here, some of the normal images included in the above-mentioned plurality of learning images are referred to as first normal images, and the other parts of the normal images are referred to as second normal images. In this case, the first selection unit 153 compares, for example, the difference between the feature amount of the first normal image and the feature amount of the second normal image (hereinafter referred to as the first difference) with the difference between the feature amount of the first normal image and the feature amount of the abnormal image (hereinafter referred to as the second difference) in each of n dimensions, and selects k dimensions in which the second difference is larger than the first difference.

[0059] Specifically, for example, the n-dimensional feature amounts of the learning images (first and second normal images and the abnormal image) are assumed to be a feature map of H×W×n, and the feature amount of one of the n dimensions (one feature vector consisting of feature amounts of H×W elements in one dimension of the feature map) is defined as a dimensional feature amount. Note that Fig. 7 conceptually illustrates the dimensional feature amount.

[0060] In this case, the first difference is calculated for each dimension as the average value of the Mahalanobis distance between a feature distribution in H×W dimensions calculated from one dimensional feature of the first normal image and the dimensional feature of the second normal image (hereinafter referred to as the first Mahalanobis distance). Similarly, the second difference is calculated for each dimension as the average value of the Mahalanobis distance between a feature distribution in H×W dimensions calculated from one dimensional feature of the first normal image and the dimensional feature of the abnormal image (hereinafter referred to as the second Mahalanobis distance). The Mahalanobis distance corresponds to a distance calculated taking into account the correlation of data.

[0061] In this case, the first selection unit 153 calculates, for each dimension, the absolute value of the difference between the first Mahalanobis distance and the second Mahalanobis distance as the difference in feature amounts between the normal image and the abnormal image, and selects k dimensions with the largest absolute value. Fig. 8 conceptually illustrates the dimension selection based on the first and second Mahalanobis distances. It can be said that a dimension with a large absolute value of the difference between the first Mahalanobis distance and the second Mahalanobis distance (i.e., distance difference) makes it easier to distinguish between normal images and abnormal images based on feature amounts than a dimension with a small distance difference. Therefore, in this embodiment, a dimension with a large distance difference is selected for learning and inference.

[0062] Here, it has been described that k dimensions are selected, but k may be a constant or may be dynamically determined. If k is a constant, the k dimensions may be selected in descending order of the absolute value of the difference between the first Mahalanobis distance and the second Mahalanobis distance. If k is dynamically determined, all dimensions for which the absolute value of the difference between the first Mahalanobis distance and the second Mahalanobis distance is equal to or greater than a threshold may be selected.

[0063] Furthermore, although the Mahalanobis distance has been described here as being used to calculate the first and second differences, the first and second differences may be calculated using a Euclidean distance or the like that does not take into account the correlation between data.

[0064] Furthermore, the difference in feature amounts between normal and abnormal images may be calculated by other methods. Specifically, for example, the feature amounts of normal and abnormal images may be clustered in each dimension, and the distance between the cluster to which the feature amounts of the normal and abnormal images belong may be used as the difference in feature amounts between the normal and abnormal images.

[0065] When the processing of step S6 is executed, dimension selection information indicating the k dimensions selected in step S6 is stored in the dimension selection information storage unit 13. Note that if other dimension selection information has already been stored in the dimension selection information storage unit 13 because the learning processing has already been executed, the dimension selection information indicating the k dimensions selected in step S6 overwrites the other dimension selection information already stored.

[0066] Next, learning unit 154 performs learning of the anomaly detection model using the k-dimensional feature quantities selected in step S6 from the n-dimensional feature quantities of the learning image (step S5). The processing of step S5 is as described above, and therefore a detailed description thereof will be omitted here. Note that when the processing of step S6 described above is executed, abnormal images are included in the learning images, but the learning images used to learn the anomaly detection model are assumed to be normal images.

[0067] 3 described above (learning process), an anomaly detection model can be generated by performing learning using k-dimensional feature quantities out of n-dimensional feature quantities extracted from learning images including one or more normal images and zero or more abnormal images obtained from image database 11. Furthermore, according to the process shown in Fig. 3, for example, it is possible to perform learning using only normal images before operating the anomaly detection model, and then to update the anomaly detection model using abnormal images after operating the anomaly detection model.

[0068] In the above-described learning process, for example, some of the multiple learning images acquired in step S1 may be used for dimension selection, and other parts of the multiple learning images may be used for training the anomaly detection model. In other words, the learning images used for dimension selection and training the anomaly detection model may be the same, or at least some of the learning images may be different.

[0069] Next, an example of the processing procedure of the above-mentioned inference processing will be described with reference to the flowchart of Fig. 9. Fig. 10 shows an overview of the inference processing shown in Fig. 9.

[0070] In the inference process, the second acquisition unit 161 acquires an inspection image (an image of an abnormality detection target) including an inspection target (step S11). For ease of explanation, step S11 will be described as acquiring one inspection image, but multiple inspection images may be acquired. If multiple inspection images are acquired in step S11, the following steps S12 to S16 may be performed for each of the inspection images.

[0071] Next, the second extraction unit 162 extracts features from the test image acquired in step S11 using a feature extraction model (e.g., CNN) stored in the first model storage unit 12 (step S12). In this case, the test image is input to the feature extraction model, and features output from the feature extraction model (intermediate layer) are extracted. The features extracted from the test image in step S12 are n-dimensional features (e.g., an H×W×n feature map). Note that the processing in step S12 is similar to the processing in step S2 shown in FIG. 3, and therefore a detailed description thereof will be omitted here.

[0072] When the process of step S12 is executed, the second selection unit 163 acquires the dimension selection information stored in the dimension selection information storage unit 13. Based on the acquired dimension selection information (i.e., the k dimensions selected by the first selection unit 153 in the above-described learning process), the second selection unit 163 selects the k dimensions from the n dimensions for which feature amounts have been extracted in step S12 (step S13).

[0073] When the process of step S13 is executed, the inference unit 164 acquires k-dimensional feature quantities selected in step S13 from the n-dimensional feature quantities of the test image extracted in step S12. Note that these k-dimensional feature quantities correspond to a feature map (H×W×k) obtained by excluding (i.e., reducing dimensions) all the n-dimensional feature quantities of the test image except for the k-dimensional feature quantities selected in step S13.

[0074] The inference unit 164 infers the state of the inspection object included in the inspection image by inputting the k-dimensional feature quantities of the acquired inspection image into the anomaly detection model stored in the second model storage unit 14 (step S14). Note that the processing of step S14 corresponds to processing for detecting an anomaly in the inspection object included in the inspection image.

[0075] If the k-dimensional feature quantities of the test image are assumed to be an H×W×k feature map, then in step S14, the anomaly detection model calculates, for each spatial element (each of the H×W elements) of the feature map, the Mahalanobis distance between the feature quantities of the test image and the feature quantity distribution of the normal image learned by the anomaly detection model in the learning process described above. The inference unit 164 determines the maximum value of the Mahalanobis distances for each element calculated in this way as the anomaly score indicating the degree of abnormality of the test object contained in the test image.

[0076] The learning and inference techniques (anomaly detection techniques) described in this embodiment are merely examples, and other techniques may be applied to this embodiment. Specifically, for example, instead of calculating the feature distribution of normal images for each vertical and horizontal element (H × W elements) of the feature map during learning, a technique may be applied in which the feature values ​​of normal images themselves are stored for each element, and during inference, the distance between the feature values ​​of the test image and the feature values ​​of the normal image is calculated, and the greater the distance to nearby feature values, the higher the anomaly score. Furthermore, a technique using One Class SVM may be applied to this embodiment.

[0077] The inference unit 164 stores in advance a threshold value for detecting an abnormality in the test object (a threshold value set as the boundary between normality and abnormality), and can detect an abnormality in the test object by comparing the abnormality score with the threshold value. Specifically, the inference unit 164 determines whether the abnormality score is equal to or greater than the threshold value, and does not detect an abnormality in the test object if the abnormality score is less than the threshold value, and detects an abnormality in the test object if the abnormality score is equal to or greater than the threshold value.

[0078] Here, the maximum value of the Mahalanobis distances for each element described above is described as the abnormality score, but the abnormality score does not necessarily have to be the maximum value of the Mahalanobis distances. Specifically, the Mahalanobis distance is calculated for each element in the spatial direction, and since the Mahalanobis distance can be used to identify areas of the inspection image with a high degree of abnormality (e.g., areas where the inspection object is damaged), the abnormality score may be a statistical value such as the average value of the Mahalanobis distances calculated for each element corresponding to that area. In other words, the abnormality score may be a value calculated from the Mahalanobis distance, for example.

[0079] The anomaly score may be any score suitable for a given anomaly detection method. The anomaly detection model may be constructed to output, for example, the anomaly score described above, or may be constructed to output the state of the test object (whether normal or abnormal).

[0080] When the processing of step S14 is executed, the test image and the inference result of step S14 are stored in image database 11 (step S15). Note that the inference result may include whether the test object is normal or abnormal, and the test image is stored in image database 11 with a label attached according to the inference result, and is used in the learning process executed later.

[0081] Note that, although it is assumed here that a test image is stored in image database 11 each time an inference process is executed, it is not necessary for all of the test images to be stored in image database 11. Specifically, for example, only test images in which an abnormality is detected (i.e., abnormal images) may be stored in image database 11, or only test images with high anomaly scores may be stored in image database 11. Furthermore, only some, rather than all, of test images in which no abnormality is detected (i.e., normal images) may be stored in image database 11. With this configuration, abnormal images, which are relatively difficult to collect, can be preferentially stored in image database 11, and the number of images (number of records) stored in image database 11 can be prevented from becoming enormous.

[0082] When the process of step S15 is executed, the output unit 165 outputs the above-mentioned inference result (step S16). Note that the inference result may be output to the communication device 10d for transmission to a server device or the like external to the information processing device 10, for example, or may be output to a display device (display) for presentation to the user.

[0083] The inference result output in step S16 may include at least whether the test object is normal or abnormal, but may also include, for example, the abnormality score described above, an abnormality score map in which the Mahalanobis distance calculated for each element in the spatial direction of the feature map is assigned to that element as an abnormality score, or a combination of these. Furthermore, in step S16, the inference result described above may be processed and output.

[0084] According to the process (inference process) shown in FIG. 9 described above, the state of the inspection object included in the inspection image can be inferred using the anomaly detection model generated by executing the learning process.

[0085] In FIG. 9, it has been described that the processing of step S16 is executed after the processing of step S15, but the order of the processing of step S15 and the processing of step S16 may be reversed, or the processing of step S15 and the processing of step S16 may be executed in parallel.

[0086] As described above, the information processing device 10 according to this embodiment acquires training images from the image database 11, inputs the acquired training images into a feature extraction model to extract n-dimensional features (first features) of the training images output from the feature extraction model, selects k dimensions from the n dimensions, and performs learning using the selected k-dimensional features from the n-dimensional features of the training images, thereby generating an anomaly detection model used to infer the state of the inspection object (to detect anomalies).

[0087] With the above-described configuration, the information processing device 10 according to this embodiment can learn an anomaly detection model regardless of whether or not the learning images contain abnormal images, thereby realizing efficient learning of the anomaly detection model.

[0088] For example, if there are no abnormal images among the training images obtained from the image database 11, k dimensions may be selected randomly from among the n dimensions, or k dimensions with small variation in features among the training images may be selected from among the n dimensions.

[0089] With this configuration, even if it is not possible to collect abnormal images before the anomaly detection model is put into operation, it is possible to generate an anomaly detection model using only normal images and detect unknown anomalies.

[0090] Furthermore, for example, if there is an abnormal image among the learning images acquired from the image database 11, k dimensions are selected from the n dimensions such that the difference (second difference) between the feature amount of a part of the normal image and the feature amount of the abnormal image is larger than the difference (first difference) between the feature amount of a part of the normal image (first normal image) and the feature amount of the other part of the normal image (second normal image).

[0091] With this configuration, for example, when an abnormal image is collected by operating the anomaly detection model, the anomaly detection model can be updated (re-learned) using the abnormal image, thereby improving the accuracy of the anomaly detection model (the accuracy of detecting anomalies in an inspection target) without changing the method for detecting anomalies in an inspection target.

[0092] As described above, in this embodiment, it is possible to realize flexible learning of an anomaly detection model depending on whether or not there is an abnormal image among the learning images.

[0093] In this embodiment, as described above, the anomaly detection model is trained using features of k dimensions selected from n dimensions, which makes it possible to improve the accuracy of the anomaly detection model compared to a configuration in which the anomaly detection model is trained using features simply extracted from training images.

[0094] Furthermore, the information processing device 10 according to this embodiment acquires an inspection image, inputs the acquired inspection image into a feature extraction model to extract n-dimensional features (second features) of the inspection image output from the feature extraction model, and inputs k-dimensional features selected from the n-dimensional features of the inspection image into an anomaly detection model to infer the state of the inspection object.

[0095] In this embodiment, with such a configuration, it becomes possible to detect an abnormality in an inspection object with high accuracy using the anomaly detection model that has been trained as described above.

[0096] Furthermore, in this embodiment, the inspection images on which the above-described inference has been performed are stored together with the inference results in the image database 11. This makes it possible to collect abnormal images while operating the anomaly detection model, and to re-train (update) the anomaly detection model using the abnormal images as learning images.

[0097] In the present embodiment, the information processing device 10 has been described as including the image database 11, the first model storage unit 12, the dimension selection information storage unit 13, the second model storage unit 14, the learning processing unit 15, and the inference processing unit 16. However, the information processing device 10 may be configured to include only some of the units 11 to 16. Specifically, the information processing device 10 according to the present embodiment may be configured to execute only the learning processing, for example, omitting the inference processing unit 16. Furthermore, the information processing device 10 according to the present embodiment may be configured such that at least some of the image database 11, the first model storage unit 12, the dimension selection information storage unit 13, and the second model storage unit 14 are arranged externally.

[0098] Furthermore, although the present embodiment has been described assuming that the information processing device 10 is a single device, the information processing device 10 may be realized as an information processing system or the like realized by a plurality of devices. Specifically, the present embodiment may be an information processing system including, for example, a learning processing device that executes processing equivalent to the learning processing unit 15 included in the information processing device 10, and an inference processing device (anomaly detection device) that executes processing equivalent to the inference processing unit 16 included in the information processing device 10.

[0099] (Second embodiment) Next, a second embodiment will be described. In this embodiment, detailed descriptions of the same parts as in the first embodiment will be omitted, and the description will focus mainly on the parts that are different from the first embodiment.

[0100] This embodiment differs from the first embodiment described above in that weighting is performed in the dimension directions of k-dimensional feature quantities, and anomaly detection is performed with emphasis on feature dimensions that are more different between normal images and abnormal images.

[0101] Fig. 11 is a block diagram showing an example of the functional configuration of an information processing device according to this embodiment. In Fig. 11, the same parts as those in Fig. 1 are denoted by the same reference numerals, and detailed description thereof will be omitted.

[0102] 11, the information processing device 10 according to this embodiment includes a weight storage unit 17. The learning processing unit 15 included in the information processing device 10 includes a first weighting unit 155. The inference processing unit 16 included in the information processing device 10 includes a second weighting unit 166.

[0103] The first weighting unit 155 determines weights for each of the k dimensions based on the k-dimensional feature quantities (features after dimension reduction) selected by the first selection unit 153 from among the n-dimensional feature quantities of the learning image extracted by the first extraction unit 152, and weights the k-dimensional feature quantities using the determined weights. Note that weighting by the first weighting unit 155 is performed when there is an abnormal image among the learning images.

[0104] Specifically, when k dimensions are selected using the absolute value of the difference between the first and second Mahalanobis distances described in the first embodiment, the magnitude of the absolute value of the difference between the first and second Mahalanobis distances calculated for each of the k dimensions is determined as a weight, and the feature of that dimension is multiplied by the weight determined for that dimension.

[0105] In this case, the learning unit 154 uses the feature amounts of k dimensions weighted by the first weighting unit 155 as described above to learn the anomaly detection model.

[0106] For example, it is possible to prepare an attention network (model) that dynamically identifies noteworthy data, and input k-dimensional features into the attention network, thereby achieving weighting of the k-dimensional features (weighting in the dimension direction).

[0107] Here, it has been explained that weighting is performed when there is an abnormal image among the training images, but if there is no abnormal image among the training images, weighting may not be performed, or equal weighting may be performed on all of the k-dimensional feature quantities.

[0108] The weighting method described here is just an example, and weighting may be performed using other methods.

[0109] The weights determined for the k dimensions as described above (that is, the weights for each dimension) are stored in the weight storage unit 17.

[0110] The second weighting unit 166 weights the k-dimensional feature quantities (features after dimension reduction) selected by the second selecting unit 163 from the n-dimensional feature quantities of the test image extracted by the second extracting unit 162, based on the weights for each dimension stored in the weight storage unit 17. In this case, the inference unit 164 inputs the k-dimensional feature quantities weighted by the second weighting unit 166 as described above into the anomaly detection model, thereby inferring the state of the test object included in the test image (detecting an anomaly).

[0111] Although the functional configuration of the information processing device 10 according to this embodiment has been described above, the hardware configuration of the information processing device 10 is the same as that of the first embodiment described above, and therefore detailed description thereof will be omitted. In this embodiment, the weight storage unit 17 shown in Fig. 11 is realized by, for example, the nonvolatile memory 10b shown in Fig. 2 described above or another storage device.

[0112] The learning process and inference process executed in the information processing device 10 according to this embodiment will be briefly described below with reference to specific examples.

[0113] First, the learning process will be described. The first acquisition unit 151 included in the learning processing unit 15 acquires learning images, as in the first embodiment described above. Here, it is assumed that the learning images include both normal and abnormal images. Next, the first extraction unit 152 uses a CNN that has been trained on a large number of images as a feature extraction model to extract a feature map (H×W×n) output from the intermediate layer of the CNN. Next, the first selection unit 153 selects a dimension in which there is a large difference in feature amount between normal images and abnormal images as the dimension (feature dimension) to be used for anomaly detection. As a result, a feature map (a dimension-reduced H×W×k feature map) is obtained in which dimensions other than the selected k dimensions are excluded.

[0114] Here, the first weighting unit 155 outputs a weighted feature map by inputting a dimension-reduced feature map (H×W×k) into the attention network and multiplying the feature map by the weight obtained as shown in Fig. 12. The attention network is composed of a GAP (Global Averaging Pooling) layer and an FC (Fully Connected) layer, but may be composed of any combination of layers other than these layers.

[0115] It is assumed that the attention network is prepared (generated) through prior training. As shown in Figure 13, training of the attention network is performed by connecting a classifier consisting of a GAP layer and an FC layer to a feature map obtained by multiplying the weights obtained by inputting a dimension-reduced feature map to the attention network, and then classifying normal and abnormal conditions. In other words, the attention network is trained to output weights for each dimension that enable the classifier to correctly distinguish normal and abnormal conditions. This makes it possible to obtain an attention network that has learned the weights for dimensions that make it easy to distinguish normal and abnormal conditions.

[0116] The learning unit 154 uses the feature map of the learning image (normal image) weighted by the first weighting unit 155 to learn the anomaly detection model in the same way as in the first embodiment.

[0117] Next, the inference process will be described. In the inference process of this embodiment, after k dimensions are selected (a dimension-reduced feature map is obtained) as described in the first embodiment, a weighted feature map is obtained by multiplying the feature map by the weights used in the learning process (weights stored in the weight storage unit 17). Note that the weighted feature map may also be obtained by multiplying the dimension-reduced feature map by the weights obtained by inputting the dimension-reduced feature map into the attention network trained as described above. In this embodiment, the feature map obtained by weighting in this way is input into the anomaly detection model, whereby the state of the inspection object is inferred (an inference result is obtained).

[0118] As described above, the information processing device 10 according to this embodiment determines weights for each of the k dimensions based on the feature quantities of the k dimensions of a training image, and weights the feature quantities of the k dimensions of the training image using the determined weights. Note that the weights may be determined based on the first and second Mahalanobis distances (first and second differences) described in the first embodiment, or may be determined using a prepared attention network (i.e., by inputting the feature quantities of the k dimensions of a normal image into the attention network). The attention network is generated by performing training to output weights for each dimension that can distinguish between normal and abnormal.

[0119] In this embodiment, the above-described configuration makes it possible to realize anomaly detection that places emphasis on feature dimensions in which there is a difference in feature amounts between normal images and abnormal images, thereby improving the accuracy of detecting anomalies in the test subject.

[0120] (Third embodiment) Next, a third embodiment will be described. In this embodiment, detailed descriptions of the same parts as those in the second embodiment will be omitted, and the description will focus mainly on the parts that are different from the second embodiment.

[0121] This embodiment differs from the second embodiment described above in that, after weighting of feature amounts in dimension directions, dimensions with small weights are excluded (dimensionality reduction is performed).

[0122] Fig. 14 is a block diagram showing an example of the functional configuration of an information processing device according to this embodiment. In Fig. 14, the same parts as those in Fig. 11 are denoted by the same reference numerals, and detailed description thereof will be omitted.

[0123] 14, the learning processing unit 15 included in the information processing device 10 according to this embodiment includes a first reduction unit 156. Furthermore, the inference processing unit 16 included in the information processing device 10 according to this embodiment includes a second reduction unit 167.

[0124] Here, as explained in the second embodiment, the first weighting unit 155 determines the weights of each of the k dimensions based on the feature quantities of the k dimensions of the training image (normal image), but the first reduction unit 156 reduces (excludes) the dimensions with smaller weights from the feature quantities of the k dimensions.

[0125] In this case, the learning unit 154 uses the feature amounts of the dimensions that have not been reduced by the first reduction unit 156 (weighted feature amounts) to learn the anomaly detection model.

[0126] The second reduction unit 167 reduces (excludes) dimensions with small weights from the feature quantities of the k dimensions for which weights have been determined by the second weighting unit 166.

[0127] In this case, the inference unit 164 infers the state of the inspection object contained in the inspection image by inputting the feature quantities (weighted feature quantities) of the dimensions that have not been reduced by the second reduction unit 167 into the anomaly detection model.

[0128] The learning process and inference process executed in the information processing device 10 according to this embodiment are the same as those in the second embodiment, except that dimensions with small weights are excluded when performing learning and inference as described above. Therefore, detailed description thereof will be omitted here.

[0129] As described above, the information processing device 10 according to this embodiment generates an anomaly detection model by reducing dimensions with low weights from the feature quantities of k dimensions of the training image and learning the feature quantities of the remaining dimensions. Furthermore, the information processing device 10 according to this embodiment performs inference by reducing dimensions with low weights from the feature quantities of k dimensions of the inspection image and inputting the feature quantities of the remaining dimensions into the anomaly detection model.

[0130] In this embodiment, due to such a configuration, the number of feature dimensions used for anomaly detection is smaller than that in the second embodiment described above, and therefore the processing time for the anomaly detection can be shortened.

[0131] (Fourth embodiment) Next, a fourth embodiment will be described. In this embodiment, detailed descriptions of the same parts as those in the first embodiment will be omitted, and the description will focus mainly on the parts that are different from the first embodiment.

[0132] This embodiment differs from the first embodiment described above in that it has a function for correcting labels (inference results) attached to test images stored in an image database as learning images.

[0133] Fig. 15 is a block diagram showing an example of the functional configuration of an information processing device according to this embodiment. In Fig. 15, the same parts as those in Fig. 1 are denoted by the same reference numerals, and detailed description thereof will be omitted.

[0134] 15, the information processing device 10 according to this embodiment includes a correction processing unit 18. The correction processing unit 18 includes a display unit 181, a reception unit 182, and a correction unit 183.

[0135] Display unit 181 displays the test image stored in image database 11 and the label attached to the test image (the inference result for the test object included in the test image) on, for example, a display device.

[0136] Here, the reception unit 182, together with the display unit 181, provides a function equivalent to a GUI (Graphical User Interface), and receives user operations (inputs) for the labels displayed by the display unit 181. Here, the user visually recognizes the inspection object included in the inspection image displayed by the display unit 181, and performs an operation to indicate an appropriate label to be attached to the inspection image.

[0137] Correction unit 183 corrects the label attached to the test image stored in image database 11 based on the user's operation received by reception unit 182. Specifically, for example, if the label attached to the test image is "normal" and a user operation is received indicating that the appropriate label to be attached to the test image is "abnormal," correction unit 183 corrects the label attached to the test image from "normal" to "abnormal." Also, for example, if the label attached to the test image is "abnormal," and a user operation is received indicating that the appropriate label to be attached to the test image is "normal," correction unit 183 corrects the label attached to the test image from "abnormal" to "normal."

[0138] Note that the user may perform an operation to indicate whether the label attached to the test image is correct, rather than an operation to indicate the appropriate label that should be attached to the test image.

[0139] Here, the functional configuration of the information processing device 10 according to this embodiment has been described, but the hardware configuration of the information processing device 10 is the same as that of the first embodiment described above, and therefore a detailed description thereof will be omitted. In this embodiment, part or all of the correction processing unit 18 shown in Fig. 15 may be realized by causing the CPU 10a shown in Fig. 2 described above to execute a predetermined program, that is, may be realized by software, hardware, or a combination of software and hardware.

[0140] The learning process and inference process executed in the information processing device 10 according to this embodiment are the same as those in the first embodiment, and therefore a detailed description thereof will be omitted here. In this embodiment, the process by the correction processing unit 18 (the process of correcting the label attached to the test image) may be executed after the test image is stored in the image database 11 as a result of the execution of the inference process, but before the execution of the learning process using the test image as a learning image.

[0141] As described above, the information processing device 10 according to this embodiment displays an inspection image and a label (inference result), accepts a user operation on the label, and corrects the label based on the accepted user operation. In this embodiment, this configuration makes it possible to prevent dimension selection and anomaly detection model training from being performed using an erroneous inference result (a result of an erroneous determination in anomaly detection).

[0142] According to at least one of the above-described embodiments, it is possible to provide an information processing device, an information processing method, and a program that are capable of realizing efficient learning of an anomaly detection model.

[0143] Although several embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are included within the scope and spirit of the invention, as well as within the scope of the invention described in the claims and their equivalents. [Explanation of symbols]

[0144] 10...information processing device, 10a...CPU, 10b...non-volatile memory, 10c...main memory, 10d...communication device, 11...image database, 12...first model storage unit, 13...dimension selection information storage unit, 14...second model storage unit, 15...learning processing unit, 16...inference processing unit, 17...weight storage unit, 18...correction processing unit, 151...first acquisition unit, 152...first extraction unit, 153...first selection unit, 154...learning unit, 155...first weighting unit, 156...first reduction unit, 161...second acquisition unit, 162...second extraction unit, 163...second selection unit, 164...inference unit, 165...output unit, 166...second weighting unit, 167...second reduction unit, 181...display unit, 182...reception unit, 183...correction unit.

Claims

1. a first acquisition means for acquiring a training image from an image database storing at least one training image including the inspection object; a first extraction means for inputting the training image into a feature extraction model and extracting first feature quantities of n dimensions (n ​​is an integer equal to or greater than 2) from the training image output from the feature extraction model; a selection means for selecting k dimensions (k is an integer equal to or greater than 1 and less than n) from among the n dimensions; a learning means for generating an anomaly detection model to be used for inferring the state of the inspection object by performing learning using the selected k-dimensional first feature amounts among the n-dimensional first feature amounts of the learning image; An information processing device comprising:

2. a second acquisition means for acquiring an inspection image including the inspection object; second extraction means for inputting the test image into the feature extraction model and extracting second feature quantities of n dimensions from the test image output from the feature extraction model; an inference means for inferring a state of the inspection object by inputting the selected k-dimensional second feature quantities from among the n-dimensional second feature quantities of the inspection image into the anomaly detection model; The information processing apparatus according to claim 1 , further comprising:

3. The information processing apparatus according to claim 2 , wherein the test image is stored in the image database together with the inference result as the learning image.

4. 3. The information processing device according to claim 2, wherein the selection means randomly selects k dimensions from the n dimensions when the acquired learning images include one or more normal images including an inspection object in a normal state and no abnormal images including an inspection object in an abnormal state.

5. 3. The information processing device according to claim 2, wherein the selection means, when there are normal images including an inspection object in a normal state and no abnormal images including an inspection object in an abnormal state among the acquired learning images, selects k dimensions out of the n dimensions that have small variations in the first feature between the learning images.

6. The information processing device according to any one of claims 2 to 5, wherein, when the acquired learning images include first and second normal images including an inspection object in a normal state and an abnormal image including an inspection object in an abnormal state, the selection means selects, from the n dimensions, k dimensions having a second difference between the first feature amount of the first normal image and the first feature amount of the abnormal image that is larger than a first difference between the first feature amount of the first normal image and the first feature amount of the second normal image.

7. a first weighting means for determining weights for each of the k dimensions based on first feature amounts of the k dimensions of the learning image, and weighting each of the first feature amounts of the k dimensions of the learning image using the determined weights; second weighting means for weighting the k-dimensional second feature amounts of the test image using the weights; Further comprising: the learning means generates the anomaly detection model by performing learning using the weighted k-dimensional first feature amounts; The inference means performs the inference by inputting the weighted k-dimensional second feature amounts to the anomaly detection model.

7. The information processing device according to claim 6.

8. The information processing apparatus according to claim 7 , wherein the weight is determined based on the first and second differences.

9. The information processing apparatus according to claim 7 , wherein the weights are determined by inputting first feature quantities of k dimensions of the learning images to a prepared attention network.

10. The information processing apparatus according to claim 9 , wherein the attention network is generated by performing learning that outputs weights for each dimension that can distinguish between normality and abnormality.

11. a first reduction means for reducing a dimension having a small weight from the k-dimensional first feature quantities of the learning image; second reduction means for reducing the dimension having the small weight from the k-dimensional first feature amounts of the test image; Further comprising: the learning means generates the anomaly detection model by performing learning using the first feature amount of the unreduced dimension; The inference means performs the inference by inputting the unreduced dimensional second feature amount to the anomaly detection model.

8. The information processing device according to claim 7.

12. a display means for displaying the test image and the inference result; a receiving means for receiving a user's operation regarding the inference result; a correcting means for correcting the inference result based on the user's operation; The information processing apparatus according to claim 3, further comprising:

13. An information processing method executed by an information processing device, acquiring a training image from an image database storing at least one training image including the object to be inspected; inputting the training image into a feature extraction model to extract first feature quantities of n dimensions (n ​​is an integer equal to or greater than 2) of the training image output from the feature extraction model; selecting k dimensions (k is an integer equal to or greater than 1 and less than n) from among the n dimensions; generating an anomaly detection model to be used for inferring the state of the inspection object by performing learning using the selected k-dimensional first feature amounts from among the n-dimensional first feature amounts of the learning image; An information processing method comprising:

14. A program executed by a computer of an information processing device, The computer, acquiring a training image from an image database storing at least one training image including the object to be inspected; inputting the training image into a feature extraction model to extract first feature quantities of n dimensions (n ​​is an integer equal to or greater than 2) of the training image output from the feature extraction model; selecting k dimensions (k is an integer equal to or greater than 1 and less than n) from among the n dimensions; generating an anomaly detection model to be used for inferring the state of the inspection object by performing learning using the selected k-dimensional first feature amounts from among the n-dimensional first feature amounts of the learning image; A program to execute.

Images