Data density-based anomaly detection method that allows data validation

Abstract

The present invention relates to a data density-based anomaly detection method capable of data verification, and more specifically, to a data density-based anomaly detection method capable of data verification that facilitates identifying the reason for a model's judgment and making it easier to improve. To achieve the above objective, the present invention comprises the following steps: inputting an image; dividing the input image into blocks of a predetermined size; dividing each of the divided blocks into a plurality of layers; extracting characteristic data points for each selected layer among the divided layers; integrating the data points extracted for each layer; matching an index for the integrated data points and storing it in a memory bank; performing sampling on the data points matched with the index and storing the sample values ​​formed by the sampling in a memory bank; comparing the sample values ​​stored in the memory bank with the data points extracted from the test image to determine whether the test image is abnormal; and verifying and correcting the result of determining the abnormality using the index.

Description

Technology Field

[0001] The present invention relates to a data density-based anomaly detection method capable of data verification, and more specifically, to a data density-based anomaly detection method capable of data verification that facilitates identifying the reason for a model's judgment and making it easier to improve. Background Technology

[0003] In the era of big data, the technology to detect anomalies in large-scale datasets has become a critical challenge across various fields, including finance, cybersecurity, healthcare, and industrial systems. Anomalies, or outliers, refer to data points or patterns that deviate significantly from the normal range; detecting them is essential for uncovering fraudulent activities, diagnosing system defects, predicting failures, and ensuring system integrity. Such anomaly detection technologies play a major role in diverse sectors, and their importance is steadily growing.

[0004] Existing anomaly detection technologies are primarily based on statistical techniques or machine learning algorithms. Statistical methods define a normal range using data distributions and identify data outside that range as outliers. However, this method has limitations in detecting complex data patterns. Additionally, machine learning techniques utilize both supervised and unsupervised learning. Supervised learning trains models using labeled data to detect anomalies, while unsupervised learning identifies anomalies by analyzing unlabeled data. However, these methods also struggle to perform well on large-scale datasets and face challenges in tailoring models to specific data characteristics.

[0005] Furthermore, AI-based anomaly detection technologies are attempting to detect more complex patterns by utilizing advanced techniques such as deep learning and reinforcement learning. However, these methods suffer from a lack of interpretation regarding why the model made a specific prediction. This is known as the "black box" problem, which makes it difficult to identify the cause or suggest methods for improvement when an anomaly detection system makes an incorrect decision. This lack of interpretation undermines the reliability of the system and prevents the provision of feedback for practical improvement.

[0006] Therefore, interpretability of analysis results is essential in AI-based anomaly detection technology. Error analysis and improvement become possible only when the decisions made by the model can be understood and the reasons behind them clearly identified. The process of analyzing the causes of erroneous judgments and improving the model based on this analysis is crucial for enhancing the accuracy of anomaly detection systems. Against this backdrop, there is a need for technology that increases the interpretability of analysis results and enables improvements through error analysis. Prior art literature

[0008] U.S. Registered Patent 9,812,791 The problem to be solved

[0009] The objective of the present invention to solve the above-mentioned problem is to provide a data density-based abnormality detection method capable of data verification to facilitate identifying the reason for the model's judgment and improving it.

[0010] The technical problems that the present invention aims to solve are not limited to those mentioned above, and other unmentioned technical problems will be clearly understood by those skilled in the art to which the present invention belongs from the description below. means of solving the problem

[0012] The present invention provides a data density-based abnormality detection method capable of data verification, characterized by comprising the following steps to achieve the above-mentioned purpose: a step of inputting an image; a step of dividing the input image into blocks of a certain size; a step of dividing each of the divided blocks into a plurality of layers; a step of extracting a data point characteristic of each selected layer among the divided layers; a step of integrating the data points extracted for each layer; a step of matching an index for the integrated data points and storing it in a memory bank; a step of performing sampling on the data points matched with the index and storing a sample value formed by the sampling in a memory bank; a step of determining whether there is an abnormality in the test image by comparing the sample value stored in the memory bank with the data point extracted from the test image; and a step of verifying and correcting the result of determining the abnormality using the index.

[0013] In an embodiment of the present invention, in the step of dividing each of the divided blocks into a plurality of layers, the block may be characterized by being divided into: a first layer including edges, lines, points, and color changes; a second layer including patterns and textures of an object; a third layer including partial shapes and structures of an object; and a fourth layer including the entire shapes and structures of an object.

[0014] In an embodiment of the present invention, in the step of extracting data points for each selected layer among the divided layers, the second layer and the third layer may be selected so that data points are extracted.

[0015] In an embodiment of the present invention, in the step of integrating the data points extracted by each layer, the data points extracted from the second layer and the third layer may be integrated through an average pooling operation.

[0016] In an embodiment of the present invention, the step of sampling the data point matched to the index and storing the sample value formed by the sampling in a memory bank may be characterized by comprising: a step of selecting any data point as an initial starting point; a step of calculating the distance between the initial starting point and other data points; a step of selecting the data point calculated to be the farthest from the initial starting point as a new starting point; a step of repeating the process of selecting the farthest data point by repeatedly calculating the distance between the new starting point and other data points until a predetermined number of data points are selected; and a step of storing the sample value composed of the selected data points in the memory bank.

[0017] In an embodiment of the present invention, the step of determining whether there is an abnormality in the test image by comparing the sample value stored in the memory bank with the data point extracted from the test image may be characterized by comprising: a step of extracting a data point that is characteristic of the test image; a step of calculating a distance-based score by calculating the distance between the data point extracted from the test image and the data point included in the sample value stored in the memory bank that corresponds to the distance-based score; and a step of determining whether there is an abnormality in the test image using the distance-based score.

[0018] In an embodiment of the present invention, in the step of making a judgment on the test image by comparing the sample value stored in the memory bank with the data point extracted from the test image, the test image may be characterized as having already been determined to be abnormal.

[0019] In an embodiment of the present invention, the step of verifying and correcting the result of determining abnormality using the index may be characterized by including: a step of verifying that the normal test image is over-checked when it is determined to be abnormal; a step of searching for the index of a data point having a high absolute value among the distance-based scores; and a step of correcting the data point corresponding to the index.

[0020] In an embodiment of the present invention, the step of verifying and correcting the result of determining abnormality using the index may be characterized by including: a step of verifying that the abnormal test image is not detected when it is determined to be normal; a step of searching for an index of a data point of the sample value for an abnormal part of the test image that is already known; and a step of correcting the data point corresponding to the searched index.

[0021] In an embodiment of the present invention, the distance-based score may be characterized by being calculated as the minimum distance value of the Euclidean distance or L2 distance between the data point of the test image and the data point stored in the memory bank. Effects of the invention

[0023] The effect of the present invention according to the above configuration is that even when an incorrect judgment is made, the model's training data can be modified by checking which part of the memory bank is likely to have been incorrectly judged through distance-based scores and indices.

[0024] The effects of the present invention are not limited to the effects described above, and should be understood to include all effects that can be inferred from the configuration of the invention described in the detailed description of the invention or the claims. Brief explanation of the drawing

[0026] FIG. 1 is a flowchart of a verifiable data density-based abnormal detection method according to an embodiment of the present invention. FIG. 2 is an example diagram of the step of dividing blocks of an image and dividing them into layers according to an embodiment of the present invention. FIG. 3 is a flowchart of the steps in which sampling is performed on data points matched to an index according to an embodiment of the present invention, and sample values ​​formed by sampling are stored in a memory bank. FIG. 4 is a flowchart of a step in which a test image is determined to have an abnormality by comparing a sample value stored in a memory bank according to an embodiment of the present invention with a data point extracted from a test image. FIG. 5 is an example diagram of a step in which a test image is determined to have an abnormality by comparing a sample value stored in a memory bank according to an embodiment of the present invention with a data point extracted from a test image. FIG. 6 is a flowchart showing the process when an over-inspection occurs during the step of verifying and correcting the result of determining whether an abnormality is found using an index according to an embodiment of the present invention. FIG. 7 is a flowchart showing the process when a failure to detect occurs during the step of verifying and correcting the result of determining whether an abnormality is detected using an index according to an embodiment of the present invention. FIGS. 8 to 10 are exemplary diagrams of the steps for verifying and correcting the result of determining whether an abnormality is found using an index according to an embodiment of the present invention. FIG. 11 is an example diagram showing an embedding area of ​​a step in which verification and correction of an abnormality determination result are performed using an index according to an embodiment of the present invention. Specific details for implementing the invention

[0027] The present invention will be described below with reference to the attached drawings. However, the present invention can be implemented in various different forms and is therefore not limited to the embodiments described herein. Furthermore, in order to clearly explain the present invention in the drawings, parts unrelated to the explanation have been omitted, and similar parts throughout the specification have been given similar reference numerals.

[0028] Throughout the specification, when it is stated that a part is "connected (connected, in contact, combined)" with another part, this includes not only cases where they are "directly connected," but also cases where they are "indirectly connected" with other members interposed between them. Furthermore, when it is stated that a part "includes" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but rather allows for the inclusion of additional components.

[0029] The terms used herein are merely for describing specific embodiments and are not intended to limit the invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this specification, terms such as “comprising” or “having” are intended to indicate the presence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0030] Additionally, terms such as "...part," "...unit," and "...module" described in the specification refer to a unit that processes at least one function or operation, and this may be implemented in hardware, software, or a combination of hardware and software.

[0031] Additionally, in this specification, when a step is described as being located "before" or "after" another step, this includes not only cases where a step is in a direct chronological relationship with another step, but also cases where there is an indirect chronological relationship in which the chronological order of the two steps may change, such as a mixing step following each step.

[0032] Embodiments of the present invention will be described in detail below with reference to the attached drawings.

[0033] FIG. 1 is a flowchart of a verifiable data density-based abnormal detection method according to an embodiment of the present invention, and FIG. 2 is an example diagram of a step of dividing blocks of an image and dividing them into layers according to an embodiment of the present invention.

[0034] Referring to FIGS. 1 and 2, a verifiable data density-based abnormality detection method may include the steps of: inputting an image (S100); dividing the input image into blocks of a certain size (S200); dividing each divided block into multiple layers (S300); extracting a characteristic data point for each selected layer among the divided layers (S400); integrating the data points extracted for each layer (S500); matching an index for the integrated data points and storing it in a memory bank (S600); sampling of the data points matched with the index and storing the sample values ​​formed by the sampling in a memory bank (S700); comparing the sample values ​​stored in the memory bank with the data points extracted from the test image to determine whether there is an abnormality in the test image (S800); and verifying and correcting the result of determining the abnormality using the index (S900).

[0035] Below, each step will be explained in detail.

[0036] First, the step (S100) of inputting an image can be arranged so that an image (10) for training is input into a data density-based abnormal detection model.

[0037] At this time, the image (10) may be configured to be converted into a format optimized for the model, such as RGB or Grayscale.

[0038] In addition, if the size of the image (10) is not constant, it may be arranged to be resized or padded to fit the model.

[0039] Next, in the step (S200) where the input image is divided into blocks of a certain size, the image (10) may be divided into blocks (20) of a certain size.

[0040] At this time, the block (20) may be provided as a single pixel. However, the size of the block (20) is not limited thereto and may be provided in units of 32x32 pixels or 64x64 pixels.

[0041] Next, in the step (S300) where each divided block is divided into multiple layers, each of the divided blocks (20) may be divided into multiple layers (30).

[0042] At this time, the block (20) may be arranged to be divided into four layers.

[0043] Specifically, the first layer may be configured to include basic information, such as edges, lines, points, and color changes, as low-level features.

[0044] The second layer can be configured to include the pattern and texture structure of an object that combines low-level features as an intermediate feature.

[0045] The third layer is a high-level feature and may be configured to include information that can identify an object, such as the partial shape or structure of the object.

[0046] The fourth layer relates to the final feature and may be configured to include the overall shape and structure of the object.

[0047] In the step (S400) where characteristic data points are extracted for each selected layer among the divided layers, the second layer and the third layer may be selected to extract data points.

[0048] Specifically, the aforementioned Layer 1 is limited in classification and identification because it includes only very simple features. Furthermore, since the aforementioned Layer 4 includes too many high-dimensional features by including full object information, there is a high possibility that specific details will be lost.

[0049] In comparison, layers 2 and 3 strike a balance between low-level features (lines, points) and high-level features (object semantics), and contain information useful for classification or anomaly detection.

[0050] In particular, the above-mentioned Layer 2 effectively represents structural information such as patterns and textures, thereby providing basic clues that constitute an object. Additionally, the above-mentioned Layer 3 learns the partial shape and structural features of an object, making it effective for distinguishing objects.

[0051] Therefore, combining the above 2nd layer and the above 3rd layer allows for the extraction of more powerful and useful features.

[0052] In addition, if only the above-mentioned 2nd layer and the above-mentioned 3rd layer are selected, the amount of features is appropriately limited, which can improve the calculation speed, while at the same time, important information is preserved, which can ensure accuracy.

[0053] Conversely, including up to layer 4 requires more complex calculations, and layer 1 may be inefficient due to a lot of unnecessary details.

[0054] Accordingly, in the step (S400) where data points that are characteristic of each selected layer among the divided layers are extracted, the above-mentioned 2nd layer and the above-mentioned 3rd layer may be selected to extract data points that are characteristic of each layer.

[0055] At this time, the above data points can be quantified and extracted as vector values, etc.

[0056] Next, in the step (S500) where data points extracted by layer are integrated, the data points extracted from the second layer and the third layer may be integrated through an average pooling operation.

[0057] For example, it may be prepared by averaging the values ​​of the extracted data points and combining them into a single value. For example, if the average value of the data points [1, 2, 3, 4] in the second layer is 2.5 and the average value of the data points [5, 6, 7, 8, 9] in the third layer is 7, then the average values ​​are combined to get a final pooling result of 4.75.

[0058] In this way, the values ​​of data points extracted from each layer can be merged under preset conditions to obtain a value that integrates the data features extracted from each layer.

[0059] Next, in the step (S600) where an index for integrated data points is matched and stored in a memory bank, the index may be arranged to be matched for each integrated data point. At this time, the index may be provided in the form of an embedding index so as to distinguish the location of the data point among adjacent images.

[0060] FIG. 3 is a flowchart of the steps in which sampling is performed on data points matched to an index according to an embodiment of the present invention, and sample values ​​formed by sampling are stored in a memory bank.

[0061] Referring to FIG. 3, the step (S700) of sampling for data points matched to an index and storing the sample values ​​formed by sampling in a memory bank may include: a step (S710) in which 10 arbitrary data points are selected and the average of the distances between each data point and the remaining points is calculated and the point with the greatest distance is selected as the starting point; a step (S720) in which the distance between the initial starting point and other data points is calculated; a step (S730) in which the data point calculated to be the furthest from the initial starting point is selected as the new starting point; a step (S740) in which the process of selecting the furthest data point by repeatedly calculating the distance between the new starting point and other data points is repeated until a preset number of data points are selected; and a step (S750) in which the sample values ​​consisting of the selected data points are stored in a memory bank.

[0062] First, in the step (S710) where a starting point is selected, a random data point may be selected from among all data points integrated in the preceding step and stored in the memory bank. This selection may be based on random sampling or specific preset conditions, such as a point close to the median value of the data points.

[0063] For example, 10 random data points may be selected, and each data point may be averaged with the remaining data points, and the data point that is farthest from the others may be selected as the initial starting point.

[0064] Next, in the step (S720) where the distance between the initial starting point and other data points is calculated, the distance between the initial starting point and all data points may be calculated. Then, this calculated distance value may be used to select a new starting point in the next step. At this time, the distance may be calculated using the Euclidean distance formula, the Manhattan distance formula, the cosine similarity formula, etc.

[0065] Next, in the step (S730) where the data point calculated to be the farthest from the initial starting point is selected as the new starting point, the data point having the largest value among the distances calculated in the step (S720) where the distance between the initial starting point and other data points is calculated may be selected as the new starting point. This process can contribute to ensuring that the data points are evenly distributed.

[0066] Next, in the step (S740) where the process of repeatedly calculating the distance between a new starting point and other data points and selecting the farthest data point is repeated until a preset number of data points are selected, the distance between the newly selected starting point and all data points within the memory bank may be recalculated. This process may be used to re-evaluate the data points within the memory bank and select the next starting point. Then, the data point having the largest value among the calculated distances may be selected again as the new starting point. In this way, until the number of selected data points reaches a preset number, the process of calculating the distance between the new starting point and the data points within the memory bank and selecting the data point calculated to be the farthest distance as the new starting point may be repeated. When the number of selected data points reaches a preset number through this process, the selected data points may be arranged to form sample values.

[0067] Next, in the step (S750) where a sample value consisting of selected data points is stored in a memory bank, the sample value derived in this way can be stored in a memory bank.

[0068] The present invention can be configured to perform learning by storing sample values ​​in a memory bank through the aforementioned process.

[0069] FIG. 4 is a flowchart of a step in which a test image is determined to have an abnormality by comparing a sample value stored in a memory bank and a data point extracted from a test image according to an embodiment of the present invention, and FIG. 5 is an example diagram of a step in which a test image is determined to have an abnormality by comparing a sample value stored in a memory bank and a data point extracted from a test image according to an embodiment of the present invention.

[0070] Referring to FIGS. 4 and 5, the step (S800) of determining whether there is an abnormality in the test image by comparing the sample value stored in the memory bank with the data point extracted from the test image may include the step (S810) of extracting a characteristic data point from the test image, the step (S820) of calculating a distance-based score by calculating the distance between the data point extracted from the test image and the data point included in the sample value stored in the memory bank that corresponds to it, and the step (S830) of determining whether there is an abnormality in the test image using the distance-based score.

[0071] In the step (S810) where feature data points are extracted from the test image, the method may be configured to extract feature data points for the test image.

[0072] At this stage, the above test image is for training purposes and may be an image for which anomalies have already been determined.

[0073] In addition, the step (S810) of extracting characteristic data points from the above test image can be prepared as integrated data points by forming a plurality of blocks and layers as described above, extracting data from layers 2 and 3, and then pooling.

[0074] In the step (S820) where a distance-based score is calculated by calculating the distance between a data point extracted from a test image and a data point included in a sample value stored in a corresponding memory bank, the distance between the data point extracted from the test image and the data point of the corresponding sample value can be calculated to calculate the distance-based score.

[0075] The calculation of the distance-based score described above may be arranged to be calculated as the minimum distance value of the Euclidean distance or L2 distance between the data point of the test image and the data point stored in the memory bank. Here, the L2 distance may refer to the same concept as the Euclidean distance.

[0076] And, in the step (S820) where a distance-based score is calculated by calculating the distance between a data point extracted from a test image and a data point included in a sample value stored in a corresponding memory bank, a score map for the distance-based score can be formed as shown in FIG. 5. The score map thus prepared can be prepared in a form that includes index values.

[0077] The step (S830) of determining whether there is an anomaly in a test image using a distance-based score can determine whether there is an anomaly in the test image using a distance-based score. Specifically, both the data points extracted from the test image and the data points of the sample values ​​may be provided as vector values ​​for features, etc. Therefore, the degree of similarity can be checked based on the magnitude of the distance-based score value. For example, it may be configured to determine that an anomaly has occurred if the distance-based score value has a magnitude greater than or equal to a preset value.

[0078] FIG. 6 is a flowchart showing the process when an over-inspection occurs during the step of verifying and correcting the result of determining whether an abnormality is determined using an index according to an embodiment of the present invention, and FIG. 7 is a flowchart showing the process when an under-inspection occurs during the step of verifying and correcting the result of determining whether an abnormality is determined using an index according to an embodiment of the present invention.

[0079] FIGS. 8 to 10 are exemplary diagrams of a step in which verification and correction of an abnormality determination result are performed using an index according to an embodiment of the present invention, and FIG. 11 is an exemplary diagram showing an embedding area of ​​a step in which verification and correction of an abnormality determination result are performed using an index according to an embodiment of the present invention.

[0080] Referring to FIGS. 6 to 11, in the step (S900) where verification and correction of the abnormality determination result is performed using an index, it may be arranged to determine whether the abnormality determination result of the test image is the same as that known in advance.

[0081] At this time, if the above test image is determined to be abnormal, a step (S910) of verifying that it is overtested may be performed.

[0082] In other words, if a normal test image is determined to be abnormal, the distance-based score can be seen as having been calculated excessively high.

[0083] For example, as shown in Fig. 8, the distance-based score may be calculated to be excessively high even though there are actually no defects in the product's test image due to light reflection, etc.

[0084] Next, in the step (S920) where the index of a data point having a high absolute value among the distance-based scores is searched, the index of a data point having a high absolute value among the distance-based scores in the score map may be searched. For example, as shown in FIG. 9, the data point may be found in the memory bank using the value of the index according to the location.

[0085] Next, in the step (S930) of correcting the data point corresponding to the index, the data point corresponding to the index that appears as a high value can be removed from the memory bank, or a method can be taken to correct for light reflection so that the same problem does not occur.

[0086] On the other hand, if the above test image is determined to be normal, a step (S950) of verifying that it is not inspected may be performed.

[0087] In this case, it can be seen that a distance-based score was assigned low to the parts of the partially defective image. Therefore, it may be determined to be uninspected and configured to perform verification.

[0088] For example, as shown in Fig. 10, a distance-based score may be calculated low even though there are defective parts in the product.

[0089] Next, a step (S960) may be performed in which the index of the data point of the sample value corresponding to the abnormal part of the above-mentioned test image, which is already known, is searched. That is, the index corresponding to the abnormal part may be searched to check whether the value of the data point corresponding to that index has been set low.

[0090] Next, in the step (S970) where the data point corresponding to the searched index is corrected, if the distance-based score value of the data point corresponding to the index is calculated low despite having defects, it may be arranged to perform correction through retraining, etc.

[0091] As such, even when an incorrect judgment is made, the present invention allows for easy modification of the model's training data by identifying where the incorrect judgment may have occurred through distance-based scores and indices.

[0092] Furthermore, a model trained sufficiently through test images in this way can effectively and accurately determine whether an actual product is defective.

[0093] In addition, as shown in FIG. 11, when the embeddings are excessively concentrated on one side, the present invention makes it easy to remove data points formed close to each other within a certain distance and distribute the data points so that they can be learned in the less selected parts.

[0094] Although the foregoing description of the present invention has been illustrated with reference to the drawings, it is for illustrative purposes only, and those skilled in the art will understand that other specific forms can be easily modified without altering the technical spirit or essential features of the invention. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive. For example, each component described as a single unit may be implemented in a distributed manner, and components described as distributed may likewise be implemented in a combined form. Furthermore, the described techniques may be performed in a different order than the described method.

[0095] The embodiments described in this specification and the accompanying drawings are merely illustrative of some of the technical ideas included in the present invention. Accordingly, the scope of the present invention is defined by the claims set forth below, and all modifications or variations derived from the meaning and scope of the claims and equivalent concepts thereof should be interpreted as being included within the scope of the present invention. Explanation of the symbols

[0097] 10: Image 20: Block 30: Layer

Claims

Claim 1 A data density-based abnormality detection method capable of data verification, characterized by comprising: a step of inputting an image; a step of dividing the input image into blocks of a certain size; a step of dividing each of the divided blocks into a plurality of layers; a step of extracting a characteristic data point for each selected layer among the divided layers; a step of integrating the data points extracted for each layer; a step of matching an index for the integrated data points and storing it in a memory bank; a step of performing sampling on the data points matched with the index and storing a sample value formed by the sampling in a memory bank; a step of comparing the sample value stored in the memory bank with the data points extracted from the test image to determine whether there is an abnormality in the test image; and a step of verifying and correcting the result of determining the abnormality using the index. Claim 2 A data density-based abnormality detection method capable of data verification, characterized in that, in the step of dividing each of the divided blocks into a plurality of layers, the block is divided into: a first layer including edges, lines, points, and color changes; a second layer including patterns and textures of an object; a third layer including partial shapes and structures of an object; and a fourth layer including the entire shapes and structures of an object. Claim 3 A data density-based abnormality detection method capable of data verification, characterized in that, in the step of extracting data points for each selected layer among the divided layers, the second layer and the third layer are selected to extract data points. Claim 4 A data density-based abnormality detection method capable of data verification, characterized in that, in the step of integrating the data points extracted by each layer, the data points extracted from the second layer and the third layer are integrated through an average pooling operation. Claim 5 A data density-based abnormality detection method capable of data verification, characterized in that, in claim 1, the step of sampling the data point matched to the index and storing the sample value formed by the sampling in a memory bank comprises: a step of selecting an arbitrary data point as an initial starting point; a step of calculating the distance between the initial starting point and other data points; a step of selecting the data point calculated to be the farthest from the initial starting point as a new starting point; a step of repeating the process of selecting the farthest data point by repeatedly calculating the distance between the new starting point and other data points until a preset number of data points are selected; and a step of storing the sample value composed of the selected data points in the memory bank. Claim 6 A data density-based abnormality detection method capable of data verification according to claim 1, wherein the step of determining whether there is an abnormality in the test image by comparing the sample value stored in the memory bank with the data point extracted from the test image comprises: a step of extracting a characteristic data point from the test image; a step of calculating a distance-based score by calculating the distance between the data point extracted from the test image and the data point included in the sample value stored in the memory bank that corresponds to the distance-based score; and a step of determining whether there is an abnormality in the test image using the distance-based score. Claim 7 A data density-based abnormality detection method capable of data verification, characterized in that, in the step of making a judgment on the test image by comparing the sample value stored in the memory bank with the data point extracted from the test image, the test image has already been determined to be abnormal. Claim 8 A data density-based abnormality detection method capable of data verification, characterized in that, in claim 7, the step of verifying and correcting the result of determining abnormality using the index comprises: a step of verifying that the normal test image is over-checked when it is determined to be abnormal; a step of searching for the index of a data point having a high absolute value among the distance-based scores; and a step of correcting the data point corresponding to the index. Claim 9 A data density-based abnormality detection method capable of data verification, characterized in that, in claim 7, the step of verifying and correcting the result of determining abnormality using the index comprises: a step of verifying that the abnormal test image is not detected when it is determined to be normal; a step of searching for an index of a data point of the sample value for an abnormal part of the test image that is already known; and a step of correcting the data point corresponding to the searched index. Claim 10 A data density-based abnormality detection method capable of data verification, characterized in that, in claim 6, the distance-based score is calculated as the minimum distance value of the Euclidean distance or L2 distance between the data point of the test image and the data point stored in the memory bank.

Images