Anomaly detection using image-based physical characterization

• US 10,664,966 B2
• Filed: 01/25/2018
• Issued: 05/26/2020
• Est. Priority Date: 01/25/2018
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Guo, Dechao, Jiang, Liying, Liu, Derrick, Zhang, Jingyun, Zhou, Huimei
• Primary Examiner(s): Dulaney, Kathleen Y
• Application Number:
  US15/879,530
• Publication Number:
  US 20190228519A1
• Time in Patent Office: 852 Days
• Field of Search: 382141, 382159, 382254, 706 45
• CPC Class Codes: G06F 16/285 Clustering or classification
• CPC Class Codes: G06F 18/2433 Single-class perspective, e.g. one-against-all classification; Novelty detection; Outlier detection
• CPC Class Codes: G06F 2218/12 Classification; Matching
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06T 2207/20081 Training; Learning
• CPC Class Codes: G06T 7/0004 Industrial image inspection
• CPC Class Codes: G06T 7/001 using an image reference approach
• CPC Class Codes: G06V 10/443 by matching or filtering
• CPC Class Codes: G06V 10/764 using classification, e.g. of video objects

Abstract

An aspect of the invention includes reading a scale in image data representing an image of physical characteristics and resizing at least a portion of the image data to align with target image data representing a target image based at least in part on the scale to form resized image data representing one or more resized images. Noise reduction is applied to the resized image data to produce test image data representing one or more test images. A best fit analysis is performed on the test image data with respect to the target image data. Test image data having the best fit are stored with training image data representing classification training images indicative of one or more recognized features. An anomaly in unclassified image data representing an unclassified image is identified based at least in part on an anomaly detector as trained using the classification training images.

Claims

1. A computer-implemented method for anomaly detection, the method comprising:

reading, by a processor, a scale in image data representing an image of a plurality of physical characteristics, wherein the scale is determined by reading pixel data of the image data into a two-dimensional matrix, dissecting the two-dimensional matrix to retain a scaling portion of the image data expected to graphically depict scaling information based on a known format, and analyzing the scaling portion of the image data expected to graphically depict scaling information to identify a legend label;

resizing, by the processor, at least a portion of the image data to align with target image data representing a target image of one or more structures based at least in part on the scale to form resized image data representing one or more resized images;

applying, by the processor, noise reduction to the resized image data to produce test image data representing one or more test images;

performing, by the processor, a best fit analysis on the test image data with respect to the target image data to determine a best fit;

storing the test image data of at least one of the test images having the best fit with training image data representing a plurality of classification training images indicative of one or more recognized features; and

identifying an anomaly in unclassified image data representing an unclassified image based at least in part on an anomaly detector as trained using the classification training images.

2. The computer-implemented method of claim 1, wherein the anomaly detector is trained for structural and defect recognition using machine learning based at least in part on the training image data representing the classification training images.

3. The computer-implemented method of claim 1, wherein analyzing the portion of the image data expected to graphically depict scaling information comprises building a training set of data to recognize a plurality of different legend labels.

4. The computer-implemented method of claim 1, further comprising:

extracting, by the processor, a plurality of image cuts from the image data by identifying a central portion of the image and randomly selecting a plurality of image blocks in proximity to the central portion of the image.

5. The computer-implemented method of claim 4, wherein resizing at least a portion of the image data comprises selecting a size of the image blocks to match a pixel count of the target image data.

6. The computer-implemented method of claim 1, wherein the resized image data comprise gray-scale image data representing one or more gray-scale images and applying noise reduction comprises:

performing edge detection with filtering on the gray-scale image data to identify a plurality of features comprising one or more of the physical characteristics that are at least partially observable in the gray-scale image data; and

applying two or more different colors to the plurality of features through re-coloring to convert the resized image data into smoothed color image data to highlight the plurality of features.

7. The computer-implemented method of claim 1, wherein performing the best fit analysis on the test image data with respect to the target image data comprises:

determining a difference value between the test image data and the target image data;

incrementally rotating either the test image data or the target image data and determining the difference value after rotation; and

identifying the best fit as the test image data representing one of the test images having a lowest difference value in comparison to the target image data after rotation.

8. A computer program product for anomaly detection, the computer program product comprising:

a non-transitory computer readable storage medium readable by a processing circuit and storing program instructions for execution by the processing circuit for performing;

reading a scale in image data representing an image of a plurality of physical characteristics, wherein the scale is determined by reading pixel data of the image data into a two-dimensional matrix, dissecting the two-dimensional matrix to retain a scaling portion of the image data expected to graphically depict scaling information based on a known format, and analyzing the scaling portion of the image data expected to graphically depict scaling information to identify a legend label;

resizing at least a portion of the image data to align with target image data representing a target image of one or more structures based at least in part on the scale to form resized image data representing one or more resized images;

applying noise reduction to the resized image data to produce test image data representing one or more test images;

performing a best fit analysis on the test image data with respect to the target image data to determine a best fit;

storing the test image data of at least one of the test images having the best fit with training image data representing a plurality of classification training images indicative of one or more recognized features; and

identifying an anomaly in unclassified image data representing an unclassified image based at least in part on an anomaly detector as trained using the classification training images.

9. The computer program product of claim 8, wherein the anomaly detector is trained for structural and defect recognition using machine learning based at least in part on the training image data representing the classification training images.

10. The computer program product of claim 8, wherein analyzing the portion of the image data expected to graphically depict scaling information comprises building a training set of data to recognize a plurality of different legend labels.

11. The computer program product of claim 8, wherein the program instructions are further executable to cause the processing circuit to:

extract a plurality of image cuts from the image data by identifying a central portion of the image and randomly selecting a plurality of image blocks in proximity to the central portion of the image.

12. The computer program product of claim 8, wherein the resized image data comprise gray-scale image data representing one or more gray-scale images and applying noise reduction comprises:

performing edge detection with filtering on the gray-scale image data to identify a plurality of features comprising one or more of the physical characteristics that are at least partially observable in the gray-scale image data; and

applying two or more different colors to the plurality of features through re-coloring to convert the resized image data into smoothed color image data to highlight the plurality of features.

13. The computer program product of claim 8, wherein performing the best fit analysis on the test image data with respect to the target image data comprises:

determining a difference value between the test image data and the target image data;

incrementally rotating either the test image data or the target image data and determining the difference value after rotation; and

identifying the best fit as the test image data representing one of the test images having a lowest difference value in comparison to the target image data after rotation.

14. A processing system for anomaly detection, comprising:

one or more types of memory; and

at least one processor communicatively coupled with the one or more types of memory, the at least one processor configured to;

read a scale in image data representing an image of a plurality of physical characteristics, wherein the scale is determined by reading pixel data of the image data into a two-dimensional matrix, dissecting the two-dimensional matrix to retain a scaling portion of the image data expected to graphically depict scaling information based on a known format, and analyzing the scaling portion of the image data expected to graphically depict scaling information to identify a legend label;

resize at least a portion of the image data to align with target image data representing a target image of one or more structures based at least in part on the scale to form resized image data representing one or more resized images;

apply noise reduction to the resized image data to produce test image data representing one or more test images;

perform a best fit analysis on the test image data with respect to the target image data to determine a best fit;

store the test image data of at least one of the test images having the best fit with training image data representing a plurality of classification training images indicative of one or more recognized features; and

identify an anomaly in unclassified image data representing an unclassified image based at least in part on an anomaly detector as trained using the classification training images.

15. The processing system of claim 14, wherein the anomaly detector is trained for structural and defect recognition using machine learning based at least in part on the training image data representing the classification training images.

16. The processing system of claim 14, wherein the at least one processor is configured to extract a plurality of image cuts from the image data by identifying a central portion of the image and randomly selecting a plurality of image blocks in proximity to the central portion of the image.

17. The processing system of claim 14, wherein the resized image data comprise gray-scale image data representing one or more gray-scale images and the noise reduction comprises:

performing edge detection with filtering on the gray-scale image data to identify a plurality of features comprising one or more of the physical characteristics that are at least partially observable in the gray-scale image data; and

applying two or more different colors to the plurality of features through re-coloring to convert the resized image data into smoothed color image data to highlight the plurality of features.

18. The processing system of claim 14, wherein the best fit analysis performed on the test image data with respect to the target image data is determined by:

determining a difference value between the test image data and the target image data;

incrementally rotating either the test image data or the target image data and determining the difference value after rotation; and

identifying the best fit as the test image data representing one of the test images having a lowest difference value in comparison to the target image data after rotation.