Granted Patent B2

US 12,456,182 · App. 18/184,911 · Granted Oct 28, 2025

Inventors: Eliyahu Schwartz (Haifa, IL); Leonid Karlinsky (Acton, MA); Sivan Harary (Manof, IL); Assaf Arbelle (Lehvot Haviva, IL)

Assignee: International Business Machines Corporation

G06T7/0002G06N3/0455G06T2207/10024G06T2207/20021G06T2207/20081

Abstract

Claims (34)

1 . A system, comprising a processor to:

randomly mask tokens using different masks to generate different subsets of masked tokens;

process the different subsets of masked tokens via a pretrained masked auto-encoder (MAE) encoder to output intermediate representations;

process the intermediate representations via a pretrained MAE decoder to output reconstructed images; and

compare an input image with the output reconstructed images to generate an anomaly score, wherein generating the anomaly score comprises channel-wise filtering squared error maps with a Gaussian kernel to remove noise.

2 . The system of claim 1 , wherein the processor is to split the input image into non-overlapping patches and flatten the patches into the tokens.

3 . The system of claim 1 , wherein the MAE encoder and the MAE decoder are pretrained on a large collection of general images.

4 . The system of claim 1 , wherein the processor is to fine-tune the pretrained MAE encoder and the pretrained MAE decoder, process the different subsets of masked tokens via the fine-tuned MAE encoder to output the intermediate representations, and process the intermediate representations via the fine-tuned MAE decoder to output the reconstructed images.

5 . The system of claim 1 , wherein generating the anomaly score further comprises summing the squared error maps over three color channels.

6 . The system of claim 1 , wherein generating the anomaly score comprises calculating a mean of a plurality of error maps to generate a single error map.

7 . The system of claim 1 , wherein the anomaly score comprises an image-level anomaly score, and wherein generating the image-level anomaly score comprises calculating a max error of pixel-level anomaly scores of an error map.

8 . The system of claim 1 , wherein the anomaly score comprises a pixel-level anomaly score.

9 . A computer-implemented method, comprising:

randomly masking, via a processor, tokens using different masks to generate different subsets of masked tokens;

processing, via the processor, the different subsets of masked tokens via a pretrained masked auto-encoder (MAE) encoder to output intermediate representations;

processing, via the processor, the intermediate representations via a pretrained MAE decoder to output reconstructed images;

fine-tuning, via the processor, the pretrained MAE encoder and the pretrained MAE decoder, and processing the different subsets of masked tokens via the fine-tuned MAE encoder to output the intermediate representations; and

comparing, via the processor, an input image with the output reconstructed images to generate an anomaly score.

10 . The computer-implemented method of claim 9 , further comprising splitting, via the processor, the input image into non-overlapping patches and flatten the patches into the tokens.

11 . The computer-implemented method of claim 9 , further comprising processing the intermediate representations via the fine-tuned MAE decoder to output the reconstructed images.

12 . The computer-implemented method of claim 11 , wherein fine-tuning the pretrained MAE encoder and the pretrained MAE decoder comprises using an attention mechanism of a transformer to share information between reference tokens and query tokens.

13 . The computer-implemented method of claim 9 , wherein generating the anomaly score comprises channel-wise filtering squared error maps with a Gaussian kernel to remove noise and summing the squared error maps over three color channels.

14 . The computer-implemented method of claim 9 , wherein generating the anomaly score comprises calculating a mean of a plurality of error maps to generate a single error map.

15 . The computer-implemented method of claim 9 , wherein generating the anomaly score comprises calculating a max error of a single error map.

16 . The computer-implemented method of claim 9 , comprising detecting, via the processor, a foreign object in the input image from which the tokens are generated based on the anomaly score.

17 . A computer program product for generating anomaly scores, the computer program product comprising a computer-readable storage medium having program code embodied therewith, the program code executable by a processor to cause the processor to:

randomly mask tokens using different masks to generate different subsets of masked tokens;

process the different subsets of masked tokens via a pretrained masked auto-encoder (MAE) encoder to output intermediate representations;

process the intermediate representations via a pretrained MAE decoder to output reconstructed images;

fine-tune, via the processor, the pretrained MAE encoder and the pretrained MAE decoder by using an attention mechanism of a transformer to share information between reference tokens and query tokens; and

compare an input image with the output reconstructed images to generate an anomaly score.

18 . The computer program product of claim 17 , further comprising program code executable by the processor to split the input image into non-overlapping patches and flatten the patches into the tokens.

19 . The computer program product of claim 17 , further comprising program code executable by the processor to process the different subsets of masked tokens via the fine-tuned MAE encoder to output the intermediate representations, and process the intermediate representations via the fine-tuned MAE decoder to output the reconstructed images.

20 . The computer program product of claim 17 , further comprising program code executable by the processor to generate the anomaly score by channel-wise filtering squared error maps with a Gaussian kernel to remove noise.

Assignments (1)

Continuity (1)

References Cited (34)