US 12,260,538 B2
Defect detection in image space
Jierong Cheng, Singapore (SG); Ying Sun, Singapore (SG); Wei Xiong, Singapore (SG); Wenyu Chen, Singapore (SG); Yusha Li, Singapore (SG); and Ying Quan, Singapore (SG)
Assigned to AGENCY FOR SCIENCE, TECHNOLOGY AND RESEARCH, Singapore (SG)
Appl. No. 17/782,015
Filed by AGENCY FOR SCIENCE, TECHNOLOGY AND RESEARCH, Singapore (SG)
PCT Filed Dec. 8, 2020, PCT No. PCT/SG2020/050727
§ 371(c)(1), (2) Date Jun. 2, 2022,
PCT Pub. No. WO2021/118463, PCT Pub. Date Jun. 17, 2021.
Claims priority of application No. 10201911839Y (SG), filed on Dec. 9, 2019.
Prior Publication US 2023/0014823 A1, Jan. 19, 2023
Int. Cl. G06T 7/11 (2017.01); G06T 5/20 (2006.01); G06T 5/70 (2024.01); G06T 7/00 (2017.01); G06T 7/149 (2017.01); G06V 10/34 (2022.01); G06V 10/82 (2022.01)
CPC G06T 7/0004 (2013.01) [G06T 5/20 (2013.01); G06T 5/70 (2024.01); G06T 7/0002 (2013.01); G06T 7/11 (2017.01); G06T 7/149 (2017.01); G06V 10/34 (2022.01); G06V 10/82 (2022.01); G06T 2207/20021 (2013.01); G06T 2207/20076 (2013.01); G06T 2207/20081 (2013.01); G06T 2207/20084 (2013.01)] 17 Claims
OG exemplary drawing
 
1. A method for training a neural network, comprising:
transforming each training image of a plurality of training images by, for each training image:
detecting a hole in the training image; and
transforming the training image into a transformed image, to suppress non-crack information; and
training a neural network using the transformed images, to detect cracks in images,
wherein transforming each training image further comprises, for each training image:
applying a set of filters, the set of filters being designed to enhance image features based on a shape of each said feature, segmenting the training image to produce a segmentation mask, and combining the segmentation mask with responses from the set of filters;
identifying, based on the segmentation mask and the responses from the set of filters, a surrounding region; and
applying a filter to the surrounding region to enhance crack features to produce a filter response, applying Gaussian blur to the surrounding region to produce a Gaussian blurred segmentation mask of the hole, and attributing to the transformed image a maximum of the filter response and Gaussian blurred segmentation mask.