US 12,423,771 B2
Multi-scale autoencoder generation method, electronic device and readable storage medium
Wan-Jhen Lee, New Taipei (TW); and Chin-Pin Kuo, New Taipei (TW)
Assigned to HON HAI PRECISION INDUSTRY CO., LTD., New Taipei (TW)
Filed by HON HAI PRECISION INDUSTRY CO., LTD., New Taipei (TW)
Filed on Aug. 26, 2022, as Appl. No. 17/896,650.
Claims priority of application No. 202210654040.2 (CN), filed on Jun. 9, 2022.
Prior Publication US 2023/0401670 A1, Dec. 14, 2023
Int. Cl. G06T 3/40 (2024.01); G06N 3/08 (2023.01); G06T 7/00 (2017.01); G06V 10/74 (2022.01); H04N 19/33 (2014.01); H04N 19/60 (2014.01)
CPC G06T 3/40 (2013.01) [G06N 3/08 (2013.01); G06T 7/0002 (2013.01); G06V 10/761 (2022.01); H04N 19/33 (2014.11); H04N 19/60 (2014.11)] 20 Claims
OG exemplary drawing
 
1. A multi-scale autoencoder generation method applied to an electronic device, the method comprising:
acquiring a plurality of product images and acquiring an annotation of each of the plurality of product images, an image size of each of the plurality of product images being the same and the plurality of product images comprising a plurality of non-defective images;
constructing latent spaces having a plurality of scales;
constructing a plurality of autoencoders according to the plurality of latent spaces and the image size of each of the plurality of product images;
obtaining a plurality of learners by training each of the plurality of autoencoders based on the plurality of non-defective images;
inputting the plurality of product images into each of the plurality of learners and obtaining a plurality of reconstructed images reconstructed by each learner;
obtaining a plurality of detection results by detecting whether each of the plurality of product images has defects according to the plurality of reconstructed images corresponding to each learner;
determining similar images for each learner based on a comparison result between each of the plurality of detection results corresponding to each learner and a corresponding annotation result;
calculating a correct rate of each learner according to the similar images corresponding to each learner, and determining a learner from the plurality of learners as a multi-scale autoencoder according to the correct rate of each learner.