US 12,079,696 B2
Machine learning model training method and device, and expression image classification method and device
Longpo Liu, Shenzhen (CN); Wei Wan, Shenzhen (CN); and Qian Chen, Shenzhen (CN)
Assigned to TENCENT TECHNOLOGY (SHENZHEN) COMPANY LIMITED, Shenzhen (CN)
Filed by Tencent Technology (Shenzhen) Company Limited, Shenzhen (CN)
Filed on Oct. 19, 2022, as Appl. No. 17/969,631.
Application 17/969,631 is a continuation of application No. 16/735,236, filed on Jan. 6, 2020, granted, now 11,537,884.
Application 16/735,236 is a continuation of application No. PCT/CN2018/090676, filed on Jun. 11, 2018.
Claims priority of application No. 201710566325.X (CN), filed on Jul. 12, 2017.
Prior Publication US 2023/0037908 A1, Feb. 9, 2023
This patent is subject to a terminal disclaimer.
Int. Cl. G06N 3/096 (2023.01); G06F 18/214 (2023.01); G06F 18/2415 (2023.01); G06N 3/08 (2023.01); G06N 20/00 (2019.01); G06V 10/764 (2022.01); G06V 10/774 (2022.01); G06V 10/82 (2022.01); G06V 20/00 (2022.01); G06V 40/16 (2022.01)
CPC G06N 20/00 (2019.01) [G06F 18/2155 (2023.01); G06F 18/2415 (2023.01); G06N 3/08 (2013.01); G06N 3/096 (2023.01); G06V 10/764 (2022.01); G06V 10/774 (2022.01); G06V 10/82 (2022.01); G06V 20/35 (2022.01); G06V 40/175 (2022.01)] 13 Claims
OG exemplary drawing
 
1. A machine learning model training method, applied to a computing device having one or more processors and memory storing a plurality of programs to be executed by the one or more processors, the method comprising:
obtaining a trained general-purpose machine learning model according to a general-purpose image training set;
obtaining a representative special-purpose image corresponding to a known classification label;
selecting, via the trained general purpose machine learning model, a sample set of special-purpose images from a special-purpose image library meeting a similarity standard with the representative special-purpose image; and
fine-tuning the general-purpose machine learning model using the sample set of special-purpose images and the known classification label as the corresponding supervision signal, to obtain a special-purpose machine learning model, the fine-tuning comprising:
determining multiple classification labels for the sample set of special-purpose images using the general-purpose machine learning model, each classification label having an associated probability indicating that the sample set of special-purpose images belong to the classification label;
obtaining one of the classification labels having a maximum probability as an intermediate classification result for the known classification label; and
adjusting one or more model parameters of the general-purpose machine learning model by reducing a difference between the intermediate classification result and the known classification label,
the method further comprising:
obtaining an unclassified special-purpose image set comprising unclassified special-purpose images, wherein the unclassified special-purpose image set is determined in accordance with a determination that the special-purpose machine learning model fails in classifying the unclassified special-purpose images;
obtaining, after the unclassified special-purpose images are inputted to the special-purpose machine learning model, image features of corresponding unclassified special-purpose images outputted by an intermediate layer of the special-purpose machine learning model;
performing clustering according to the image features of the unclassified special-purpose images, to obtain a special-purpose image subset;
determining a classification label corresponding to the special-purpose image subset; and
fine-tuning the special-purpose machine learning model according to the special-purpose image subset and the corresponding classification label.