Robust machine learning for imperfect labeled image segmentation
AssigneeThe Chinese University of Hong Kong
InventorsPheng-Ann Heng, Cheng Xue, Qi Dou
Life of this patent
Estimating the term…
Abstract
To improve the performance and accuracy of an image segmentation neural network, a cascaded robust learning framework for the segmentation of noisy labeled images includes two stages: a sample selection stage, and a joint optimization stage with label correction. In the first stage, the clean annotated samples are selected for network updating, so that the influence of noisy sample can be interactively eliminated. In the second stage, the label correction module works together with the joint optimization scheme to revise the imperfect labels. Thus, the training of the whole network is supervised by the corrected labels and the original ones.
Claims20
A computer-implemented method comprising:
- receiving, by a computing system, a set of image samples and original labels for the image samples to train an image segmentation neural network model;
- performing, by the computing system, a sample selection process using a plurality of neural networks to select a subset of image samples from the set of image samples;
- training, by the computing system, the image segmentation neural network model using the subset of image samples;
- performing, by the computing system, a label correction process to generate corrected labels for remaining image samples in the set of image samples that were not selected during the sample selection process;
- and
- training, by the computing system, the image segmentation neural network model using a combination of the corrected labels and the original labels in a joint optimization process.
A computing system comprising:
- processing logic;
- and
- a memory coupled to the processing logic, the memory storing code, which when executed by the processing logic, causes the computing system to perform operations including:
- receiving a set of image samples and original labels for the image samples to train an image segmentation neural network model;
- performing a sample selection process using a plurality of neural networks to select a subset of image samples from the set of image samples;
- training the image segmentation neural network model using the subset of image samples;
- performing a label correction process to generate corrected labels for remaining image samples in the set of image samples that were not selected during the sample selection process;
- and
- training the image segmentation neural network model using a combination of the corrected labels and the original labels in a joint optimization process.
Dependencies are inferred from the claim text.