Training and distillation methods for image segmentation models, electronic devices and storage media
CN115565026BActive Publication Date: 2026-05-26SUZHOU MEGAROBO TECH CO LTD
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SUZHOU MEGAROBO TECH CO LTD
- Filing Date
- 2022-09-09
- Publication Date
- 2026-05-26
AI Technical Summary
Technical Problem
Existing image segmentation models suffer from high false negative rates and low crossover ratios in imbalanced sample and small target scenarios.
Method used
We employ both asymmetric and symmetric pre-defined loss functions, and differentiate the processing for different categories by adjusting the pre-defined logits and decay parameters. This enhances the prediction values of rare classes and reduces the weights of non-rare classes, while also optimizing model training through knowledge distillation.
Benefits of technology
It improves the generalization ability of image segmentation models in imbalanced sample and small target scenarios, reduces the false negative rate and false positive rate, and is suitable for segmentation of small targets.
✦ Generated by Eureka AI based on patent content.
Smart Images
Figure CN115565026B_ABST
Abstract
This invention provides a training and knowledge distillation method for an image segmentation model, an electronic device, and a storage medium. The training method includes: acquiring training sample images; inputting the training sample images into an image segmentation model, and iteratively training the image segmentation model based on a first preset loss function; wherein the first preset loss function includes the sum of the following cross-entropies: the cross-entropy corresponding to the first category: calculated based on the predicted value of the first category under a preset logits adjustment parameter; the cross-entropy corresponding to the second category with a decay factor: the decay factor non-linearly decays (1 - the predicted value of the second category) under a first preset decay parameter. First, it helps solve problems such as imbalanced samples and overfitting of the image segmentation model due to limited data, making it suitable for small-sample learning; second, it is very suitable for segmenting small targets, especially for small targets that require consideration of the global field of view, reducing false positives and false negatives.
Need to check novelty before this filing date? Find Prior Art