Adaptive Pseudo-Label Generation for Balanced Image Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing semi-supervised object detection methods suffer from imbalanced learning due to the generation of pseudo labels that favor easy categories over difficult categories, leading to poor classification accuracy and error accumulation.
Innovation Solution
A label generation method that dynamically evaluates the learning status of each category by a teacher and student model, determining a probability threshold for each category to generate pseudo labels, thereby alleviating the imbalance and improving the training effect of the image classification model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If pseudo labels are generated using a fixed threshold for all categories, then the generation process is simple and fast, but the learning becomes imbalanced with easy categories favored over difficult categories
Solution Approach 1:
The patent applies local quality by differentiating the probability threshold for each category based on its learning difficulty. Instead of using a uniform threshold, the system dynamically adjusts thresholds locally for each category, allowing easy categories to have higher thresholds and difficult categories to have lower thresholds, thus achieving balanced learning across different category types
Solution Approach 2:
The patent implements dynamics by making the probability threshold adaptive rather than static. The threshold for each category is dynamically adjusted during training based on the student model's learning status for that category, enabling the system to respond to changing learning conditions and maintain balanced learning throughout the training process
2Measurement precision
If manual annotation is used to obtain labeled data, then the data quality is high and accurate, but the process is time-consuming and costly
Solution Approach 1:
The patent applies self-service by enabling the student model to generate its own training labels (pseudo labels) from unlabeled data through semi-supervised learning. The model learns to identify detection objects and categories autonomously without requiring manual annotation, thus eliminating the time-consuming and costly manual labeling process while still achieving acceptable data quality for training
Solution Approach 2:
The patent uses a composite approach by combining a small amount of manually annotated labeled data with a large amount of unlabeled data in a semi-supervised learning framework. This composite data strategy leverages the high quality of manual labels for initialization while utilizing the abundance of unlabeled data for scalable training, reducing overall annotation requirements
3Device complexity
If a single probability threshold is used for all categories, then the implementation is simple, but the learning balance across categories deteriorates
Solution Approach 1:
The patent applies local quality by differentiating the probability threshold for each category based on its learning difficulty. Instead of using a uniform threshold, the system dynamically adjusts thresholds locally for each category, allowing easy categories to have higher thresholds and difficult categories to have lower thresholds, thus achieving balanced learning across different category types
Solution Approach 2:
The patent implements feedback by using the student model's classification performance and learning status for each category to adjust the probability threshold dynamically. The system monitors how well the model learns each category and uses this feedback to adaptively modify thresholds, ensuring that categories that are harder to learn receive more favorable threshold conditions
Data Source
Figure 1~2
Figure 3~5
Figure 6~8
AI summary
The present application discloses a label generation method, an image classification model, an image classification method and a device. The label generation method includes: classifying and identifying an unlabeled image by a teacher model and a student model respectively to obtain classification reference information of the teacher model and classification reference information of the student model; the classification reference information of each model includes N detection objects included in the unlabeled image and a probability of each detection object in each category of a plurality of categories; determining a number of detection objects of each category in the teacher model according to the classification reference information of the teacher model, and determining a number of detection objects of each category in student model according to the classification reference information of the student model; determining a probability threshold of each category according to the number of detection objects of each category in the teacher model and the number of detection objects in the student model; generating a pseudo label of the unlabeled image for the student model according to the second classification reference information and the probability threshold of each predicted category.