Image Classification Training With Iterative Pseudo-Label Updates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image classification methods for medical images rely heavily on manual labeling, which is labor-intensive, inefficient, and prone to errors.
Innovation Solution
An AI-based image processing method that utilizes pseudo classification labels during training, where pseudo classification labels are generated and updated iteratively, reducing the need for manual labeling and improving training efficiency and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual labeling is used for training image classification models, then the accuracy of classification labels is improved, but the labor cost and time consumption increase significantly
Solution Approach 1:
The system enables self-service by allowing the image classification model to generate its own training labels through automatic classification. The model classifies unlabeled images and uses these automatic classification results as training labels, eliminating the need for manual labeling while maintaining continuous model improvement through iterative retraining.
Solution Approach 2:
The patent introduces automatic classification results as an intermediary between unlabeled images and training labels. These intermediate classification results serve as pseudo-labels that bridge the gap between raw images and required training annotations, enabling the system to utilize unlabeled data without manual intervention.
2Reliability
If manual labeling is used for all training images, then the quality of training data is improved, but the quantity of training samples is limited due to high costs
Solution Approach 1:
The system applies local quality by differentiating the quality requirements for different subsets of training data. Manual labeling is applied selectively to a portion of training images to establish high-quality baseline labels, while automatic classification is used for the remaining images. This heterogeneous approach allows the system to scale training data quantity while maintaining acceptable quality through the combination of manually labeled and automatically labeled samples.
3Productivity
If pseudo classification labels are used for training, then the productivity and efficiency are improved, but the measurement precision of labels deteriorates
Solution Approach 1:
The system implements feedback by continuously using the image classification model to generate pseudo-labels for unlabeled images, retraining the model with these newly labeled images, and iteratively improving classification accuracy. The feedback loop ensures that errors in pseudo-labeling are progressively reduced as the model learns from the expanding set of labeled training samples, thereby improving label precision over time while maintaining high productivity.
Data Source
AI summary
An image processing method can reduce costs related to manual labeling, improve training efficiency, and increase a quantity of training samples, thereby improving the accuracy of an image classification model. First images and second images are processed using an image classification model to obtain predicted classification results. The first images include a classification label and the second images include a pseudo classification label. A first loss value indicating accuracy is acquired based on the predicted classification results, the corresponding classification labels, and the corresponding pseudo classification labels. A second loss value indicating accuracy is acquired based on the predicted classification results and the corresponding pseudo classification labels. A model parameter of the image classification model is updated based on the first loss value and the second loss value. Classification processing and acquisition is performed until a target image classification model is obtained.