Image Abnormality Model Training With Feedback Label Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image abnormality detection models suffer from low accuracy due to the subjectivity of binary classification labels, leading to noise in training labels and reduced model performance.
Innovation Solution
An iterative training method that adjusts both model parameters and mapping labels using model feedback data, iteratively refining the mapping labels and model parameters to improve accuracy by reducing noise and enhancing the model's ability to detect image abnormalities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If simple binary classification labels are used for training images, then the labeling process is simple and fast, but the training labels contain noise and model accuracy is low
Solution Approach 1:
The patent implements a feedback mechanism where the trained model's prediction results are used to iteratively optimize the training labels. The model outputs are fed back to adjust and refine the binary classification labels, transforming static noisy labels into dynamic, self-correcting training data that improves model accuracy without requiring manual relabeling
Solution Approach 2:
The system enables self-service by allowing the model to automatically optimize its own training labels through the feedback mechanism. The model uses its own prediction capabilities to identify and correct labeling errors, eliminating the need for external expert intervention in the label refinement process
2Ease of operation
If binary classification labels with subjectivity are used, then labeling is easier to perform, but model training accuracy deteriorates due to label noise
Solution Approach 1:
The feedback loop uses model prediction results to identify discrepancies between subjective binary labels and actual image characteristics. This feedback enables automatic correction of labeling subjectivity while preserving the simplicity of binary classification, thereby maintaining ease of operation while improving reliability
Solution Approach 2:
The patent changes the parameters of training labels from fixed binary values to dynamic, iteratively optimized values. By adjusting label parameters based on model feedback, the system transforms static noisy labels into adaptive training signals that improve model accuracy without complicating the labeling process
Data Source
AI summary
An image abnormality detection model training method includes acquiring a noise-containing training label as a current mapping label of a training image input into an initial image abnormality detection model to obtain a prediction label. The method further includes generating model feedback data based on the current mapping label and the prediction label, generating a label loss based on a data change of the model feedback data, and adjusting the current mapping label based on the label loss. The method also includes adjusting model parameters of the initial image abnormality detection model based on the model feedback data, and iteratively performing the inputting the training image, the generating the model feedback data, the generating the label loss, and the adjusting the model parameters until a training end condition is satisfied to obtain a trained image abnormality detection model.