Image Label Error Detection Using Cross-Subset Model Voting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Error image labels in training data negatively impact the performance of deep learning models used for object detection and fine-grained image classification, leading to reduced prediction accuracy.
Innovation Solution
A method involving the division of a training dataset into subsets, training multiple prediction models on these subsets, and using them to generate multiple labels for each image, allowing identification of potential error image labels based on inconsistencies among the generated labels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If training data contains manually labeled images, then the model can be trained for image classification, but error labels inevitably reduce prediction accuracy
Solution Approach 1:
The training dataset is divided into N subsets, and N different prediction models are trained on these subsets. Each model independently labels the images in other subsets, generating N-1 prediction labels for each image. This segmentation approach allows multiple perspectives to identify potential error labels through comparison.
Solution Approach 2:
The system uses the prediction labels generated by N-1 models as feedback to evaluate the original image labels. By comparing the original labels with the prediction labels from multiple models, the system identifies potential error labels, creating a feedback loop that improves label quality.
2Difficulty of detecting and measuring
If multiple prediction models are trained on different subsets to identify error labels, then error detection capability improves, but computational complexity and training time increase
Solution Approach 1:
The dataset is segmented into N subsets for parallel training of multiple models. This segmentation enables distributed computation where each model trains on a smaller subset, making the overall process more manageable and allowing for parallel execution to reduce total training time.
Solution Approach 2:
The prediction models serve themselves by using their own predictions to identify potential error labels in the training data. The models generate prediction labels that are then used to evaluate original labels, creating a self-service mechanism for error detection without requiring external annotation.
Data Source
AI summary
A method, computer system, and program product facilitate identification of error image labels in training data. The method comprises: evenly dividing a training dataset into N subsets, where the training dataset includes M data items each comprising a pair of image and its original image label; training a prediction model to label images by respectively using each of the N subsets as training data to generate N respective trained prediction models; respectively using each of the N trained prediction models trained by using one of the N subsets as training data to label the images in other N−1 subsets of the N subsets to generate N−1 prediction labels for each of the M images in the training dataset. For each image in the M data items, whether the original image label of the image is a potential error image label is based on the N−1 prediction labels of the image.