Image Semantic Segmentation Model Training with Transformed Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Image semantic segmentation models trained with noisy labels struggle to produce accurate semantic segmentation results, as noisy labels can significantly impact the model's performance and precision.
Innovation Solution
The method involves obtaining a first image and its transformed versions, generating semantic segmentation labels based on the semantic prediction results of the transformed images, and using these labels to train the image semantic segmentation model, thereby reducing the impact of noisy labels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If training data with noisy labels is used to train the image semantic segmentation model, then the training process can proceed with available data, but the precision and accuracy of the model's semantic segmentation results deteriorate
Solution Approach 1:
The patent introduces transformed images as an intermediary element between the original noisy training data and the model training process. By generating multiple transformed versions of each training image and aggregating their semantic predictions, the system creates a mediator that filters out noisy labels while preserving genuine semantic information, thus resolving the contradiction between using available training data and maintaining segmentation precision
Solution Approach 2:
The patent creates multiple copies of the original training image through geometric transformations (rotation, flipping, scaling). These copied and transformed images serve as additional validation perspectives, allowing the system to cross-verify semantic predictions and identify consistent patterns that are robust against label noise, thereby improving model accuracy without sacrificing training data utilization
Data Source
AI summary
Embodiments of the present disclosure relate to a method, device and computer-readable storage medium for image processing. A method for image processing comprises: obtaining a first image and at least one second image, wherein each image of the at least one second image is a transformed image of the first image; and generating a first semantic segmentation label for the first image based on at least one semantic prediction result of the at least one second image. By using the first image and the first semantic segmentation label as training data to train an image semantic segmentation model, it helps to increase the precision of the trained image semantic segmentation model.