Seg-Augmix Data Augmentation for Semantic Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for semantic segmentation face challenges in data efficiency, particularly when the labeled training dataset is small, as they require significant human effort to label and often neglect data augmentation techniques effectively.
Innovation Solution
The proposed solution involves a novel data augmentation technique called seg-Augmix, which extends Augmix to semantic segmentation by removing position-sensitive augmentations and maintaining a pool of data augmentation methods. This technique generates augmented images without changing labels and applies Jensen-Shannon divergence consistency loss to stabilize training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If data augmentation techniques are applied to semantic segmentation, then training data diversity is improved, but label consistency becomes difficult to maintain due to position sensitivity
Solution Approach 1:
The patent extracts and removes position-sensitive augmentation operations from the augmentation pipeline. Specifically, it eliminates augmentations that change spatial coordinates (like rotation, flipping, cropping) while retaining position-insensitive operations (like color jittering, brightness adjustment) that do not affect label alignment.
Solution Approach 2:
The patent changes the parameters of data augmentation by selecting only those transformation parameters that preserve spatial relationships. It adjusts the augmentation strategy to use parameters that modify appearance properties rather than geometric properties, ensuring labels remain consistent with augmented images.
2Reliability
If more labeled training data is collected, then model performance is improved, but human effort and time consumption increase significantly
Solution Approach 1:
The patent creates multiple copies of existing labeled images through position-insensitive data augmentation. Instead of manually labeling new images, it generates synthetic training samples by applying transformations that preserve label validity, effectively copying and adapting existing labeled data.
Solution Approach 2:
The patent performs data augmentation during the training process itself rather than requiring all labeled data to be prepared in advance. The augmentation is applied dynamically to create training batches, eliminating the need for time-consuming manual labeling of additional datasets.
3Quantity of substance
If standard data augmentation methods are used for semantic segmentation, then data quantity is increased, but training stability deteriorates due to label mismatch
Solution Approach 1:
The patent changes the selection criteria for augmentation parameters to exclude position-sensitive transformations. By modifying which parameters are applied (using only appearance-based augmentations rather than geometric ones), it increases training data quantity while maintaining label-image consistency and training stability.
Data Source
AI summary
A method and system for training a neural network are provided. The method includes receiving an input image, selecting at least one data augmentation method from a pool of data augmentation methods, generating an augmented image by applying the selected at least one data augmentation method to the input image, and generating a mixed image from the input image and the augmented image.