Class-Specific Data Augmentation for AI Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data augmentation techniques apply the same transformation policies in a class-agnostic fashion, leading to suboptimal performance as different transformations are beneficial or detrimental for different classes of objects.
Innovation Solution
The method involves obtaining a group of transform functions and data augmentation policies specific to each class, generating augmentation data by applying these policies to images within the training dataset, and combining the original and augmented data to create a final training dataset for an AI model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If class-agnostic data augmentation policies are applied to all classes, then the data augmentation process is simple and efficient, but the accuracy of AI models deteriorates for specific classes that require different transformations
Solution Approach 1:
The patent segments the data augmentation process by class, assigning different transformation policies to different classes based on their specific requirements. Instead of applying a uniform augmentation strategy to all classes, the system divides the augmentation process into class-specific segments, where each class receives customized transformations that optimize its performance.
Solution Approach 2:
The patent implements local quality by applying different data augmentation transformations tailored to the specific characteristics of each class. Each class receives localized augmentation policies that are optimized for its unique visual properties, rather than applying a generic augmentation strategy uniformly across all classes.
2Reliability
If class-specific data augmentation policies are implemented, then the accuracy of AI models improves for all classes, but the complexity of the data augmentation process increases
Solution Approach 1:
The patent applies preliminary action by pre-defining and storing class-specific augmentation policies before the actual data augmentation process. The system prepares a library of transformation policies organized by class, allowing the augmentation process to simply retrieve and apply the appropriate pre-defined policies rather than computing custom transformations in real-time.
Solution Approach 2:
The patent uses copying by creating and storing templates of class-specific augmentation policies that can be replicated and applied to multiple images of the same class. Instead of manually designing augmentation strategies for each image, the system copies and applies pre-validated transformation policies that are optimized for each class.
3Ease of manufacture
If uniform data augmentation transformations are applied to all images, then the implementation is straightforward, but the diversity of training data is insufficient for classes with specific transformation requirements
Solution Approach 1:
The patent introduces dynamics into the data augmentation process by making the transformation policies adaptive to the specific class of each image. The system dynamically selects and applies different augmentation transformations based on the class label, allowing the augmentation strategy to adapt its behavior according to the visual characteristics and requirements of different classes.
Data Source
AI summary
A method for controlling an artificial intelligence (AI) device can include obtaining a group of transform functions, a first training data set, a plurality of data augmentation policies including a first data augmentation policy including a first transform function from the group of transform functions for a first class and a second data augmentation policy including a second transform function from the group of transform functions for a second class, the first data augmentation policy being different than the second data augmentation policy. Also, the method can include generating first augmentation data by transforming images within the first training data set having the first class based on the first transform function, generating second augmentation data by transforming images within the first training data set having the second class based on the second transform function, generating a final training data for training an AI model.