Neural Image Augmentation for Ambiguous Sample Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Collecting clean and comprehensible samples for a dataset is time-consuming and computationally expensive, and machine learning systems often struggle with ambiguous samples that are not accurately recognized, leading to ineffective training and inference.
Innovation Solution
A data augmentation system uses a separate neural network to generate augmented images that are correctly classified by a classification network, reducing the need for manual data collection and retraining, and enhances the learning capability of the prediction network through curriculum learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual data collection is used to gather clean samples, then data quality is improved, but time consumption and computational cost increase
Solution Approach 1:
The patent uses an augmentation network to generate synthetic augmented samples that copy and modify existing samples in the dataset. Instead of manually collecting new clean samples, the system creates artificial augmented versions of existing samples, thereby reducing time consumption while maintaining data quality.
Solution Approach 2:
The system performs self-service by automatically generating augmented samples using the augmentation network without requiring manual intervention. The process is automated end-to-end, from selecting ambiguous samples to generating and validating augmented versions, eliminating the need for manual data collection efforts.
2Quantity of substance
If ambiguous samples are included in the training dataset, then dataset completeness is improved, but machine learning system performance deteriorates
Solution Approach 1:
The augmentation network acts as an intermediary between the ambiguous samples and the classification network. It processes ambiguous samples by generating augmented versions that are easier to classify, thereby mediating the conflict between including ambiguous samples for completeness and maintaining system performance.
Solution Approach 2:
The system changes the parameters of ambiguous samples by generating augmented versions with modified characteristics. The augmentation network transforms ambiguous samples into clearer versions by adjusting visual parameters, thereby improving classification performance while maintaining dataset completeness.
3Measurement precision
If retraining the classification network is performed to handle ambiguous samples, then accuracy on ambiguous inputs is improved, but computational cost increases
Solution Approach 1:
The patent segments the system into two independent networks: a classification network for accurate classification and an augmentation network for generating augmented samples. This segmentation allows the classification network to remain stable while the augmentation network handles the computational burden of generating augmented versions of ambiguous samples.
Solution Approach 2:
The augmentation network performs preliminary action by pre-generating augmented samples before they are used for training or inference. This preliminary generation of easy-to-classify samples from ambiguous inputs reduces the computational cost of retraining the classification network, as the augmented samples can be prepared in advance.
Data Source
AI summary
Systems and methods for data augmentation are provided. One aspect of the systems and methods include receiving an image that is misclassified by a classification network; computing an augmentation image based on the image using an augmentation network; and generating an augmented image by combining the image and the augmentation image, wherein the augmented image is correctly classified by the classification network.