Image Identifying Model Training via Distribution-Based Data Augmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Image identification models face challenges due to data imbalance during training, where small-sample categories have insufficient data, leading to poor identification rates and potential overfitting, and existing solutions like data augmentation and loss function weighting are not universally applicable across tasks.
Innovation Solution
A training method that involves obtaining image samples from multiple categories, extracting feature vectors, calculating statistical characteristics of actual distribution functions, establishing augmented distribution functions, and using these to generate additional sample features for supervised learning, thereby balancing data and improving model accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If data enhancement operations (rotation, flip, noise addition) are applied to small-sample categories, then the number of training samples increases, but information loss occurs and the model may overfit on the small-sample category
Solution Approach 1:
The patent creates virtual copies of small-sample category data by learning the distribution characteristics from multi-sample categories and generating synthetic samples that follow the same distribution. This copying approach increases sample quantity while maintaining data fidelity and avoiding the information loss and overfitting issues associated with traditional enhancement operations like rotation and noise addition.
2Measurement precision
If weighted parameters are adjusted to emphasize small-sample category loss, then identification accuracy of small-sample categories improves, but a large number of experiments are required to search for parameters and the same parameters cannot be generalized to all tasks
Solution Approach 1:
The patent develops a distribution learning-based data augmentation method that is universally applicable across different tasks and categories. By learning the underlying distribution characteristics rather than using task-specific weighted parameters, the method achieves good identification accuracy for small-sample categories without requiring extensive parameter search or task-specific tuning, thus providing a generalized solution.
3Adaptability or versatility
If simple data enhancement operations are applied to small-sample categories, then sample imbalance is alleviated to a certain extent, but information loss occurs and the model may overfit on the small-sample category
Solution Approach 1:
The patent replaces mechanical data enhancement operations (rotation, flip, noise addition) with a distribution learning-based synthetic data generation approach. Instead of mechanically transforming existing samples, the method learns the statistical distribution from multi-sample categories and generates new samples that naturally follow this distribution, thereby maintaining information fidelity while achieving data balance.
Data Source
AI summary
The present disclosure provides a training method and device for an image identifying model, and an image identifying method. The training method comprises: obtaining image samples of a plurality of categories; inputting image samples of each category into a feature extraction layer of the image identifying model to extract a feature vector of each image sample; calculating a statistical characteristic information of an actual distribution function corresponding to each category according to the feature vector of each image sample of the each category; establishing an augmented distribution function corresponding to the each category according to the statistical characteristic information; obtaining augmented sample features of the each category based on the augmented distribution function; and inputting feature vectors of the image samples and the augmented sample features into a classification layer of the image identifying model for supervised learning.