Phase Image Generator Training for Low-Data Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional classification methods require a large number of samples to achieve high accuracy, leading to low classification accuracy when the number of samples is insufficient.
Innovation Solution
A training method for a phase image generator and classifier that involves generating phase images, determining their difference from original images, calculating a loss value, and selecting a stable phase image generation mode to produce training images that are then used to train the classifier, allowing for high classification accuracy with fewer original images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large number of samples are used for training, then classification accuracy is improved, but the requirement for sample quantity increases
Solution Approach 1:
The patent uses a phase image generator to create synthetic phase images that copy the characteristics of original images. These generated images serve as artificial training samples, allowing the classifier to be trained with fewer real samples while maintaining high classification accuracy. The generator produces multiple copies/variations of limited original images to expand the training dataset.
Solution Approach 2:
The patent performs preliminary generation of phase images before the actual classification task. By pre-generating synthetic training samples using the phase image generator and selecting optimal generation modes through loss value evaluation, the system prepares high-quality training data in advance, enabling the classifier to achieve high accuracy without requiring large numbers of real samples.
2Loss of time
If the number of original images is reduced, then data collection time is reduced, but classification accuracy deteriorates
Solution Approach 1:
Instead of collecting large numbers of real images, the system copies existing images through the phase image generator to create synthetic training samples. This copying process is computationally efficient and does not require additional data collection time, while the generated images maintain the essential characteristics needed for accurate classification.
Solution Approach 2:
The system changes parameters of the phase image generation process (such as phase modulation depth, frequency, and other optical parameters) to generate diverse training samples from limited originals. By varying these parameters, the system creates sufficient diversity in the training dataset without needing to collect more real images, thus maintaining accuracy while reducing data collection time.
Data Source
AI summary
The training method for the phase image generator includes the following steps. Firstly, in each iteration, a loss value is generated, including: (1). the phase image generator generates a plurality of generated phase images using a phase image generation mode; (2). the phase image determiner determines a degree of difference between the generated phase images and original phase images; (3). the loss value of the generated phase images is generated according to the degree of difference. Then, a selector selects a stable loss value from the loss values, and uses the phase image generation mode in the iteration corresponding to the stable loss value as a selected phase image generation mode of the phase image generator.