Image Classification Device Using Noise Augmentation and Manifold Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image classification systems using deep learning struggle to capture important features in small image regions and are sensitive to noise, leading to reduced accuracy and increased man-hours in training data creation.
Innovation Solution
An image classification device and method that generates noise-variant image groups to differentiate features, employing manifold learning for two- or three-dimensional mapping, allowing robust feature extraction and visualization of important features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If principal component analysis is performed to map features two-dimensionally or three-dimensionally, then dimensionality reduction is achieved, but small features within the image are ignored and noise may be taken as axes
Solution Approach 1:
Instead of directly performing dimensionality reduction on the original image data, the patent inverts the approach by first generating augmented images with added noise, then using these noisy versions to train the feature extraction model. This inversion allows the model to learn features that are robust to noise while still achieving effective dimensionality reduction through manifold learning.
Solution Approach 2:
The patent applies preliminary noise augmentation to the training images before feature extraction. By pre-processing the images with various noise types and intensities, the model is prepared to handle noise during inference, ensuring that small important features are not lost and that noise does not dominate the feature space during dimensionality reduction.
2Reliability
If a large number of input images and training information are used for deep learning training, then model accuracy is improved, but training cost becomes enormous
Solution Approach 1:
The patent creates multiple copies of the same image with different noise augmentations applied. Instead of requiring many different original images, the system generates synthetic variations by adding noise to existing images, effectively multiplying the training data utility without proportionally increasing data collection costs.
Solution Approach 2:
The patent changes parameters of existing images by applying different noise types and intensities. This parameter transformation allows a single image to serve multiple training purposes, reducing the need for large volumes of unique training images while maintaining model accuracy through diverse training examples.
Data Source
AI summary
Provided are an image classification device and method that are capable of extracting and mapping an important feature in an image. The image classification device includes: a feature extraction unit 101 that generates a first image group generated by applying different noises to the same image among images included in an image group and a second image group including different images, is trained such that features obtained from the first image group are approximate, is trained such that features obtained from the second image group are more different, and extracts features; a feature mapping unit 102 that maps the extracted plurality of features two-dimensionally or three-dimensionally using manifold learning; and a display unit 103 that displays a mapping result and constructs a training information application task screen.