Image Processing Neural Networks With Pseudo-Label Self-Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image classification methods using machine learning require substantial teacher data and resources for neural network learning, and may fail to perform appropriate class classification and labeling, especially for unlabeled images.
Innovation Solution
An image processing apparatus that includes a feature extractor, class classifier, domain classifier, and self-learner to assign pseudo labels to unlabeled images, performing self-learning and back propagation to enhance neural network learning efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional neural network learning is used for image classification, then classification accuracy can be achieved, but substantial teacher data and learning resources are required
Solution Approach 1:
The system enables unlabeled images to serve themselves by automatically generating pseudo-labels through the classification model. The model classifies unlabeled images, generates pseudo-labels based on classification confidence, and uses these pseudo-labels for self-learning, eliminating the need for manual teacher data annotation.
Solution Approach 2:
Pseudo-labels act as an intermediary between the classification model and unlabeled images. The system generates pseudo-labels for unlabeled images based on model predictions, and these pseudo-labels serve as intermediate training data that bridges the gap between labeled and unlabeled data.
2Reliability
If more teacher data is used for learning, then neural network learning completeness improves, but learning costs (time, data) increase
Solution Approach 1:
The system performs continuous self-learning by continuously generating pseudo-labels for unlabeled images and using them to train the classification model. This continuous process allows the model to progressively improve without requiring additional time for manual data annotation or repeated full retraining cycles.
Solution Approach 2:
The system performs preliminary classification of unlabeled images to generate pseudo-labels before actual training. By pre-processing unlabeled data into pseudo-labeled data, the system prepares training materials in advance, reducing the time required during the actual learning process.
3Adaptability or versatility
If conventional classification methods are applied, then labeled images can be classified, but unlabeled images cannot be appropriately labeled
Solution Approach 1:
The classification model serves unlabeled images by automatically generating pseudo-labels through its own prediction capability. The model classifies unlabeled images and uses its confidence scores to create pseudo-labels, enabling the system to handle both labeled and unlabeled images without external intervention.
Data Source
AI summary
An image processing apparatus has a first image acquisitor that acquires a source image, a second image acquisitor that acquires a first target image, a label acquisitor that acquires a label, a feature extractor including a first neural network that extracts a feature of the source image and a feature of the first target image, a class classifier including a second neural network that performs a class classification of the source image and the first target image, a domain classifier including a third neural network that performs a domain classification of the source image and the first target image, a processor that assigns a pseudo label to the first target image, a self-learner that performs a self-learning of the first neural network, the second neural network, and the third neural network, and a learner that learns the first, second and third neural networks, by performing a back propagation process.