Image Recognition Pseudo-Labeling With Gaussian Mixture Modeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing self-training methods for image recognition face challenges in generating accurate reference label information due to noise and lack of attention to data attributes, leading to suboptimal performance.
Innovation Solution
Utilizing a Gaussian mixture model (GMM) with basis transformation to generate characterization information and reference label information, incorporating orthogonal regularization and EM algorithms to improve the accuracy and robustness of image recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If a large amount of data without reference label information is directly used for model pre-processing, then the data requirement burden is reduced, but the model performance cannot achieve desired results
Solution Approach 1:
The patent introduces an intermediary process (self-training mechanism with Gaussian mixture model and basis transformation) that bridges the gap between unlabeled data and reliable model performance. The intermediary generates pseudo-labels and transforms data attributes to create intermediate training signals that enable effective learning without requiring abundant reference-labeled data.
Solution Approach 2:
The system performs self-service by automatically generating reference label information through self-training mechanisms. The model uses its own predictions and the Gaussian mixture model to create pseudo-labels for unlabeled data, enabling the system to improve itself without external annotation resources.
2Ease of manufacture
If conventional self-training methods are used to generate reference label information, then the process is simple, but the generated label information contains noise and lacks accuracy
Solution Approach 1:
The patent transforms the parameter representation of data attributes using basis transformation. By changing the parameter space and applying orthogonal regularization, the method enhances the quality of generated reference labels while maintaining computational feasibility. This parameter transformation approach improves label accuracy without significantly complicating the generation process.
3Reliability
If more reference label information is generated to improve performance, then model accuracy improves, but the complexity of label generation and verification increases
Solution Approach 1:
The patent segments the label generation process into distinct modular components: Gaussian mixture model for clustering, basis transformation for attribute transformation, orthogonal regularization for noise reduction, and EM algorithm for iteration. This segmentation allows each component to be optimized independently while working together to generate accurate reference labels efficiently.
Data Source
AI summary
Embodiments of the present disclosure relate to a method, a device, and a computer program product for image recognition. In some embodiments, characterization information for a first reference image in a reference image set is generated in an image recognition engine by using a Gaussian mixture model. First reference label information for the first reference image is generated based on the characterization information for the first reference image, the first reference label information being associated with a category of a first object in the first reference image. The image recognition engine is updated by determining the accuracy of the first reference label information for the first reference image. In this way, good characterization of images and generation of reference label information for the images can be achieved, thus both improving the robustness of the generated reference label information and significantly improving the accuracy of image recognition.