Image Classifier Learning Device Pseudo Labeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognizer training methods face challenges with over-fitting and inability to handle images without label information, particularly when pseudo labeling is limited to surrounding areas with the same label, and when training data is scarce.
Innovation Solution
An image recognizer training device and method that uses a feature representation model to extract feature vectors from images with both labeled and unlabeled pixels, estimating distributions to assign pseudo labels efficiently, and repeatedly refining the training process to ensure all pixels are labeled, thereby minimizing the objective function based on distances between feature vectors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If pseudo labeling is performed only for pixels surrounding labeled pixels based on contour information, then labeling accuracy is improved, but images with no label information cannot be processed
Solution Approach 1:
The system performs preliminary action by training a feature representation model on available labeled data before attempting to label unlabeled pixels. This pre-trained model serves as a foundation for subsequent pseudo-labeling operations, enabling the system to handle images with no initial labels by leveraging features learned from the training set.
Solution Approach 2:
The feature representation model acts as an intermediary between labeled training data and unlabeled target images. By extracting and comparing feature vectors through this intermediate representation, the system can transfer knowledge from labeled to unlabeled pixels, bridging the gap between supervised training data and unsupervised target images.
2Productivity
If a recognizer is trained using only some labeled samples, then training speed is improved, but the risk of over-fitting increases
Solution Approach 1:
The system implements feedback by iteratively performing pseudo-labeling, re-training the feature representation model with newly labeled data, and evaluating performance. This cyclic process allows the model to progressively incorporate more training samples, reducing over-fitting risk while maintaining efficient training through incremental updates rather than complete re-training.
Solution Approach 2:
The system applies partial action by initially using only a subset of available labeled samples for training, then progressively incorporating more samples through pseudo-labeling. This approach balances training speed (using fewer samples initially) with reliability (gradually increasing sample size to reduce over-fitting), avoiding the need to process all samples from the outset.
3Loss of information
If all pixels in images are to be labeled through repeated pseudo-labeling iterations, then labeling completeness is improved, but computational time increases
Solution Approach 1:
The system performs partial action by labeling only a portion of unlabeled pixels in each iteration rather than all pixels. By selecting and labeling a subset of high-confidence pixels based on the feature representation model's predictions, the system progressively achieves complete labeling while controlling computational burden through incremental processing.
Solution Approach 2:
The system prepares for complete labeling by performing preliminary feature representation learning on the available data before initiating the pseudo-labeling process. This preliminary action creates a robust feature space that accelerates subsequent labeling iterations, reducing the overall computational time required to achieve complete labeling coverage.
Data Source
AI summary
An object is to make it possible to train an image recognizer by efficiently using training data that does not include label information. A determination unit 180 causes repeated execution of the followings. A feature representation model for extracting feature vectors of pixels is trained such that an objective function is minimized, the objective function being expressed as a function that includes a value that is based on a difference between a distance between feature vectors of pixels labeled with a positive example label and a distance between a feature vector of a pixel labeled with the positive example label and a feature vector of an unlabeled pixel, and a value that is based on a difference between a distance between a feature vector of a pixel labeled with the positive example label and a feature vector of an unlabeled pixel and a distance between a feature vector of a pixel labeled with the positive example label and a feature vector of a pixel labeled with a negative example label, and based on a distribution of feature vectors corresponding to the positive example label, a predetermined number of labels are given based on the likelihood that each unlabeled pixel is a positive example.