Object Region Segmentation Using Confidence-Guided Pseudo Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep-learning network models for object region segmentation in images require large amounts of labeled data, which is costly and time-consuming to collect, leading to inefficiencies in training and performance.
Innovation Solution
A deep-learning network model comprising a pseudo label generation network, a confidence map generation network, and a segmentation network, trained using a semi-supervised learning technique, generates pseudo labels and confidence maps to segment object regions efficiently with a small amount of labeled data, leveraging unlabeled data through pseudo labeling and confidence-guided training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If deep-learning network models are trained using large amounts of labeled data, then segmentation performance is improved, but data collection costs and time increase significantly
Solution Approach 1:
The system performs preliminary actions by generating pseudo-labels for unlabeled data using a pre-trained first network model before actual training. This preliminary labeling allows the third network model to be trained with augmented labeled data, improving segmentation performance without the time cost of manual annotation for all training samples
Solution Approach 2:
The system creates copies of labeled data by generating pseudo-labels that mimic true labels for unlabeled images. These pseudo-labels serve as synthetic training targets, effectively multiplying the available labeled data without requiring additional manual annotation time
2Reliability
If deep-learning network models are trained using large amounts of labeled data, then segmentation performance is improved, but training costs increase
Solution Approach 1:
The system creates synthetic labeled data copies by generating pseudo-labels for unlabeled images through the first network model. This copying approach effectively increases the quantity of available labeled training data without the resource-intensive process of manual annotation, enabling better segmentation performance with reduced labeled data requirements
Solution Approach 2:
The system changes the parameter of data labeling status by transforming unlabeled data into pseudo-labeled data through automated prediction. This parameter change allows unlabeled data to be utilized as training samples, reducing dependence on manually labeled data while maintaining segmentation performance
3Productivity
If pseudo labels with low confidence are used for training, then more data can be utilized, but training accuracy decreases
Solution Approach 1:
The system applies local quality by differentiating between high-confidence and low-confidence pseudo-labels through the confidence map generated by the second network model. High-confidence regions are used for training the third network model while low-confidence regions are excluded or down-weighted, ensuring that only reliable pseudo-labels contribute to training and maintaining high training accuracy
4Productivity
If a simple network model is used, then training is faster and less resource-intensive, but segmentation performance suffers
Solution Approach 1:
The system segments the training process into multiple specialized network models: a first model for pseudo-label generation, a second model for confidence assessment, and a third model for final segmentation. This segmentation allows each model to be optimized for its specific function, achieving both training efficiency and high segmentation performance through divided computational tasks
Solution Approach 2:
The system performs preliminary actions by using the first and second network models to generate pseudo-labels and confidence maps before training the third segmentation model. This preliminary processing prepares high-quality training data in advance, enabling the third model to train faster with pre-processed inputs while maintaining high segmentation accuracy
Data Source
AI summary
An object region segmentation device and an object region segmentation method thereof are provided. The object region segmentation device includes a processor and storage. The storage stores a deep-learning network model for segmenting an object region in an image. The deep-learning network model includes a first network model for generating a pseudo label, a second network model for generating a confidence map for the pseudo label, and a third network model for segmenting the object region in the image. The processor inputs an unlabeled image to the first network model to generate the pseudo label, inputs the pseudo label to the second network model to generate the confidence map, and trains the third network model using a pseudo label corresponding to at least one pixel, a confidence level of which is greater than or equal to a threshold, on the confidence map.