Self-Supervised Image Segmentation Using Meta-Learning and Multi-Backbone Feature Fusion

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Few-shot segmentation (FSS) reduces manual annotation needs but requires substantial labeled data for base classes.
• Existing FSS methods struggle with generalization and reliance on extensive labeled datasets.
• Addressing the high cost and time of data annotation is crucial for practical FSS applications.

Purpose of the Study:

• To propose a novel self-supervised few-shot segmentation method (MBFSS) that minimizes reliance on labeled data.
• To enhance feature representation by integrating multiple backbone networks.
• To improve model generalization and reduce annotation effort in FSS.

Main Methods:

• Developed a multi-backbone few-shot segmentation (MBFSS) approach.
• Employed self-supervised learning utilizing unsupervised saliency for pseudo-labeling on unlabeled data.
• Integrated features from multiple backbones (ResNet, ResNeXt, PVT v2) for richer representations.

Main Results:

• Achieved 54.3% and 25.1% accuracy in one-shot segmentation on PASCAL-5i and COCO-20i datasets, respectively.
• Outperformed baseline methods by 13.5% and 4% in one-shot segmentation tasks.
• Demonstrated significant performance gains with negligible labeling effort.

Conclusions:

• The proposed MBFSS method effectively reduces the need for manual annotation in few-shot segmentation.
• Integrating multiple backbones and self-supervised learning enhances model generalization and performance.
• This approach offers a practical solution for real-world FSS applications with limited labeled data.