Two-Branch Attention Network via Efficient Semantic Coupling for One-Shot Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) have advanced one-shot image classification.
• Current CNNs lack effective attention modeling, limiting performance.
• Attention mechanisms are crucial for enhancing feature representation in image classification tasks.

Purpose of the Study:

• To propose a novel Two-branch (Content-aware and Position-aware) Attention (CPA) Network for improved attention modeling in one-shot image classification.
• To enhance the learning of attention mechanisms by mutually utilizing support and query images.
• To introduce a local-global optimizing framework to boost recognition accuracy.

Main Methods:

• Developed a CPA Network with an Efficient Semantic Coupling module for attention modeling.
• Implemented content-aware attention to capture characteristic features (color, shape, texture).
• Implemented position-aware attention to model spatial position weights.
• Designed a local-global optimizing framework for enhanced recognition.
• Evaluated the CPA module integrated into a local-global Two-stream framework (CPAT).

Main Results:

• The proposed CPA module, as CPAT, achieved state-of-the-art performance on four benchmark datasets (miniImageNet, tieredImageNet, CUB-200-2011, CIFAR-FS).
• Demonstrated significant accuracy improvements, particularly a 3.16% increase on the CUB-200-2011 dataset.
• Validated effectiveness across three popular one-shot learning networks (DPGN, RelationNet, IFSL).

Conclusions:

• The CPA Network effectively addresses limitations in attention modeling for CNN-based one-shot image classification.
• The proposed CPAT method offers a robust approach to enhance recognition accuracy in few-shot learning scenarios.
• The study highlights the importance of integrating content and position awareness for superior attention mechanisms in computer vision.