EMNet: A Novel Few-Shot Image Classification Model with Enhanced Self-Correlation Attention and Multi-Branch Joint...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Bio-inspired Computing

Background:

• Few-shot image classification requires models to recognize new categories with limited data.
• Traditional methods demand extensive labeled datasets, limiting their applicability.
• Bio-inspired mechanisms offer potential for optimizing feature extraction and generalization.

Purpose of the Study:

• To introduce the Enhanced Self-Correlation Attention and Multi-Branch Joint Module Network (EMNet) for few-shot image classification.
• To address challenges in effective feature extraction and generalization to new categories.
• To leverage biological visual attention and swarm intelligence principles.

Main Methods:

• Developed the Enhanced Self-Correlated Attention (ESCA) module for precise local feature extraction.
• Integrated the Multi-Branch Joint Module (MBJ Module) to focus on inter-class similarities and intra-class differences.
• Employed bio-inspired algorithms for feature optimization and enhanced generalization.

Main Results:

• EMNet demonstrated superior performance in one-shot and five-shot learning tasks.
• Achieved higher classification accuracies than existing models on mini-ImageNet, CUB-200, and CIFAR-FS datasets.
• Showcased significant improvements, e.g., 1.27% higher accuracy on CUB-200-2011 in five-way one-shot experiments.

Conclusions:

• EMNet is an efficient end-to-end solution for few-shot image classification.
• The proposed model effectively enhances feature extraction and generalization capabilities.
• Bio-inspired approaches show significant promise for advancing few-shot learning.