An Improved Capsule Network for Image Classification Using Multi-Scale Feature Extraction.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Capsule networks offer powerful image classification but suffer from high computational costs and poor generalization on complex datasets.
• Standard capsule networks can overfit on basic datasets due to their complex structure and large number of parameters.

Purpose of the Study:

• To propose a novel enhanced capsule network topology that overcomes the limitations of standard capsule networks.
• To improve feature extraction capabilities and classification performance while reducing computational overhead.

Main Methods:

• Incorporated a multi-scale feature extraction module with star structure convolution into the capsule network.
• Integrated optimization techniques including dense connections, attention mechanisms, and low-rank matrix operations.
• Evaluated the enhanced network on various datasets like CIFAR-10, CIFAR-100, CUB, ISIC, and Forged Face EXP.

Main Results:

• The enhanced capsule network demonstrated strong classification performance across multiple datasets.
• Achieved high accuracy rates, including 98.21% on the ISIC dataset and 95.38% on the Forged Face EXP dataset.
• The proposed network effectively balances feature extraction ability with computational efficiency.

Conclusions:

• The novel enhanced capsule network topology significantly improves image classification performance.
• The integration of multi-scale feature extraction and optimization techniques addresses key limitations of existing capsule networks.
• This research offers a promising direction for developing more efficient and accurate deep learning models for image classification.