Brain Cognition-Inspired Dual-Pathway CNN Architecture for Image Classification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning Architectures

Background:

• Traditional Convolutional Neural Networks (CNNs) often struggle with "texture bias" and "semantic confusion" due to limitations in processing global context.
• The human visual system effectively integrates local details with global context for robust image understanding.

Purpose of the Study:

• To propose a novel CNN architecture, Cognition-Inspired Network (CogNet), that mimics the human visual system's global-local information processing.
• To enhance feature extraction by combining local details with global contextual information for improved image recognition performance.

Main Methods:

• Developed CogNet with three key components: a local pathway (CNN blocks) for fine feature extraction, a global pathway (transformer encoder) for contextual understanding, and a top-down modulator for feature integration.
• Encapsulated the dual-pathway mechanism into a reusable "global-local block" (GL block) for flexible network construction.
• Validated CogNet's efficacy through extensive experiments on six benchmark datasets.

Main Results:

• CogNets achieved state-of-the-art performance accuracies across all evaluated benchmark datasets.
• Demonstrated significant effectiveness in mitigating the "texture bias" and "semantic confusion" issues prevalent in existing CNN models.
• The proposed architecture shows superior ability in capturing both intricate local features and overarching global context.

Conclusions:

• The proposed CogNet architecture offers a powerful new approach to image recognition by effectively integrating global and local information.
• CogNets provide a robust solution to common challenges in deep learning for computer vision, leading to enhanced accuracy and reliability.