A Compound-Eye-Inspired Multi-Scale Neural Architecture with Integrated Attention Mechanisms

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biologically Inspired Computing

Background:

• Effective integration of multi-scale features and contextual information is crucial for neural system structure modeling and complex visual tasks.
• Existing models often struggle with balancing global and local feature representation efficiency.

Purpose of the Study:

• To propose CompEyeNet, a biologically inspired hybrid neural network architecture.
• To enhance multi-scale information representation and reconstruction capabilities for complex visual tasks.

Main Methods:

• Developed a hybrid architecture combining transformers (MATBN) and lightweight convolutional structures (CENN).
• MATBN utilizes multiple attention mechanisms for local and long-range dependencies.
• CENN enhances high-resolution feature layers and attention fusion for multi-scale representation.

Main Results:

• CompEyeNet demonstrated superior performance on medical image segmentation datasets (MICCAI-CVC-ClinicDB, ISIC2018, MICCAI-tooth-segmentation).
• Achieved better performance with fewer parameters compared to Deeplab, Unet, and YOLO series.
• Reduced parameters by 38.31% compared to YOLOv11, improving Dice, Jaccard, Precision, and Recall.

Conclusions:

• CompEyeNet offers significant advantages in parameter efficiency and accuracy for neural system modeling and image analysis.
• Bio-inspired attention-fusion hybrid neural networks show broad application potential.