Learning Dendritic-Neuron-Based Motion Detection for RGB Images: A Biomimetic Approach

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Biomimetic systems

Background:

• Biological visual systems offer advanced processing capabilities.
• Current artificial visual systems often struggle with multi-channel integration and generalization.
• Limited annotated datasets pose a challenge for training conventional models.

Purpose of the Study:

• To design a biomimetic artificial visual system (AVS) that mimics biological visual processing for RGB images.
• To improve multi-channel integration and generalization in artificial vision.
• To create a robust model for low-data scenarios.

Main Methods:

• Mimicked biological photoreceptor cone cells for initial image processing.
• Employed a learnable dendritic neuron model for ganglion cell simulation.
• Utilized a nonlearnable dendritic neuron model to simulate the lateral geniculate nucleus (LGN) for multi-channel integration.

Main Results:

• The AVS demonstrated strong generalization across diverse object-background configurations.
• Achieved superior accuracy compared to traditional models like EfN-B0, ResNet50, and ConvNeXt.
• Maintained over 96% test accuracy with limited training data, showing robustness in low-data scenarios.

Conclusions:

• The AVS advances biologically inspired multi-channel processing.
• Provides a practical framework for efficient, integrated visual processing in computational models.
• Offers a viable solution for applications with limited annotated datasets.