Multi-layer Feature Cascade Fusion Spiking Neural Network for Object Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Computational Neuroscience

Background:

• Spiking Neural Networks (SNNs) are biologically inspired models known for low-power, event-driven computation.
• Conventional object detection networks often use residual structures, introducing nonspiking operations that challenge SNN implementation.
• Integrating SNNs into object detection requires overcoming the incompatibility of residual connections with pure spike-based processing.

Purpose of the Study:

• To develop a Spiking Neural Network architecture for object detection that eliminates nonspiking operations.
• To enhance gradient propagation and feature preservation in deep SNNs for improved detection accuracy.
• To propose a novel SNN model that maintains pure spike-based computation throughout the network.

Main Methods:

• Introduced a multi-level cascaded feature extraction module to replace residual connections with cascade operations.
• Developed a pooling-convolution module combining max-pooling and spiking convolution for effective downsampling.
• Ensured pure spike-based computation by redesigning feature extraction and downsampling processes.

Main Results:

• The proposed multi-layer feature cascade fusion SNN (MFCF-SNN) demonstrated state-of-the-art performance on object detection tasks.
• Elimination of nonspiking computations through cascade operations enhanced gradient propagation.
• The pooling-convolution module effectively preserved feature information and improved gradient flow in deep SNNs.

Conclusions:

• The MFCF-SNN effectively advances SNN-based object detection by enabling deep network training with pure spike-based computation.
• The novel modules successfully address the challenge of residual structures in conventional networks for SNNs.
• The approach validates the potential of SNNs for high-performance, low-power object detection applications.