Bidirectional dynamic threshold SNN for enhanced object detection with rich spike information.

Area of Science:

• Artificial Intelligence
• Neuroscience-inspired Computing
• Computer Vision

Background:

• Spiking Neural Networks (SNNs) offer energy efficiency but struggle with accuracy in complex tasks like object detection due to limited information capacity.
• Directly trained SNNs often underperform Artificial Neural Networks (ANNs) because of binary spike feature maps.

Purpose of the Study:

• To enhance the accuracy and information encoding capacity of directly trained Spiking Neural Networks for object detection.
• To introduce a novel SNN architecture, BD-SNN, that overcomes the limitations of traditional SNNs.

Main Methods:

• Proposed Bidirectional Dynamic Threshold neurons (BD-LIF) that emit bipolar spikes (+1/-1) and dynamically adjust thresholds.
• Introduced two novel all-spike residual blocks (BD-Block1 and BD-Block2) for efficient feature extraction and multi-scale fusion.
• Implemented and evaluated the BD-SNN architecture on COCO and Gen1 object detection datasets.

Main Results:

• BD-SNN demonstrated improved accuracy over the state-of-the-art EMS-YOLO method by 3.1% on the COCO dataset.
• BD-SNN achieved a 2.8% accuracy improvement on the Gen1 dataset compared to EMS-YOLO.
• The proposed BD-LIF neurons and BD-Blocks effectively enhanced information encoding and network performance.

Conclusions:

• The novel BD-SNN architecture significantly improves object detection accuracy compared to existing SNN methods.
• Bidirectional Dynamic Threshold neurons and specialized residual blocks are effective in boosting SNN performance.
• BD-SNN presents a promising advancement for energy-efficient and accurate SNN-based object detection.