Target Detection Method Based on Lightweight Spiking Neural Network

To address the issue of balancing model size, detection accuracy, and energy efficiency in existing models for object detection tasks, we propose Spiking-Lite, a lightweight SNN-based object detection model. The model employs a lightweight feature extraction structure that utilizes hybrid convolution to replace traditional convolution, thereby reducing the number of model parameters. We design a multi-dimensional attention module that dynamically allocates differential weights to features, enhancing the fusion of multi-scale features. A hybrid training method is employed to enhance both training efficiency and detection accuracy. Experimental results demonstrate that the Spiking-Lite model achieves a $58.1 \%$ mAP@ 0.5 on PASCAL VOC2012 and a $62.9 \% m A P \text{@} 0 . 5$ on MSCOCO. At the same level of precision, Spiking-Lite consumes only $20 \%$ of the energy of equivalent ANN models and is $30 \%$ smaller in size compared to traditional SNN models. The Spiking-Lite model provides an efficient and accurate object detection solution for embedded devices with limited computational power.