Res-SRNN: Residual Spiking Recurrent Neural Network for Efficient Video Recognition

Action recognition in videos is a critical task in computer vision, with a wide range of applications including security surveillance, behavior analysis, and cooperative control. Despite significant advancements in action recognition accuracy achieved by existing video recognition methods, they commonly suffer from high computational complexity and substantial consumption of memory and computational resources. To address the urgent demand for concise, energy-efficient, and high-performance video recognition methods, we propose a residual spiking recurrent neural network (Res-SRNN). Firstly, we design the residual convolution module based on 2D convolution operations to effectively extract spatial information from input video frames. Moreover, drawing inspiration from the recurrent connections in the visual cortex, we incorporate the microscale spiking mechanism of the brain and put forward the spiking recurrent neural network (SRNN). Our SRNN not only excels at capturing complex temporal dependencies but also conserves computational energy costs. Experimental results demonstrate that our proposed Res-SRNN method significantly outperforms most comparative methods on benchmark datasets, achieving the best balance between parameter efficiency and performance.