基于动态视觉运动特征的脉冲神经网络识别方法

Spiking Neural Network Recognition Method Based on Dynamic Visual Motion Features

针对现有脉冲神经网络(SNN)对动态视觉事件流识别精度低与实时性差等问题,该文提出一种基于动态视觉运动特征的脉冲神经网络识别方法。首先利用基于事件的运动历史信息表示与梯度方向计算提取事件流中的动态运动特征;然后引入时空池化操作来消除事件在时间和空间上的冗余,保留显著的运动特征;最后,将特征事件流输入脉冲神经网络进行学习与识别。在基准的动态视觉数据集上的实验结果表明,动态视觉运动特征可显著提升SNN对于事件流的识别精度与计算速度。

Considering the shortcomings of the low recognition accuracy and poor real-time performance of existing Spiking Neural Networks(SNN)for dynamic visual event streams,a SNN recognition method based on dynamic visual motion features is proposed in this paper.First,the dynamic motion features in the event stream are extracted using the event-based motion history information representation and gradient direction calculation.Then,the spatiotemporal pooling operation is introduced to eliminate the redundancy of events in the temporal and spatial domain,further retaining the significant motion features.Finally,the feature event streams are fed into the SNN for learning and recognition.Experiments conducted on benchmark dynamic visual datasets show that dynamic visual motion features can significantly improve the recognition accuracy and computational speed of SNN for event streams.