Design and implementation of neural network computing framework on Zynq SoC embedded platform
Xingying Li, Zhenyu Yin, Fulong Xu, Feiqing Zhang, Guangyuan Xu
Abstract
Limited resources and low computing power of embedded platform make it difficult to apply neural network technology. To overcome this problem, a new neural network computing framework “Zynq-Darknet” was proposed. The framework is based on Darknet, which constructs depthwise separable convolution and a lightweight classification model MobileNetV2 and was deployed to Xilinx Zynq-7000 System-on-Chip (SoC) with Linux operating system (OS). In order to verify the performance of the framework and model, experiments were conducted on imagenet-1k dataset using different network structures. The results show that the MobileNetV2 network model based on Zynq-Darknet can effectively perform image classification, and ensure a certain real-time and accuracy while reducing the computational complexity and storage overhead, assuming promising application prospects.