用HLS技术与BP神经网络实现的手语识别系统

Sign Language Recognition Based on HLS and BP Neural Network

提出一种通过HLS(High Level Synthesis,高层次综合)技术快速将BP神经网络硬件化的方法,并结合曲度传感器和惯性测量单元设计了一套识别率较高的手语识别系统。利用HLS技术将用于识别的BP神经网络实现于XilinxZynq全可编程SoC的PL(可编程逻辑)端,并在资源和时序上做出优化。实验结果表明,使用HLS技术设计的手语识别系统能够在更高的抽象层面完成算法设计和实现,在FPGA中可以高效地实现BP神经网络,相较于传统的RTL设计而言,设计周期大大缩短,系统对10种手语的识别情况达到了理论准确率的水平。

In the paper,a method for quickly implementing BP neural network by HLS (High Level Synthesis) is proposed.Combined with flex sensor and inertial measurement unit,a set of sign language recognition system with high recognition rate is designed.In this paper, the BP neural network for identification is implemented on the PL(Programmable Logic) side of the Xilinx Zynq All Programmable SoC by HLS technology,and optimized in terms of resources and timing.The experiment results show that the design of HLS technology can complete the algorithm design and implementation at a higher level of abstraction.BP neural network can be realized efficiently in FPGA.Compared with the traditional RTL design,the design cycle is greatly shortened.The recognition of sign language reaches the level of theoretical accuracy.