| The landing of neural networks is one of the important research directions in recent years.Convolutional neural networks are widely used in the field of image identification and image processing because of their special characteristics.As the accuracy of deep convolutional neural networks is higher and higher on various data sets,their network frameworks are gradually becoming more complex,which is not conducive to the practical application of mobile and embedded terminals.Therefore,a number of lightweight neural networks for mobile terminals have emerged,providing possibilities for mobile and embedded applications.The acceleration and deployment of corresponding lightweight neural networks have become a hot spot for current development.In this paper,we combine the lightweight neural network with FPGA and accelerate the typical lightweight neural network Mobilenet V2 by FPGA.First of all,this paper introduces the research status and development trend of lightweight neural network at home and abroad,and enumerates the efforts made at home and abroad for the design and development of lightweight convolutional neural networks.Then we analyze and introduce the research status of the accelerated realization of lightweight neural network at home and abroad,and propose the research direction of this paper,which is the FPGA acceleration implementation of Mobilenet V2.Then,we analyze the network framework of Mobilenet V2 in depth,and analyze the knowledge related to FPGA acceleration.On this basis,we design the acceleration framework of each operator,and propose the corresponding external storage structure to improve the bandwidth utilization of external storage.Then we give the specific implementation process of the framework.Finally,we test the designed Mobilenet V2 in detail.The test results show that the FPGA-based Mobilenet V2 convolutional neural network acceleration framework proposed and implemented in this paper works normally,and can fully realize the image classification function of Mobilenet V2.In the case of limited bandwidth,it can still achieve similar or faster image classification speed than traditional processing units(ARM,CPU),which is 44% faster than optimized ARM implementation and near performance compared to optimized CPU implementation,but lower power consumption.Meanwhile,the framework has great compatibility and portability,when the bandwidth is not limited,the computational utilization rate of the framework is 94%,when the resources are not limited,the framework can repeatedly instantiate the minimum computing unit to achieve higher parallelism and then improve the overall calculation speed. |
