Research On Key Problems Of Fixed-point For Convolutional Neural Network

Convolutional neural network is one of the artificial neural networks.Object detection research based on convolutional neural network has drawn increasing attention due to the high recognition accuracy,high robustness,and easy implementation.However,with the increasing requirement for recognition accuracy,the parameter scale of convolutional neural network is getting larger.Massive convolution operations seriously affect the calculation speed.At the same time,the convolution operation of floating-point numbers has high computation complexity.Therefore,the fixed-point accelerated algorithm for convolutional neural network has become the key research of convolutional neural network optimization calculations on embedded devices.In this paper,a fixed-point quantitative accelerometer convolution calculation method based on Field Programmable Gate Array(FPGA)is proposed for the problems of large parameter scale of convolutional neural network,high computation complexity,and slow speed on resource-constrained devices.This paper analyzes the reasons of slow calculation speed of convolutional neural network,the model parameter distribution and the characteristics of convolution operation,then proposes a FPGA-based fixed-point quantitative accelerometer convolution calculation method.In order to further reduce the fixed-point quantization bit width and ensure the less precision loss,this paper designs an optimization strategy for the quantization of input feature map parameters of convolutional neural network.In addition,this paper optimizes the FPGA fixed-point computation,so that the model is adaptable to FPGA hardware resource allocation,maximize acceleration of convolution calculation and fully utilized resources such as multipliers and adders.To evaluate the effectiveness of fixed-point quantization methods and fixed-point calculation models,this paper selects multiple convolutional neural network models as the measured objects and designs dynamic fixed-point quantization methods to quantify the weights and inputs of the model.Then this paper compares and analyzes the parameter file size,accuracy and speed before and after quantization respectively.The results demonstrate the FPGA-based fixed-point quantization accelerated convolution calculation method is feasible and effective under the premise of satisfying the basic convolutional neural network model structure.Compared with the traditional floating-point convolution calculation,the proposed method ensures when the weight and input feature map parameters are quantized to 7Bit,the storage reduce by about 4.5 times,the convolution operation accelerate by about 18.69 times,under the premise of the small accuracy loss of the convolutional neural network.Moreover,the proposed method improves the FPGA resource utilization and demonstrates it is an effective fixed-point acceleration method for convolutional neural networks.