基于深度可分离卷积神经网络的水声目标分类研究及FPGA实现  

Underwater Acoustic Target Classification Based on Depthwise Separable Convolution Neural Network and FPGA Implementation

作　　者：张天帅 刘金涛[1] 王良[1] Zhang Tianshuai;Liu Jintao;Wang Liang(Faculty of Information Science and Engineering,Ocean University of China,Qingdao 266100,China)

机构地区：[1]中国海洋大学信息科学与工程学部,山东青岛266100

出　　处：《中国海洋大学学报（自然科学版）》2024年第8期152-165,共14页Periodical of Ocean University of China

基　　金：国家自然科学基金项目(52001296)资助。

摘　　要：针对传统声纳处理器算力受限,能效比低,难以支撑水声目标识别实时推理的问题,本文基于异构SoC平台设计了面向被动声纳水下目标实时计算处理系统。该系统具有较低资源开销和较小分类精度损失等优点,是一种低时延、高能效比的硬件加速器解决方案。本文以MobileNetV1网络模型为基础并对其进行结构优化,在现场可编程门阵列(Field programmable gate array,FPGA)上通过并行流水线的加速结构实现它的前向推理过程,并对其权值参数进行二值化的处理,以达到减少存储量和计算量的同时加快其推理速度的目的。同时,根据在输入通道维度以及输出图像高度上分块并行的优化思想,设计了深度可分离卷积的流水优化策略,采用并行流水的结构极大减少了前向推理的时间。实验表明,在利用出海实际采集得到的水声数据集上,本文实现的系统识别精度为88.5%,在的分辨率的图像上,时间延迟达到4.23 ms。对比CPU速度提升了70.68倍,是GPU速度的68%。能效比分别为CPU的0.08%,GPU的2.12%。本文为神经网络在硬件资源有限以及功耗存在限制的轻量型移动端或者边缘设备上的应用与部署,以及对促进融合水下勘探网络的建设和水下信息的快速获取提供了设计思路。Aiming at the problem that the traditional sonar processor has limited computing power and low energy efficiency ratio,which is difficult to support real-time deduction of underwater acoustic target recognition,this paper designs a real-time computing and processing system for passive sonar underwater targets based on heterogeneous SoC platform.This system provides an efficient accelerator solution with low resource cost and nearly lossless network accuracy to reduce deduction delay and power consumption.This paper optimizes the structure of MobileNetV1,implements its forward deduction process on the field programmable logic gate array(FPGA)through the parallel pipeline acceleration structure,and binarizes its weight parameters,so as to reduce the amount of storage and computation while speeding up its deduction speed.In addition,according to the optimization idea of block parallel in channel dimension and input image height,a pipeline optimization strategy with depthwise separable convolution is designed to reduces the time of forward deduction.The experiment shows that using the underwater acoustic data set collected at sea,the recognition accuracy of the system realized in this paper is 0.875,the time delay is 4.23 ms on the image with a resolution of 3×128×128.Compared with the CPU speed,it is 70.68 times faster,which is 68%of the GPU speed.The power consumption is 10.6%of CPU and 1.44%of GPU respectively.This paper provides a design idea for the application and deployment of neural networks on lightweight mobile terminals or edge devices with limited hardware resources and power consumption,and for promoting the construction of integrated underwater exploration networks and rapid acquisition of underwater information.

关 键 词：水声目标分类 深度可分离卷积 定点量化 FPGA 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TH89[自动化与计算机技术—控制科学与工程]