基于深度学习的舰船目标检测算法与硬件加速  

Ship target detection algorithm based on deep learning and hardware acceleration

作　　者：李磊 徐国伟 李文婧 宋庆增 LI Lei;XU Guowei;LI Wenjing;SONG Qingzeng(School of Electrical Engineering and Automation,Tian Gong University,Tianjin 300387,China;School of Computer Science and Technology,Tian Gong University,Tianjin 300387,China)

机构地区：[1]天津工业大学电气工程与自动化学院,天津300387 [2]天津工业大学计算机科学与技术学院,天津300387

出　　处：《计算机应用》2021年第S01期162-166,共5页journal of Computer Applications

基　　金：天津市高等学校基本科研业务费资助项目(2019KJ019)。

摘　　要：现有基于深度学习的检测算法,虽然有效提高了高分辨率遥感图像中的舰船目标检测准确率,但是由于其网络结构非常复杂,导致计算量和参数量巨大。为了满足实际应用中的实时性要求,采用异构硬件加速,并进行了相应的算法优化。为了更好地贴合硬件,首先在YOLOV3算法的基础上,通过对主干网络进行改进,设计并实现了YOLOV3&MobileNetV3轻量化网络,这样可以极大地削减网络的参数规模和计算量。然后在现场可编程逻辑门阵列(FPGA)平台,通过设计卷积神经网络加速器,实现了高效的轻量化神经网络。最后实验结果表明,改进的神经网络在自主研发的FPGA加速架构上,在测试集中的船舰目标的检测达到了150帧每秒的检测速度以及0.872的F1值,能够更加快速并有效地检测船舰目标。The current detection algorithms based on deep learning have effectively improved the accuracy of ship target detection in high-resolution remote sensing images,however their complex network structure result in the huge amount of calculations and parameters.In order to meet the real-time requirements in practical applications,the heterogeneous hardware acceleration and corresponding algorithm optimization were designed and implemented.To better fit the hardware,firstly,the YOLOV3&MobileNetV3 lightweight network,based on the YOLOV3(You Only Look Once Version 3)algorithm,was designed and implemented by improving the backbone network,which can greatly reduce the parameter scale and calculation of the network.Then an efficient lightweight neural network was realized by designing a convolutional neural network accelerator on the Field Programmable Gate Array(FPGA)platform.Finally,as experimental results showed,the improved neural network on the self-developed FPGA acceleration architecture achieved the detection speed at 150 frames per second and F1 value at 0.872 in the detection of ship targets in the test set,which could better quickly and effectively detect the ship targets.

关 键 词：舰船目标检测 轻量化神经网络 神经网络加速器 现场可编程门阵列 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]