摘要
近年来,航空光学成像技术快速发展,机载图像处理系统对于目标检测精度和检测速度的要求越来越高,传统的目标检测算法已经无法满足要求。与此同时,基于深度学习的目标检测算法凭借更优的性能表现得到了学术界的广泛关注。但这类算法往往参数较多,时间复杂度高且移动端移植困难。针对上述问题,本文提出了一种基于Yolo V3算法的MPSOC平台实现方案。利用改进的k均值聚类算法获取新的初始锚框,之后通过改变特征图的大小提高算法对小目标的检测精度,通过基于敏感度的剪枝方法压缩算法大小,最后利用VISDRONE数据集在MPSOC平台进行了验证。实验结果表明改善的Yolo算法的MAP提高了1.3%,误检率也得到了极大降低。算法经过压缩后,检测速度提高了1倍,体积仅为原来的37%,基本满足了对航空图像目标检测的设计要求,同时为深度学习算法在MPSOC中实现提供了可行的解决方案。
In recent years,the traditional aerial image target detection algorithms have been unable to meet the requirements of detection accuracy and speed,while the rapid development of target detection algorithms based on deep learning provides a new idea for target detection.However,this kind of algorithm is often accompanied by large scale and highly dependent on GPU devices,which makes the migration of the mobile end of the algorithm difficult.Aiming at the above problems,this paper proposes a MPSOC platform implementation scheme based on Yolo V3 algorithm.Firstly,the anchor frame of the original network is re-selected by means of k-means clustering,the detection accuracy of the algorithm is increased by adjusting the convolutional layer,and then the model scale is compressed by sensity-based pruning operation.Finally,VISDRONE data set is used to verify the Xilinx ZYNQ series MPSOC platform.The experimental results show that MAP of the improved Yolo algorithm increases by 1.3%,and the false detection rate is also greatly reduced.After the model is compressed,the detection speed is doubled and the volume becomes 37%of the original.It basically meets the design requirements of aerial image target detection,and provides a feasible solution for the implementation of deep learning algorithm in MPSOC.
作者
任彬
王宇庆
丛振
聂海涛
杨航
REN Bin;WANG Yu-qing;CONG Zhen;NIE Hai-tao;YANG Hang(Changchun Institute of Optics, Final Mechanics and Physics, Chinese Academy of Sciences, Changchun 130033, China;University of Chinese Academy of Sciences, Beijing 100049, China;Department Medical Engineing, The 964th Hospital, Joint Service Support Unit, Chinese People's Liberation Army, Changchun 130033, China)
出处
《液晶与显示》
CAS
CSCD
北大核心
2021年第7期1006-1017,共12页
Chinese Journal of Liquid Crystals and Displays
基金
国家自然科学基金青年基金(No.61401425)。
