多模块融合的浮游生物检测器  

Multi module fusion plankton detector

作　　者：齐雷 胡姣婵 于双和 阎妍 赵颖 QI Lei;HU Jiaochan;YU Shuanghe;YAN Yan;ZHAO Ying(College of Marine Electrical Engineering,Dalian Maritime University,Dalian,Liaoning 116026,China;College of Environmental Science and Engineering,Dalian Maritime University,Dalian,Liaoning 116026,China)

机构地区：[1]大连海事大学船舶电气工程学院,辽宁大连116026 [2]大连海事大学环境科学与工程学院,辽宁大连116026

出　　处：《江苏大学学报（自然科学版）》2021年第6期727-732,共6页Journal of Jiangsu University：Natural Science Edition

基　　金：国家重点研发计划项目(2018YFC1406405);国家自然科学基金资助项目(62073054,62003070);中国博士后科学基金资助项目(2020M680037,2020M680930);辽宁省自然科学基金资助项目(2021-MS-142);中央高校基本科研业务费资助项目(3132020107)。

摘　　要：针对传统海洋浮游生物利用人工提取特征的传统检测方法,存在检测精度低、检测过程冗余等问题,基于深度学习技术,提出了多模块融合的浮游生物检测器(multi module fusion single shot detector,MMFSSD).首先,提出了特征信息增强模块,在不增加网络复杂性的前提下增加了网络的感受野,将下采样图像注入该模块中,以增强特征图的低级特征信息.在此基础上,进一步提出了选择性特征融合模块,在网络中学习融合时的权重,选择性地融合不同尺度的特征.有效性验证试验结果表明,在数据集PASCAL VOC和MS COCO中的平均精度均值分别为80.70%和32.20%,在浮游生物数据集PMID2019中的平均精度均值达到90.41%.To solve the problems of detection process with low detection accuracy and redundancy in the traditional detection method of Marine plankton by artificial features extraction,a multi module fusion single shot detector(MMFSSD)was proposed based on deep learning technology.The feature information enhancement module was proposed to add the receptive field of network without increasing the network complexity,and the down-sampled image was infused into the module to enhance the low-level feature information of feature graph.The selective feature fusion module was further proposed to learn the weight of fusion in the network and selectively fuse features of different scales.The results of verification test show that the mean average precision values are 80.70%and 32.20%on PASCAL VOC and MS COCO test-set,respectively.The mean average precision on PMID2019 data-set reaches 90.41%.

关 键 词：检测器 浮游生物 深度学习 感受野 低级特征信息 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]