基于改进Faster R-CNN的水母检测与识别算法  

Jellyfish Detection and Recognition Algorithm Based on Improved Faster R-CNN

作　　者：高美静[1] 李时雨 刘泽昊 张博智 白洋 关宁 王萍 常秋悦 GAO Mei-jing;LI Shi-yu;LIU Ze-hao;ZHANG Bo-zhi;BAI Yang;GUAN Ning;WANG Ping;CHANG Qiu-yue(College of Integrated Circuits and Electronics,Beijing Institute of Technology,Beijing 100081,China;The Key Laboratory for Special Fiber and Fiber Sensor of Hebei Province,School of Information Science and Engineering,Yanshan University,Qinhuangdao,Hebei 066004,China)

机构地区：[1]北京理工大学集成电路与电子学院,北京100081 [2]燕山大学信息科学与工程学院河北省特种光纤与光纤传感重点实验室,河北秦皇岛066004

出　　处：《计量学报》2023年第1期54-61,共8页Acta Metrologica Sinica

基　　金：国家自然科学基金(61971373);河北省自然科学基金(C2020203010);河北省博士在读研究生创新能力培养(CXZZBS2022148)。

摘　　要：提出一种基于改进Faster R-CNN水母检测与识别算法。首先,建立了包含7种水母的数据集;然后,针对ResNeXt(C=32)用于目标检测时出现计算量较大的问题,在保证精确度的前提下,将分支数C设置为8以降低计算量;最后,为解决水母检测时出现的检测精度低和小个体无法检测的问题,在残差网络中引入膨胀卷积。实验结果表明:该算法较VGG16、ResNet101、ResNeXt(C=32)和ResNeXt(C=8)方法,mAP值分别提高了3.15%、2.09%、3.01%和2.36%;F 1-score分别提高了2.53%、1.99%、2.01%和2.31%;loss损失函数收敛值更优,收敛精度趋近于0。P-R曲线、可视化效果分析和水母视频检测的结果证明:该算法的水母检测准确率和水母检测数量明显优于其他算法,检测精度较高,基本可以达到实时监测的要求。A jellyfish detection algorithm based on improved Faster R-CNN is proposed.Firstly,a data set containing 7 species of jellyfishes is established.Secondly,on the premise of ensuring the accuracy,the number of branches C is set to 8 to solve the problem that ResNeXt(C=32)has a high amount of calculation for target detection.Finally,to solve the problems of low detection accuracy and small individuals unable to be recognized,expansion convolution is introduced into the residual network.The experimental results shown that compared with VGG16,ResNet101,ResNeXt(C=32)and ResNeXt(C=8),the mAP value of the proposed algorithm increase by 3.15%,2.09%,3.01%and 2.36%.F 1-score increase by 2.53%,1.99%,2.01%and 2.31%.Loss function convergence value of the proposed algorithm approach to 0.Results of P-R curve,visual analysis and video detection show that the accuracy and detection number of jellyfish by the proposed algorithm is the best,the proposed algorithm has high detection accuracy and can meet the requirements of real-time monitoring.

关 键 词：计量学 水母检测与识别 Faster R-CNN ResNeXt 膨胀卷积 残差网络 

分 类 号：TB96[机械工程—光学工程] TB973[一般工业技术—计量学]