基于改进SSD的海洋生物检测算法  

Marine Organism Detection Algorithm Based on Improved SSD

作　　者：邓权 林明星[1,2,3] DENG Quan;LIN Ming-xing(School of Mechanical Engineering,Shandong University,Jinan 250061,China;Key Laboratory of High Efficiency and Clean Mechanical Manufacture,Jinan 250061,China;National Demonstration Center for Experimental Mechanical Engineering Education,Jinan 250061,China)

机构地区：[1]山东大学机械工程学院,山东济南250061 [2]高效洁净机械制造教育部重点实验室,山东济南250061 [3]机械工程国家级实验教学示范中心,山东济南250061

出　　处：《计算机技术与发展》2022年第4期51-56,共6页Computer Technology and Development

基　　金：山东省自然科学基金(ZR2020ME267);山东省重点研发计划(2019JZZY020703)。

摘　　要：为了满足海洋生物检测对精度和实时性的要求,提出了一种基于改进SSD算法(single shot multibox detector)的海洋生物检测算法。针对SSD算法浅层特征层语义信息不足、小目标检测效果差等问题,设计了特征融合模块和特征增强模块。特征融合模块通过融合不同特征层,丰富了浅层特征层的语义信息以及深层特征层的细节信息,综合上下文信息提高检测效果。为了进一步提高浅层特征层的语义信息,提出了特征增强模块,通过引入空洞卷积以及多尺度的卷积核,综合不同感受野信息以提高改进算法对小目标的检测效果。改进算法在仅增加少量计算量和参数量的情况下,全面提高了算法对海洋生物目标的检测准确率。结果表明,改进算法在海洋生物数据集中的平均精度(mAP)达80.8%,比原始网络提高了5%,检测速度(FPS)为74,略低于SSD算法,但远高于其他改进算法。改进算法能在保持实时性的同时取得较高的检测精度,能够满足海洋生物检测要求。In order to meet the accuracy and real-time requirements of marine organism detection,a marine organism detection algorithm based on improved SSD algorithm(single shot multibox detector) is proposed.For the problem of insufficient semantic information in the shallow feature layer of the SSD algorithm and poor detection of small targets,a feature fusion module and a feature enhancement module are designed.The feature fusion module enriches the semantic information of the shallow feature layer and the detailed information of the deep feature layer by fusing different feature layers,and integrates context information to improve the detection accuracy.In order to further improve the semantic information of the shallow feature layer,a feature enhancement module is proposed,which integrates different receptive field information by introducing dilated convolution and multi-scale convolution kernels to improve the detection effect of the improved algorithm on small targets.The improved algorithm comprehensively improves the accuracy of the algorithm’s detection of marine organism while only adding a small amount of calculation and parameters.Experimental results show that the average accuracy(mAP) of the improved algorithm in the marine organism data set is 80.8%,which is 5% higher than the original network,the detection speed(FPS) is 74,slightly lower than the SSD algorithm,but much higher than other improved algorithms.The improved algorithm can achieve high detection accuracy while maintaining real-time performance,which can meet the requirements of marine organism detection.

关 键 词：SSD 小目标 特征融合 特征增强 实时性 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]