基于轻量级SSD模型的夜间金蝉若虫检测  

Detecting nocturnal golden cicada nymphs using lightweight SSD model

作　　者：刘阳[1,2] 高国琴[1] Liu Yang;Gao Guoqin(School of Electrical and Information Engineering,Jiangsu University,Zhenjiang 212013,China;College of Electronical Information Engineering,Nantong Vocational University,Nantong 226007,China)

机构地区：[1]江苏大学电气信息工程学院,镇江212013 [2]南通职业大学电子信息学院,南通226007

出　　处：《农业工程学报》2022年第10期166-175,共10页Transactions of the Chinese Society of Agricultural Engineering

基　　金：国家自然科学基金项目(51375210);镇江市重点研发计划(GZ2018004);江苏高校优势学科建设工程资助项目;江苏省研究生科研与实践创新计划项目(CXZZ12_0694);南通市科技计划项目(MSZ20155)。

摘　　要：为实现夜间树上金蝉若虫的快速准确检测,该研究以自然环境图像数据集为研究对象,结合近距离实际应用场景,考虑到嵌入式系统的模型小型化和计算过程轻量化,在保持精度指标基本不变的前提下,基于适度削减模型深度、宽度的思路对已有目标检测网络Mobile Net-SSD提出改进。具体措施包括:删除骨干网络末端的小尺寸特征图卷积层,逐级裁剪模型整体宽度、适当增加中高层卷积深度,在目标检测的分类层和预测框的回归层中使用深度可分离卷积代替传统3×3卷积等措施,先后获取3种改进的精简模型以进行比较。夜间图像测试结果表明,在基本保持网络性能的前提下,改进后的模型大小及计算量均呈现大幅减小,其中最优模型大小从原MobileNet-SSD的15.22 MB减少到1.51 MB,模型的浮点运算量也由原先的1.13×10^(9)减少到1.26×10^(8),其平均准确率达90.46%,平均交并比达83.52%,F1分数达92.35%,GPU上的检测速度达179.3帧/s,CPU上的检测速度达到6.49帧/s,与改进前的模型相比具有更好的综合性能,白天图像的试验结果也显示出较好的泛化性能。该文提出的改进模型在大幅减少模型大小及其计算量的同时使模型性能保持在一个较高的水平,更适合部署在移动终端等资源受限设备上,可为金蝉的人工养殖提供有益参考。Golden cicada nymphs have been one of the most commonly edible insect species in the world,due to the extremely high medicinal and nutritional value with the unique taste.Therefore,the breeding of golden cicada nymphs can be brought the high economic value in food production.In this study,a rapid and accurate detection was proposed for the golden cicada nymphs in the natural environment.366 images were captured from the Jiangsu University campus,Zhejiang Province,China,and another 180 images were from the green belt of residential area,and the downloaded from the Internet as supplement.Three procedures were proposed to realize the rapid and accurate detection of golden cicada nymphs,while reducing the model size and computational load.The first procedure was to reduce the number of branches of the original MobileNet-Single-Shot multibox Detection(SSD) network.Six SSD branches were used to detect the objects,each of which was corresponding to the different size of the feature map in original model.A relatively simple task was implemented to avoid the very complex neural network.Therefore,three small-scale branches of 3×3,2×2,and 1×1 were removed to reduce the size of the model and the calculation amount.The second procedure was to reduce the width of MobileNet-SSD backbone,namely the input/output channels of the convolution layer.The number of features was determined in the network,where the enough width was used to ensure each layer learn the rich features.Among them,the width of the network was dominated the size of the parameters and calculation amount of network.A much wider network was used to extract much more the repetitive features,leading to the higher calculation amount without the substantial contribution to the network.Thus,the width of the backbone network was reduced significantly,due to only one type of object to identify in this case.The third procedure was to adjust the depth(the number) of the convolution layers.The depth of the neural network also determined the non-linear expression of the s

关 键 词：图像识别 目标检测 金蝉若虫 MobileNet-SSD 轻量级卷积神经网络 模型尺寸 模型运算量 

分 类 号：S126[农业科学—农业基础科学]