基于改进SSD的青瓜检测算法  

Cucumber detection algorithm based on improved SSD

作　　者：曾乾 李博 Zeng Qian;Li Bo(College of Mechanical and Electrical Engineering,University of Electronic Science and Technology of China,Chengdu 611731,China;College of Mechanical and Electrical Engineering,University of Electronic Science and Technology of China,Zhongshan Institute,Zhongshan 528402,China)

机构地区：[1]电子科技大学机械与电气工程学院,成都611731 [2]电子科技大学中山学院机电工程学院,中山528402

出　　处：《国外电子测量技术》2023年第4期158-165,共8页Foreign Electronic Measurement Technology

基　　金：2021年广东省教育厅普通高校重点领域专项(421N34);2020年广东省普通高校创新团队项目—机器人与智能装备团队(2020KCXTD035)项目资助。

摘　　要：针对复杂近色背景下青瓜目标识别率低、定位效果不佳等问题,提出一种基于SSD的循环融合特征增强(CFFE-SSD)目标检测模型。首先,对SSD的前4个有效特征层进行循环特征融合,使低层特征层和高层特征层的信息得到有效利用;其次,针对青瓜目标的特殊长宽比以及重叠现象,使用K-means算法改进先验框的默认尺寸以及长宽比,提出以DIoU-NMS替换普通NMS;最后,将ECA注意力机制引入循环特征融合模块,增强网络特征提取能力。实验结果表明,改进CFFE-SSD模型AP@0.5达到了96.63%,提升了4.61%;AP@0.75达到了89.02%,提升了7.14%,检测速度达到144 fps,边框回归精度更高,能有效满足青瓜自动采摘的需求。Aiming at the problems of low target recognition rate and poor localization effect of cucumber in complex nearcolor background,a cyclic fusion feature enhanced based on SSD(CFFE-SSD)target detection model is proposed.First,cyclic feature fusion is performed on the first four effective feature layers of SSD,so that the information of low-level feature layers and high-level feature layers can be effectively used.Secondly,in the view of special aspect ratio and overlapping phenomenon of the cucumber target,the K-means algorithm is used to improve the default size and aspect ratio of the a priori frame,and meanwhile,DIoU-NMS is proposed to replace ordinary NMS.Finally,the efficient channel attention module is introduced into the cyclic feature fusion module to enhance the feature extraction ability of the network.The ex-perimental results show that the AP@0.5 of the improved CFFE-SSD model proposed in this paper reaches 96.63%,an increase of 4.61%,and the AP@0.75 reaches 89.02%,an in-crease of 7.14%.The detection speed reaches 144 fps,and the frame regression accuracy is higher.Effectively meet the needs of automatic cucumber picking.

关 键 词：青瓜检测 图像识别 深度学习 卷积神经网络 目标检测 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]