基于改进Faster R-CNN的瓶装饮料商品目标检测方法  

Target Detection Method of Bottled Drinks Based on Improved Faster R-CNN

作　　者：陈欢欢 汪建晓[1] 王高杰 陈勇[1] CHEN Huanhuan;WANG Jianxiao;WANG Gaojie;CHEN Yong(Foshan University of Science and Technology,Foshan 528000,China;Guangdong Shunde Innovation and Design Institute,Foshan 528000,China)

机构地区：[1]佛山科学技术学院,佛山528000 [2]广东顺德创新设计研究院,佛山528000

出　　处：《集成技术》2021年第3期1-11,共11页Journal of Integration Technology

基　　金：广东省科技计划项目(2017A010102018)。

摘　　要：该文以无人售货机售卖瓶装饮料商品为研究场景,提出一种基于改进Faster R-CNN算法的瓶装饮料商品目标检测方法。首先,采用残差网络ResNet-50进行特征提取,加深网络对目标特征的提取和学习的深度;然后,根据瓶装饮料商品形态学特征,增加区域建议网络(Regional Proposal Network)的锚框数量和样式;最后,基于所提出的多维特征图融合网络,增强了网络对小目标的检测性能。实验结果表明,模型训练迭代10000次后的损失值趋于收敛,10个类别的瓶装饮料商品平均精度值均在90%以上,综合检测识别率平均精度均值(MAP)为93.26%,较改进前的模型测试精度提升20%,取得良好检测效果。This paper presents an improved faster R-CNN algorithm based on the application of unmanned vending machine selling bottled drinks.Firstly,the residual network ResNet-50 is used as the feature extraction network to deepen the depth of target feature extraction and learning.Then,the number and style of anchor frame in regional proposal network(RPN)is improved according to the morphological characteristics of bottled beverage products.Finally,a multi-dimensional feature map fusion network is proposed to enhance the detection performance of small targets.The experimental results showed that,the loss value tends to converge after 10000 iterations of model training.Average precision values of 10 categories of bottled beverage products are all larger than 90%.And the comprehensive detection recognition rate mean average precision value is 93.26%,which is improved 20%compared with the original model.

关 键 词：Faster R-CNN 目标检测 残差网络 区域提议网络 多维特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]