罗会兰,
江西理工大学信息工程学院 赣州 341000
基金项目:国家自然科学基金(61862031, 61462035)；江西省教育厅科学技术研究项目(GJJ200859, GJJ200884)；江西省赣州市“科技创新人才计划” 项目
详细信息
作者简介:陈鸿坤：男，1995年生，硕士，研究方向为目标检测
罗会兰：女，1974年生，教授，博士后，主要研究方向为机器学习、模式识别
通讯作者:罗会兰 luohuilan@sina.com
中图分类号:TN911.73; TP391.4计量
文章访问数:498
HTML全文浏览量:296
PDF下载量:152
被引次数:0
出版历程
收稿日期:2020-03-03
修回日期:2020-11-27
网络出版日期:2020-12-07
刊出日期:2021-07-10
Multi-scale Semantic Information Fusion for Object Detection
Hongkun CHEN,Huilan LUO,
School of Information Engineering, Jiangxi University of Science and Technology, Ganzhou 341000, China
Funds:The National Natural Science Foundation of China (61862031, 61462035), The Science and Technology Research Project of Jiangxi Provincial Department of Education (GJJ200859, GJJ200884), Ganzhou City, Jiangxi Province “Technology Innovation Talent Program” Project
摘要
摘要:针对当前目标检测算法对小目标及密集目标检测效果差的问题，该文在融合多种特征和增强浅层特征表征能力的基础上提出了浅层特征增强网络(SEFN)，首先将特征提取网络VGG16中Conv4_3层和Conv5_3层提取的特征进行融合形成基础融合特征；然后将基础融合特征输入到小型的多尺度语义信息融合模块中，得到具有丰富上下文信息和空间细节信息的语义特征，同时把语义特征和基础融合特征经过特征重利用模块获得浅层增强特征；最后基于浅层增强特征进行一系列卷积获取多个不同尺度的特征，并输入各检测分支进行检测，利用非极大值抑制算法实现最终的检测结果。在PASCAL VOC2007和MS COCO2014数据集上进行测试，模型的平均精度均值分别为81.2%和33.7%，相对于经典的单极多盒检测器(SSD)算法，分别提高了2.7%和4.9%；此外，该文方法在检测小目标和密集目标场景上，检测精度和召回率都有显著提升。实验结果表明该文算法采用特征金字塔结构增强了浅层特征的语义信息，并利用特征重利用模块有效保留了浅层的细节信息用于检测，增强了模型对小目标和密集目标的检测效果。
关键词:目标检测/
特征金字塔/
特征融合/
通道注意力/
单极多盒检测器模型
Abstract:Current object detection algorithms have poor detection results on small targets and dense targets. To address this challenge, a Shallow Enhanced Feature Network (SEFN) is proposed in this paper, which is based on the fusion of multiple features and enhanced shallow feature characterization capabilities. Firstly, the features extracted from the Conv4_3 layer and Conv5_3 layer are combined to form basic fusion features. Then the basic fusion features are inputted into a small multi-scale semantic information fusion module to obtain semantic features of rich contextual information and spatial detail information. The semantic features are fused into the basics features by the feature reuse module to obtain shallow enhanced features. Finally, a series of convolutions are performed based on the shallow enhanced features to obtain multiple features with different scales. Multiple detection branches are then constructed based on the features of different scales. The non-maximum suppression algorithm is used to achieve the final detection. The average accuracy of the proposed model is 81.2% and 33.7% on the PASCAL VOC2007 and MS COCO2014 datasets respectively, which is 2.7% and 4.9% higher than the classic Single Shot multibox Detector (SSD) algorithm. In addition, on detecting small targets in dense target scenes, the detection accuracy and recall rate of the proposed method are significantly improved. The experimental results show that the feature pyramid structure can enhance the semantic information of shallow features, and the feature reuse module can effectively retain shallow detail information for detection, so the proposed method can get better detection performance on small targets and dense targets.
Key words:Object detection/
Feature pyramid/
Feature fusion/
Channel attention/
Single Shot multibox Detector (SSD) model
PDF全文下载地址:
https://jeit.ac.cn/article/exportPdf?id=6cc6492c-2069-4739-b4f8-236ad7535d9c
