[author_cn_name].[cn_title][J].空军工程大学学报:自然科学版,[year_id],[volume]([issue]):[start_page]-[end_page] 基于空洞卷积和特征融合的单阶段目标检测算法- Single Stage Object Detection Algorithm Based on Dilated Convolution and Feature Fusion
文章摘要
李娟娟, 侯志强, 白玉, 程环环, 马素刚, 余旺盛, 蒲磊.基于空洞卷积和特征融合的单阶段目标检测算法[J].空军工程大学学报:自然科学版,2022,23(1):97-103
基于空洞卷积和特征融合的单阶段目标检测算法
Single Stage Object Detection Algorithm Based on Dilated Convolution and Feature Fusion
  
DOI:
中文关键词: 目标检测  SSD算法  空洞卷积  特征融合
英文关键词: object detection  SSD algorithm  dilated convolution  feature fusion
基金项目:国家自然科学基金(62072370)
作者单位
李娟娟, 侯志强, 白玉, 程环环, 马素刚, 余旺盛, 蒲磊 1.西安邮电大学计算机学院西安710121 2.西安邮电大学陕西省网络数据分析与智能处理重点实验室西安7101213.空军工程大学信息与导航学院西安710077 4.火箭军工程大学作战保障学院西安710025 
摘要点击次数: 83
全文下载次数: 141
中文摘要:
      针对SSD多尺度目标检测过程中存在的目标漏检和错检问题,提出了一种融入多维空洞卷积和多尺度特征融合的目标检测算法。在卷积神经网络输出的多尺度特征中,浅层具有更多的细节信息,深层具有更多的语义信息,根据这一特点,对浅层网络采用了3种多维空洞卷积的浅层特征增强模块,获得具有语义信息的特征图,将增强后的特征图进行下采样,融合不同层的特征;同时在深层网络引入通道注意力模块,对通道进行权重分配,抑制无用信息,提高目标的检测性能。研究结果表明:该算法在PASCAL VOC数据集上检测精度为79.7%,比SSD算法提高了2.4%;在KITTI数据集上检测精度为68.5%,比SSD算法提高了5.1%,检测速度达到了实时性的要求,有效地改善了目标的漏检和错检。
英文摘要:
      Aiming at the problem of missed detection and wrong detection in the process of SSD(Single Shot Multibox Detector)multi scale object detection, this paper proposes a object detection algorithm that incorporates multi dimensional dilated convolution and multi scale feature fusion. Among the multi scale features output by the convolutional neural network, the shallow layer has more detailed information, and the deep layer has more semantic information. According to this feature, this paper adopt three types of multi dimensional dilated convolution shallow layers for the shallow network. The feature enhancement module obtains a feature map with semantic information, down samples the enhanced feature map, and fusion features of different layers; at the same time, a channel attention module is introduced in the deep network to assign weights to channels, suppress useless information, and improve object The detection performance. The research results show that the detection accuracy of the algorithm in this paper is 79.7% on the PASCAL VOC dataset, which is 2.4% higher than the SSD algorithm; the detection accuracy on the KITTI dataset is 68.5%, which is 5.1% higher than the SSD algorithm, and the detection speed reaches real time requirements have effectively improved the missed and wrongly detected object.
查看全文   查看/发表评论  下载PDF阅读器
关闭