文章摘要
王建浩,呼子宇,张翮翔,代言,郝若欣,高泽航.基于深度信息融合的密集目标检测[J].高技术通讯(中文),2022,32(9):914~921
基于深度信息融合的密集目标检测
Dense target detection based on deep information fusion
  
DOI:10.3772/j.issn.1002-0470.2022.09.004
中文关键词: 信息融合; 图像处理; 目标检测; 深度学习; 数据聚类
英文关键词: information fusion, image processing, target detection, deep learning, data clustering
基金项目:
作者单位
王建浩 (燕山大学电气工程学院秦皇岛 066004) 
呼子宇 (燕山大学电气工程学院秦皇岛 066004) 
张翮翔 (燕山大学电气工程学院秦皇岛 066004) 
代言 (燕山大学电气工程学院秦皇岛 066004) 
郝若欣 (燕山大学电气工程学院秦皇岛 066004) 
高泽航 (燕山大学电气工程学院秦皇岛 066004) 
摘要点击次数: 679
全文下载次数: 638
中文摘要:
      针对密集行人检测精度低的问题,提出基于深度信息融合的密集目标检测方法——YOLOv4-SD。该方法通过将single scale Retinex (SSR)与目标检测算法信息融合,增强输入图像质量,凸显图像中更多的信息元素;并对YOLOv4算法中特征融合层进行改进,增加原始图像特征的利用率,深度优化特征融合层的网络结构。在VOC 2012等数据集上进行对比实验,结果表明在保持检测速度的前提下,该算法的平均检测精度和交并比分别提高了7.7%和5.2。对于数据集中边缘低像素或高重叠的行人目标,YOLOv4-SD算法能够较为准确地检测出特殊目标具体位置。
英文摘要:
      Aiming at the problem of low accuracy of dense pedestrian detection, YOLOv4-SD a dense target detection method based on deep information fusion, is proposed. The method combines the information of single scale Retinex (SSR) and the target detection algorithm to improve the quality of the input image and highlight more information elements in the image. By improving the feature fusion layer in the YOLOv4 algorithm, the network structure of the feature fusion layer is deeply optimized to achieve the purpose of improving the utilization of original image features. Comparison experiments are conducted on data sets VOC 2012, the experimental results show that the average detection accuracy and intersection ratio of the algorithm are increased by 7.7% and 5.2, while maintaining the detection speed. For predestrian targets with low pixels or high overlap at the edge of the data set images, YOLOv4-SD algorithm can accurately detect the specific location special targets.
查看全文   查看/发表评论  下载PDF阅读器
关闭

分享按钮