电子学报 ›› 2017, Vol. 45 ›› Issue (10): 2362-2367.DOI: 10.3969/j.issn.0372-2112.2017.10.008

• 学术论文 • 上一篇    下一篇

基于梯度双面互补特性的级联快速目标检测

谢昭, 吴东涛, 吴克伟, 李洋   

  1. 合肥工业大学计算机与信息学院, 安徽合肥 230009
  • 收稿日期:2016-03-29 修回日期:2016-08-10 出版日期:2017-10-25
    • 作者简介:
    • 谢昭,男,1980年6月出生,合肥工业大学计算机与信息学院副研究员,主要研究方向:计算机视觉、图像处理、模式识别.E-mail:xiezhao@hfut.edu.cn;吴东涛,男,1990年9月出生,合肥工业大学计算机与信息学院硕士研究生,主要研究方向:计算机视觉、图像处理.E-mail:wudt0901@126.com
    • 基金资助:
    • 国家自然科学基金 (No.61273237,No.61503111)

The Cascaded Rapid Object Detection with Double-Sided Complementary in Gradients

XIE Zhao, WU Dong-tao, WU Ke-wei, LI Yang   

  1. School of Computer and Information, Hefei University of Technology, Hefei, Anhui 230009, China
  • Received:2016-03-29 Revised:2016-08-10 Online:2017-10-25 Published:2017-10-25
    • Supported by:
    • National Natural Science Foundation of China (No.61273237, No.61503111)

摘要: 针对目标检测中精度和速度难以兼顾的问题,借助视觉注意理论中的目标感知与识别机制,分析目标描述中梯度幅值与梯度方向信息之间具有的互补性,提出了基于两层级联梯度特征的快速目标检测模型,可有效描述类无关和类相关检测器.一方面,采用梯度幅值特征,从滑动窗口采样中获得候选目标提议,大幅降低了验证窗口的数量,确保检测速度,另一方面,利用级联方式学习训练多个子检测器,可更好实现不同尺度变化下的目标检测精度.PASCAL数据集上的实验结果,解释了级联梯度特征对目标结构描述的有效性,表明了该文方法在与现有先进方法的检测精度相当的前提下,可极大提升检测速度.

关键词: 目标检测, 类无关和类相关, 梯度特征, 级联结构, 互补性

Abstract: To address the dilemma of trade-off between efficiency and accuracy for object detection,based on the mechanism of object perception and recognition in visual attention theory,the two sides derived from gradient feature as magnitude and direction have been revisited to manifest their complementary characteristics.The new rapid object detection model based on two-layer cascade with gradients is motivated,making two types of category-independent and category-dependent detectors efficiently described.On the one hand,gradient magnitude can be used to generate the efficient object proposal in clutter from sliding window samples which guarantees the significant decrease on the number of windows for candidate and speeds up detection.On the other hand,the cascade-architecture in form of multiple sub-detectors can well adapt to the varying scales of different objects resulting in boost of accuracy.Experimental performance in PASCAL presents the effectiveness of cascade structure for gradient features,and demonstrates that our model can dramatically speed up the detection with the advantages of comparable accuracy against the state-of-the-art.

Key words: object detection, category-independent and category-dependent, gradient feature, cascade structure, complementarity

中图分类号: