计算机工程 ›› 2012, Vol. 38 ›› Issue (22): 232-235.doi: 10.3969/j.issn.1000-3428.2012.22.058

• 工程应用技术与实现 • 上一篇    下一篇

一种积分图的分治计算方法

张羽立,戴 麟,周薇娜,邹 垚,曾晓洋   

  1. (复旦大学专用集成电路与系统国家重点实验室,上海 201203)
  • 收稿日期:2012-02-01 修回日期:2012-03-30 出版日期:2012-11-20 发布日期:2012-11-17
  • 作者简介:张羽立(1987-),男,硕士研究生,主研方向:图像处理,信息安全;戴 麟,助理研究员;周薇娜,博士;邹 垚,博士研究生;曾晓洋,教授、博士生导师
  • 基金项目:
    国家自然科学基金资助项目(60776028);教育部基金资助重点项目(109055)

A Divide and Conquer Calculation Approach of Integral Graph

ZHANG Yu-li, DAI Lin, ZHOU Wei-na, ZOU Yao, ZENG Xiao-yang   

  1. (State Key Laboratory of ASIC & System, Fudan University, Shanghai 201203, China)
  • Received:2012-02-01 Revised:2012-03-30 Online:2012-11-20 Published:2012-11-17

摘要: AdaBoost算法在处理大量图片时需要进行长时间的预处理。为此,提出一种用硬件并行计算图像各个局部积分图的方法,通过AdaBoost算法用少数特征淘汰大量无关图像,牺牲单个特征计算时间,避免对大量图片进行预处理操作。实验结果证明,该方法可以加快大量高像素图像的识别速度,在无关图像占整体98%的情况下,速度比传统算法提高31.8%。

关键词: AdaBoost算法, 积分图, 特征矩形计算, 递归, 乒乓算法

Abstract: This paper proposes a method to calculate integral images on part of the image instead of global scope by hardware, thus avoiding most time-consuming pretreatment. The cost of doing so and the speed of single rectangular feature calculation are acceptable due to the characteristic of AdaBoost that few rectangular features are needed to drop out most unrelated images. With traditional AdaBoost, this method can accelerate the analysis of a great number of high-resolution images. As result, when irrelative image occupies 98% of whole, the performance of this approach will be 31.8% higher than traditional ones.

Key words: AdaBoost algorithm, integral graph, feature rectangular calculation, recursion, ping-pong algorithm

中图分类号: