摘要
为了利用GPU强大的并行处理能力提高图像拷贝检测速度,提出一种基于GPU的图像拷贝检测方法.首先结合GPU的架构设计了尺度不变特征点提取算法——Harris-Hessian算法,通过在低尺度图像上检测特征点,在图像的一系列尺度空间中根据Hessian矩阵的行列式精确确定特征点的位置和尺度,显著地减少了像素级的计算量,并具有更好的并行性;在此基础上建立了图像拷贝检测系统,检测速度得到显著提升.实验结果表明,与基于CPU实现的传统算法相比,Harris-Hessian算法可以获得10~20倍的加速比,并可保证较高的检测精度.在11 250幅的图像库中,使用文中系统检测一幅640×480图像平均只需19.8 ms,并具有95%的正确率,满足了大规模数据下实时应用的需求.
To speed up image copy detection by exploring the powerful computing capability of GPU,a novel GPU-based image copy detection scheme is proposed.Firstly,a new scale-invariant interest point detector-Harris-Hessian(H-H) is designed according to the architecture of GPU.The H-H extracts interest points in low scale and refines their location and scale in a series of scale-space with the determinant of Hessian matrix,which significantly reduces the pixel-level computation complexity and has better parallelism.Then,an image copy detection system based on the H-H is presented,the detection speed is significantly improved.The experimental results show that,compared to the existing CPU-based methods,the H-H achieves up to a speedup factor of 10~20 times and maintains a high detection accuracy.It only takes 19.8 ms for the system to detect a 640×480 image in a dataset of 11 250 images with 95% accuracy rate,which meets the demand of real-time applications under large scale data.
出处
《计算机辅助设计与图形学学报》
EI
CSCD
北大核心
2010年第9期1483-1490,共8页
Journal of Computer-Aided Design & Computer Graphics
基金
国家"九七三"重点基础研究发展计划项目(2007CB311100)
国家"八六三"高技术研究发展计划(2007AA01Z416)
国家自然科学基金(60873165
60802028
60802067)
北京市科技新星计划项目(2007B071)
北京市教育委员会共建项目专项资助