TY - GEN
T1 - A 7-round parallel hardware-saving accelerator for Gaussian and DoG pyramid construction part of SIFT
AU - Qiu, Jingbang
AU - Huang, Tianci
AU - Ikenaga, Takeshi
PY - 2010/12/29
Y1 - 2010/12/29
N2 - SIFT, short for Scale Invariant Feature Transform, is regarded as one of the most robust feature detection algorithms. The Gaussian and DoG Pyramid Construction part, functioning as computation basis and searching spaces for other parts, proves fatal to the system. In this paper, we present an FPGA-implementable hardware accelerator for this part. Stratified Gaussian Convolution scheme and 7-Round Parallel Computation scheme are introduced to reduce the hardware cost and improve process speed, meanwhile keeping high accuracy. In our experiment, our proposal successfully realizes a system with max clock frequency of 95.0 MHz, and on-system process speed of up to 21 fps for VGA format images. Hardware cost of Slice LUTs is reduced by 12.1% compared with traditional work. Accuracy is kept as high as 98.27% against original software solution. Our proposed structure proves to be suitable for real-time SIFT systems.
AB - SIFT, short for Scale Invariant Feature Transform, is regarded as one of the most robust feature detection algorithms. The Gaussian and DoG Pyramid Construction part, functioning as computation basis and searching spaces for other parts, proves fatal to the system. In this paper, we present an FPGA-implementable hardware accelerator for this part. Stratified Gaussian Convolution scheme and 7-Round Parallel Computation scheme are introduced to reduce the hardware cost and improve process speed, meanwhile keeping high accuracy. In our experiment, our proposal successfully realizes a system with max clock frequency of 95.0 MHz, and on-system process speed of up to 21 fps for VGA format images. Hardware cost of Slice LUTs is reduced by 12.1% compared with traditional work. Accuracy is kept as high as 98.27% against original software solution. Our proposed structure proves to be suitable for real-time SIFT systems.
UR - http://www.scopus.com/inward/record.url?scp=78650424448&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=78650424448&partnerID=8YFLogxK
U2 - 10.1007/978-3-642-12297-2_8
DO - 10.1007/978-3-642-12297-2_8
M3 - Conference contribution
AN - SCOPUS:78650424448
SN - 3642122965
SN - 9783642122965
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 75
EP - 84
BT - Computer Vision, ACCV 2009 - 9th Asian Conference on Computer Vision, Revised Selected Papers
T2 - 9th Asian Conference on Computer Vision, ACCV 2009
Y2 - 23 September 2009 through 27 September 2009
ER -