Multi-Threshold Corner Detection and Region Matching Algorithm Based on Texture Classification
Zetian Tang,Zhao Ding,Ruimin Zeng,Yang Wang,Jun Wen,Lifeng Bian,Chen Yang
DOI: https://doi.org/10.1109/access.2019.2940137
IF: 3.9
2019-01-01
IEEE Access
Abstract:In order to address the unreasonable distributed corners in single threshold Harris detection and expensive computation cost incurred from image region matching performed by normalized cross correlation (NCC) algorithm, multi-threshold corner detection and region matching algorithm based on texture classification are proposed. Firstly, the input image is split into sub-blocks which are classified into four different categories based on the specific texture: flat, weak, middle texture and strong regions. Subsequently, an algorithm is suggested to decide threshold values for different texture type, and interval calculation for the sub-blocks is performed to improve operation efficiency in the algorithm implementation. Finally, based on different texture characteristics, Census, interval-sampled NCC, and complete NCC are employed to perform image matching. As demonstrated by the experimental results, corner detection based on texture classification is capable to obtain a reasonable corner number as well as a more uniform spatial distribution, when compared to the traditional Harris algorithm. If combined with the interval classification, speedup for texture classification is approximately 30%. In addition, the matching algorithm based on texture classification is capable to improve the speed of 26.9%~29.9% while maintaining the comparable accuracy of NCC. In general, for better splicing quality, the overall stitching speed is increased by 14.1%~18.4%. Alternatively, for faster speed consideration, the weak texture region which accounts for a large proportion of an image and provides less effective information can be ignored, for which 23.9%~28.4% speedup can be achieved at the cost of a 1.9%~3.9% reduction in corner points. Therefore, the proposed algorithm is made potentially suited to uniformly distributed corner point calculation and high computation efficiency requirement scenarios.