Outliers inevitably exist in image matching

and they may result in a significantly high mismatching ratio when the overlapping region of two images is small. Some normal matching algorithms cause a high mismatching ratio. A robust purifying algorithm for rough matched pairs can be designed to solve this problem by reducing difficulties in image matching. Since Hough transform was proposed

it has been used to detect certain kinds of curves. It establishes a mathematical model for curves and votes in the para-meter space and determines exact parameters of the curve via maximum value in the parameter space. Based on the same voting idea

Hough transform is introduced in this paper to purify rough matched pairs. First

we assume that those truly matched pairs obey a certain transform model equation. Then

a common transform model can be established

and Hough transform is used to obtain parameters of model equation. In particular

each matched pair votes on the corresponding hypersurface

which is in the parameter space and determined by Hough transform. Thus

parameters of the transform model equation can be determined by the global maximum value in the parameter space. Then

all matched pairs that obey model equation are saved. Thus

the rough matched pairs can be purified in this way. Compared with traditional algorithms

such as random sample consensus

the proposed algorithm is not only robust to outliers with a good recall ratio but also more efficient. Moreover

experimental results indicate that the proposed method can be robust when the ratio of outliers is as high as 85%

and even when the ratio is up to 95%

it still can work very well with a probability of 50%. Hough transform can be applied to purify matched pairs

and many experiments prove its feasibility. Corresponding models should be chosen to obtain a good performance when aiming at rigid-body transformation and affine transformation. However

the proposed method is not suitable when many parameters (more than four) exist in the model equation

given that a high-dimensional space determined by parameters of the model equation is memory expensive and time consuming when searching and voting in the high dimensional parameters space.