Grayscale Template-Matching Invariant to Rotation, Scale, Translation, Brightness and Contrast
In this paper, we consider the grayscale template-matching problem, invariant to rotation, scale, translation, brightness and contrast, without previous operations that discard grayscale information, like detection of edges, detection of interest points o
- PDF / 2,362,784 Bytes
- 14 Pages / 430 x 660 pts Page_size
- 113 Downloads / 182 Views
bstract. In this paper, we consider the grayscale template-matching problem, invariant to rotation, scale, translation, brightness and contrast, without previous operations that discard grayscale information, like detection of edges, detection of interest points or segmentation/binarization of the images. The obvious “brute force” solution performs a series of conventional template matchings between the image to analyze and the template query shape rotated by every angle, translated to every position and scaled by every factor (within some specified range of scale factors). Clearly, this takes too long and thus is not practical. We propose a technique that substantially accelerates this searching, while obtaining the same result as the original brute force algorithm. In some experiments, our algorithm was 400 times faster than the brute force algorithm. Our algorithm consists of three cascaded filters. These filters successively exclude pixels that have no chance of matching the template from further processing. Keywords: Template matching, RST-invariance, segmentation-free shape recognition.
1 Introduction In this paper, we consider the problem of finding a query template grayscale image Q in another grayscale image to analyze A, invariant to rotation, scale, translation, brightness and contrast (RSTBC), without previous “simplification” of A and Q that discards grayscale information, like detection of edges, detection of interest points and segmentation/binarization. These image-simplifying operations throw away the rich grayscale information, are noise-sensitive and prone to errors, decreasing the robustness of the matching. Moreover, these simplifications cannot be used to find smooth grayscale templates. The “brute force” solution to this problem performs a series of conventional (BCinvariant) template matchings between the image to analyze A and the query template Q. Image Q must be rotated by every angle, translated to every position and scaled by every factor (within some specified range of scale factors) and a conventional BC-invariant template matching is executed for each instance of the transformed Q. Possibly, the brute force algorithm yields the most precise solution to this problem. However, it takes too long and thus is not practical. Our technique, named Ciratefi, substantially accelerates this searching, while obtaining exactly the same result as the D. Mery and L. Rueda (Eds.): PSIVT 2007, LNCS 4872, pp. 100 – 113, 2007. © Springer-Verlag Berlin Heidelberg 2007
Grayscale Template-Matching Invariant to RST
101
original brute force algorithm (disregarding incidental numerical imprecision). In some experiments, our algorithm was 400 times faster than the brute force algorithm and obtained exactly the same results. Fast grayscale RSTBC-invariant template matching is a useful basic operation for many image processing and computer vision tasks, such as visual control [1], image registration [2], and computation of visual motion [3]. Consequently, it has been the object of an intense and thorough stu
Data Loading...
