Detecting Patterns in Various Size and Angle Using FPGA

In this paper, we describe an approach for detecting patterns in various size and angle using FPGA. In many approaches, features of a given pattern which are invariant to scaling and/or rotation are defined in advance, and those features are searched in a given image. These approaches make it possible to narrow down the candidate regions with less computational cost, but the sensitivity depends on how to define the features. In our approach, the image is downscaled by α^k and α^l along the x and y axes (k, l = 0, 1, 2, ..., n), and the regions in the downscaled images are compared with several sequences of the templates which are generated from the given pattern using direct cross-correlation. This approach requires high computational cost, but by calculating the cross-correlations incrementally starting from the nonrotated pattern, it becomes possible to detect the rotated patterns in various size and angle with one FPGA.