Real-time detection of rotated patterns using FPGA

In this paper, we describe an approach for real-time detection of rotated patterns in an image using FPGA. In many approaches, a two-dimensional pattern and the target regions in the image are transformed to one-dimensional arrays or several pattern specific values which are invariant to rotation, and the regions in which the pattern may exist are listed up using them. These approaches make it possible to narrow down the candidate regions with less computational cost, but the sensitivity depends on how to define the transform method. In our approach, each region in a given image is directly compared with a sequence of the rotated patterns using direct cross-correlation, and the positions and rotation angles of the patterns in the image can be found during a scan. This approach requires high computational cost, but by calculating the cross-correlations of a region and the sequence of the rotated patterns incrementally starting from the non-rotated pattern, and by reducing the number of the operations for the cross-correlation considering the equality of the pixel values in the pattern, it becomes possible to realize real-time detection of the rotated patterns with one FPGA.