Image pyramid generator is an important component for many computer vision applications. Most pyramid generators only produce few levels of down-sampled output images from an input image. In this paper, we present the design of an image pyramid generator optimized for FPGAs that can create a pyramid of up to 24 levels of down-sampled resolutions from the input image, with the scaling factors ranging from 8/32 to 31/32. Since image pyramid processing is essential to many coarse-to-fine solutions of computer vision problems, an image pyramid with more resolution levels can potentially lead to better accuracy for estimation/recognition problems. Furthermore, the down-sampling filters used in the proposed design is based on the Sine-windowed Sinc function filter. Therefore, it preserves more image details than popular low complexity filters such as the bilinear or bicubic interpolation filters. To maintain application flexibility, a HW-SW codesigned approach is adopted so that the granularity levels can be chosen flexibly. The proposed circuit is verified on an FPGA development board with a Xilinx Kintex-7 device. The experimental results show that the design is very promising for real-time computer vision applications.
|
|
Reference:
|
Souce code: The source code of the image pyramid generator, including the top-level controller software in C and the AXI4 master accelerator in Verilog will be available in this page when the paper review process is finished. For now, you can download the source code of the C model we used to implement the circuits here: the C model of the 24-level image pyramid generator. The HW-SW system is developped using Xilinx Vivado, and the source code of the entire project workspace for the Xilinx KC-705 FPGA platform will be release under the Apache License 2.0. |