Design of a Convolutional Two-Dimensional Filter in FPGA for Image Processing Applications

Gian Licciardo,Carmine Cappetta,Luigi Di Benedetto

doi:10.3390/computers6020019

Abstract

Exploiting the Bachet weight decomposition theorem, a new two-dimensional filter is designed. The filter can be adapted to different multimedia applications, but in this work it is specifically targeted to image processing applications. The method allows emulating standard 32 bit floating point multipliers using a chain of fixed point adders and a logic unit to manage the exponent, in order to obtain IEEE-754 compliant results. The proposed design allows more compact implementation of a floating point filtering architecture when a fixed set of coefficients and a fixed range of input values are used. The elaboration of the data proceeds in raster-scan order and is capable of directly processing the data coming from the acquisition source thanks to a careful organization of the memories, avoiding the implementation of frame buffers or any aligning circuitry. The proposed architecture shows state-of-the-art performances in terms of critical path delay, obtaining a critical path delay of 4.7 ns when implemented on a Xilinx Virtex 7 FPGA board.

Highlights

In image processing applications such as object recognition, segmentation or Visual Search (VS), a filtering stage is always required
We developed a memory-reduced implementation of the Bachet multiplier, which is capable of obtaining the other Fh,j λi coefficients starting from Fh,j λ0 and their 2’s complement counterparts
In typical VS applications a minimum size K = 3 is fixed for the kernel, this size has been chosen for the Gaussian kernel

Summary

Introduction

In image processing applications such as object recognition, segmentation or Visual Search (VS), a filtering stage is always required. During the last years the computational complexity of designs dedicated to this kind of problems underwent a critical growth, due to more complex architecture designs and an increase in the amount of data to be processed, related to higher resolution images. VS search applications are among those which require most in terms of complexity, due to the high amount of calculations needed for searching and extracting the features in the input frame [3] In these kind of applications, SW implementations are usually limited in terms of maximum achievable frequency, which does not allow for real-time processing in the case of higher resolution images unless using strong simplifications of the algorithms. These elaborations are required in the Scale Invariant Feature Transform (SIFT) [3,4]

Objectives

Results

Conclusion