Dismiss
InnovationQ will be updated on Sunday, Oct. 22, from 10am ET - noon. You may experience brief service interruptions during that time.
Browse Prior Art Database

Method for Efficient Low-Pass Filtering of Images

IP.com Disclosure Number: IPCOM000106293D
Original Publication Date: 1993-Oct-01
Included in the Prior Art Database: 2005-Mar-20
Document File: 2 page(s) / 83K

Publishing Venue

IBM

Related People

Kasson, JM: AUTHOR

Abstract

A process is disclosed that reduces the number of computations required for some kinds of low-pass filtering of images. In digital image processing, it is frequently necessary to pass an image through a low- pass filter to remove components with high spatial frequency. It is the purpose of this invention to allow efficient computation of low-pass filtered images with large kernels.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 52% of the total text.

Method for Efficient Low-Pass Filtering of Images

      A process is disclosed that reduces the number of computations
required for some kinds of low-pass filtering of images.  In digital
image processing, it is frequently necessary to pass an image through
a low- pass filter to remove components with high spatial frequency.
It is the purpose of this invention to allow efficient computation of
low-pass filtered images with large kernels.

      The main idea of the invention is captured in the flow chart in
the figure.  Repeated subsampling and low-pass filtering  with small
kernels, followed by repeated decimation and low-pass filtering with
small kernels, approximates a low-pass filter with a large kernel.

      First subsampling by a factor of 2 in both the x and y
dimension, and then oversampling by a factor of 2 in both the x and y
dimension, yields an approximation to a filter of size 3k by 3k or 3k
+ 1, by 3k + 1, with each of its entries

1 cover 9k sup 2 or 1 cover (3k + 1)sup2respectively.

      For k = 2, subsampling by a factor of 2 in both the x and y
dimension, oversampling by a factor of 2 in both the x and y
dimension, and using a 3x3 kernel, each of whose entries are 1/9,
performing the above set of operations will yield an approximation to
convolving the image with a 6x6 kernel, each of whose entries is
1/36, or a 7x7 kernel, each of whose entries is 1/49.  If k is
changed to 3, a 12x12 kernel, each of whose entries is 1/144, or a
13x13 kernel, each of whose entries is 1/169, will be approximated.
A k of 4 yields an approximation to a 24x24 kernel, each of whose
entries is 1/156, or a 25x25 kernel, each of whose entries is 1/625,
and a k of 5 yields an approximation to a 48x48 kernel, each of whose
entries is 1/2304, or a 49x49, kernel each of who...