METHOD AND DEVICE FOR DETECTING INTEREST POINTS IN IMAGE

• US 20160034780A1
• Filed: 10/12/2015
• Published: 02/04/2016
• Est. Priority Date: 04/12/2013
• Status: Active Grant

• Current Assignee: Peking University
• Original Assignee: Peking University
• Inventors: DUAN, LINGYU, CHEN, JIE, WANG, FANGKUN, HUANG, TIEJUN, GAO, WEN
• Granted Patent: US 9,779,324 B2
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G06T 2207/20021 Dividing image into blocks, subimages or windows
• CPC Class Codes: G06T 2207/20024 Filtering details
• CPC Class Codes: G06T 2207/20056 Discrete and fast Fourier transform, [DFT, FFT]
• CPC Class Codes: G06T 3/00 Geometric image transformations in the plane of the image
• CPC Class Codes: G06T 3/40 Scaling of whole images or parts thereof, e.g. expanding or contracting
• CPC Class Codes: G06T 3/4084 in the transform domain, e.g. fast Fourier transform [FFT] domain scaling
• CPC Class Codes: G06T 7/60 Analysis of geometric attributes
• CPC Class Codes: G06V 10/42 Global feature extraction by analysis of the whole pattern, e.g. using frequency domain transformations or autocorrelation
• CPC Class Codes: G06V 10/449 Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
• CPC Class Codes: G06V 10/462 Salient features, e.g. scale invariant feature transforms [SIFT]
• CPC Class Codes: G06V 10/467 Encoded features or binary features, e.g. local binary patterns [LBP]
• CPC Class Codes: G06V 10/50 by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis

Abstract

The present invention provides a method and a device for detecting interest points in an image. The method includes: acquiring an original input image; performing down-sampling processing on the original input image, so as to obtain a plurality of sampling images with different resolutions; dividing each sampling image into a plurality of small image blocks; performing filtering processing on the plurality of small image blocks in each sampling image in sequence by using Laplacian-of-Gaussian filters, so as to obtain filtered images of the plurality of small image blocks in each sampling image; and acquiring interest points in an image in filtered images of the plurality of small image blocks in each sampling image. The present invention is used for solving the problems of more memory consumption and a low detection speed in the prior art.

Claims

1. A method for detecting interest points in an image, comprising:

acquiring an original input image;

performing down-sampling processing on the original input image, so as to obtain a plurality of sampling images with different resolutions;

dividing each sampling image into a plurality of small image blocks;

performing filtering processing on the plurality of small image blocks in each sampling image in sequence by using Laplacian-of-Gaussian filters, so as to obtain filtered images of the plurality of small image blocks in each sampling image; and

acquiring interest points in an image in each sampling image, according to the filtered images of the plurality of small image blocks in each sampling image.

2. The method according to claim 1, wherein, the dividing each sampling image into the plurality of small image blocks, comprises:

dividing each sampling image into a plurality of small square image blocks having a width of X and a height of Y, wherein, both X and Y are positive integers, if the small image block at the boundary of the sampling image has a width less than X or a height less than Y, then filling the small image block at the boundary of the sampling image with pixels;

filling each small square image block having a width of X and a height of Y with pixels, so that a filled small square image block has a width of X+M−

1 and a height of Y+M−

1, wherein M is an positive integer;

performing a discrete Fourier transform on the filled small square image block, so as to obtain frequency domain small image blocks.

3. The method according to claim 2, wherein, the performing filtering processing on the plurality of small image blocks in each sampling image in sequence by using the Laplacian-of-Gaussian filters, so as to obtain filtered images of the plurality of small image blocks in each sampling image, comprises:

performing multiple filtering processing on the plurality of small image blocks in each sampling image by using frequency domain Laplacian-of-Gaussian filters, so as to obtain frequency domain Laplacian-of-Gaussian response images of the plurality of small image blocks in each sampling image;

performing an inverse discrete Fourier transform on each of the frequency domain Laplacian-of-Gaussian response images of the plurality of small image blocks in each sampling image, so as to obtain filtered images of the plurality of small image blocks in each sampling image.

4. The method according to claim 3, wherein, before performing filtering processing on the plurality of small image blocks in each sampling image in sequence by using the Laplacian-of-Gaussian filters, so as to obtain filtered images of the plurality of small image blocks in each sampling image, the method comprises:

generating a plurality of square spatial domain two-dimensional Gaussian filters, according to two-dimensional Gaussian kernel functions and preset Gaussian parameters, and the plurality of square spatial domain two-dimensional Gaussian filters have a maximum width of M;

generating a square spatial domain two-dimensional Laplacian filter, according to a second-order Laplacian operator function, if the square spatial domain two-dimensional Laplacian filter has a width less than M, filling the square spatial domain two-dimensional Laplacian filter with pixels, so that a filled square spatial domain two-dimensional Laplacian filter has a width of M;

converting the plurality of square spatial domain two-dimensional Gaussian filters into a plurality of frequency domain Gaussian filters;

converting the filled square spatial domain two-dimensional Laplacian filter into a frequency domain Laplacian filter;

generating a plurality of frequency domain Laplacian-of-Gaussian filters, according to the plurality of frequency domain Gaussian filters and the frequency domain Laplacian filter;

before converting the plurality of square spatial domain two-dimensional Gaussian filters into the plurality of frequency domain Gaussian filters, the method further comprises;

performing one-to-one correspondence replacement with a pixel value of the square spatial domain two-dimensional Gaussian filter having a width of M, for an area having a width of M on the upper left of a square matrix having a width of X+M−

1 and a height of Y+M−

1, and performing a cyclic shift to the square matrix with M/2 pixels leftwards and upwards, wherein, all initial pixel values for the square matrix are 0;

before converting the filled square spatial domain two-dimensional Laplacian filter into the frequency domain Laplacian filter, the method further comprises;

performing one-to-one correspondence replacement with a pixel value of the square spatial domain two-dimensional Laplacian filter having a width of M, for an area having a width of M on the upper left of a square matrix having a width of X+M−

1 and a height of Y+M−

1, and performing a cyclic shift to the square matrix with M/2 pixels leftwards and upwards, wherein, all initial pixel values for the square matrix are 0.

5. The method according to claim 1, wherein, the acquiring the interest points in the image in each sampling image, according to the filtered images of the plurality of small image blocks in each sampling image, comprises:

acquiring interest points in an image of each small image block, according to the filtered images of the plurality of small image blocks in each sampling image;

acquiring interest points in an image in each sampling image, according to the interest points in the image of each small image block;

or,

merging filtered images of the plurality of small image blocks in each sampling image, so as to obtain a filtered image of each sampling image;

acquiring interest points in each sampling image, according to the filtered image of each sampling image.

6. A device for detecting interest points in an image, comprising:

an acquiring module, configured to acquire an original input image;

a processing module, configured to perform down-sampling processing on the original input image, so as to obtain a plurality of sampling images with different resolutions;

a dividing module, configured to divide each sampling image into a plurality of small image blocks;

a filtering module, configured to perform filtering processing on the plurality of small image blocks in each sampling image in sequence by using Laplacian-of-Gaussian filters, so as to obtain filtered images of the plurality of small image blocks in each sampling image;

wherein the acquiring module is further configured to acquire interest points in an image in each sampling image, according to the filtered images of the plurality of small image blocks in each sampling image.

7. The device according to claim 6, further comprising:

a filling module and a converting module;

wherein

the dividing module is specifically configured to divide each sampling image into a plurality of small square image blocks having a width of X and a height of Y, wherein, both X and Y are positive integers, if the small image block at the boundary of the sampling image has a width less than X or a height less than Y, then the filling module fills the small image block at the boundary of the sampling image with pixels;

the filling module is further configured to fill each small square image block having a width of X and a height of Y with pixels, so that a filled small square image block has a width of X+M−

1 and a height of Y+M−

1, wherein M is an positive integer;

the converting module is configured to perform a discrete Fourier transform on the filled small square image block, so as to obtain frequency domain small image blocks.

8. The device according to claim 7, wherein, the filtering module is specifically configured to perform multiple filtering processing on the plurality of small image blocks in each sampling image by using frequency domain Laplacian-of-Gaussian filters, so as to obtain frequency domain Laplacian-of-Gaussian response images of the plurality of small image blocks in each sampling image;

the converting module is further configured to perform an inverse discrete Fourier transform on each of the frequency domain Laplacian-of-Gaussian response images of the plurality of small image blocks in each sampling image, so as to obtain filtered images of the plurality of small image blocks in each sampling image.

9. The device according to claim 8, further comprising:

a generating module;

wherein

before the filtering module performs filtering processing on the plurality of small image blocks in each sampling image in sequence by using a frequency domain Laplacian-of-Gaussian filters, so as to obtain filtered images of the plurality of small image blocks in each sampling image, the generating module is configured to generate a plurality of square spatial domain two-dimensional Gaussian filters, according to two-dimensional Gaussian kernel functions and preset Gaussian parameters, and the plurality of square spatial domain two-dimensional Gaussian filters have a maximum width of M;

the generating module is further configured to generate a square spatial domain two-dimensional Laplacian filter, according to a second-order Laplacian operator function, if the square spatial domain two-dimensional Laplacian filter has a width less than M, fill the square spatial domain two-dimensional Laplacian filter with pixels, so that a filled square spatial domain two-dimensional Laplacian filter has a width of M;

the converting module is further configured to convert the plurality of square spatial domain two-dimensional Gaussian filters into a plurality of frequency domain Gaussian filters;

converting the filled square spatial domain two-dimensional Laplacian filter into a frequency domain Laplacian filter;

generate a plurality of frequency domain Laplacian-of-Gaussian filters, according to the plurality of frequency domain Gaussian filters and the frequency domain Laplacian filter;

before the converting module converts the plurality of square spatial domain two-dimensional Gaussian filters into the plurality of frequency domain Gaussian filters, the filling module is further configured to perform one-to-one correspondence replacement with a pixel value of the square spatial domain two-dimensional Gaussian filter having a width of M, for an area having a width of M on the upper left of a square matrix having a width of X+M−

1 and a height of Y+M−

1, and perform a cyclic shift to the square matrix with M/2 pixels leftwards and upwards, wherein, all initial pixel values for the square matrix are 0;

before the converting module converts the filled square spatial domain two-dimensional Laplacian filter into the frequency domain Laplacian filter, the filling module is further configured to perform one-to-one correspondence replacement with a pixel value of the square spatial domain two-dimensional Laplacian filter having a width of M, for an area having a width of M on the upper left of a square matrix having a width of X+M−

1 and a height of Y+M−

1, and perform a cyclic shift to the square matrix with M/2 pixels leftwards and upwards, wherein, all initial pixel values for the square matrix are 0.

10. The device according to claim 6, wherein, the acquiring module acquires interest points in an image in each sampling image, according to the filtered images of the plurality of small image blocks in each sampling image, the acquiring module is specifically configured to:

acquire interest points in an image of each small image block, according to the filtered images of the plurality of small image blocks in each sampling image;

acquire interest points in an image in each sampling image, according to the interest points in the image of each small image block;

or, the acquiring module is specifically configured to;

merge filtered images of the plurality of small image blocks in each sampling image, so as to obtain a filtered image of each sampling image;

acquire interest points in each sampling image, according to the filtered image of each sampling image.