WO2014166377A1 - 图像兴趣点检测方法和装置 - Google Patents

图像兴趣点检测方法和装置 Download PDF

Info

Publication number
WO2014166377A1
WO2014166377A1 PCT/CN2014/074927 CN2014074927W WO2014166377A1 WO 2014166377 A1 WO2014166377 A1 WO 2014166377A1 CN 2014074927 W CN2014074927 W CN 2014074927W WO 2014166377 A1 WO2014166377 A1 WO 2014166377A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
sampled
square
images
gaussian
Prior art date
Application number
PCT/CN2014/074927
Other languages
English (en)
French (fr)
Inventor
段凌宇
王仿坤
陈杰
黄铁军
高文
Original Assignee
北京大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 北京大学 filed Critical 北京大学
Publication of WO2014166377A1 publication Critical patent/WO2014166377A1/zh
Priority to US14/880,981 priority Critical patent/US9779324B2/en

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T3/00Geometric image transformations in the plane of the image
    • G06T3/40Scaling of whole images or parts thereof, e.g. expanding or contracting
    • G06T3/4084Scaling of whole images or parts thereof, e.g. expanding or contracting in the transform domain, e.g. fast Fourier transform [FFT] domain scaling
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T3/00Geometric image transformations in the plane of the image
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T3/00Geometric image transformations in the plane of the image
    • G06T3/40Scaling of whole images or parts thereof, e.g. expanding or contracting
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/60Analysis of geometric attributes
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/40Extraction of image or video features
    • G06V10/42Global feature extraction by analysis of the whole pattern, e.g. using frequency domain transformations or autocorrelation
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/40Extraction of image or video features
    • G06V10/44Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G06V10/443Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
    • G06V10/449Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/40Extraction of image or video features
    • G06V10/46Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
    • G06V10/462Salient features, e.g. scale invariant feature transforms [SIFT]
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/40Extraction of image or video features
    • G06V10/50Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/20Special algorithmic details
    • G06T2207/20021Dividing image into blocks, subimages or windows
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/20Special algorithmic details
    • G06T2207/20024Filtering details
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/20Special algorithmic details
    • G06T2207/20048Transform domain processing
    • G06T2207/20056Discrete and fast Fourier transform, [DFT, FFT]
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/40Extraction of image or video features
    • G06V10/46Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
    • G06V10/467Encoded features or binary features, e.g. local binary patterns [LBP]

Definitions

  • the present invention relates to image processing technologies, and in particular, to an image point of interest detection method and apparatus. Background technique
  • Image point of interest detection technology refers to finding a certain number of marks in the image. These marks have the characteristics that they are also in another image containing the same object. These markers can be detected at the same location of the object, even if the two images have different sizes or are taken under different lighting conditions or angles.
  • the image point of interest detecting device downsamples the original image to generate a pyramid including a series of images whose resolution is gradually reduced; and uses a plurality of spatial Gaussian filters to acquire a multiscale image of each layer of the image pyramid.
  • a multi-scale Gaussian Laplacian response image of each layer of image is obtained using a spatial Laplacian filter, and points of interest of each layer of the image are obtained according to the multi-scale Gaussian Laplacian response image.
  • the present invention provides an image point detection method and apparatus for solving the problem that the memory resources are occupied in the prior art and the detection speed is low.
  • a first aspect of the present invention provides an image point of interest detection method, including: acquiring an original input image; Performing down-sampling processing on the original input image to obtain a plurality of sample images of different resolutions;
  • an image point of interest detecting apparatus including: an acquiring module, configured to acquire an original input image;
  • a processing module configured to perform a downsampling process on the original input image to obtain a plurality of sample images of different resolutions
  • a dividing module configured to divide each of the sampled images into a plurality of image small blocks
  • a filtering module configured to sequentially filter, by using a Gaussian Laplace filter, a plurality of image small blocks in each of the sampled images Obtaining a filtered image of a plurality of image patches in each of the sampled images
  • the acquiring module is further configured to acquire an image interest point of each of the sample images according to a filtered image of a plurality of image patches in each of the sampled images.
  • FIG. 1 is a flowchart of an embodiment of an image point of interest detection method provided by the present invention
  • FIG. 2 is an implementation of dividing an image of a plurality of image blocks into an image point of interest detection method according to the present invention
  • FIG. 3 is a schematic diagram of dividing a sample image into a plurality of square image patches of 128*128 square pixels;
  • FIG. 4 is a flowchart of an embodiment of sequentially filtering a plurality of image patches in each sampled image by using a Gaussian Laplacian filter in the image point of interest detection method provided by the present invention
  • FIG. 5 is a schematic diagram of performing image point of interest searching in a three-dimensional space composed of a plurality of filtered images corresponding to image patches;
  • FIG. 7 is a schematic structural diagram of an embodiment of an image point of interest detecting apparatus according to the present invention
  • FIG. 8 is a schematic structural diagram of still another embodiment of an image point of interest detecting apparatus according to the present invention.
  • Fig. 9 is a schematic structural view showing another embodiment of an image point of interest detecting apparatus according to the present invention. detailed description
  • FIG. 1 is a flowchart of an embodiment of an image point of interest detection method according to the present invention.
  • an execution subject of an image point of interest detection method according to the present invention may be a camera, a camera, a camera, a computer, or a mobile phone.
  • a module with processing functions in the device, such as a CPU. The method includes:
  • the image point of interest detecting device acquires an original input image.
  • the original input image can be an image obtained by a device such as a camera, a video camera, a camera, a computer, or a mobile phone that can take pictures.
  • a device such as a camera, a video camera, a camera, a computer, or a mobile phone that can take pictures.
  • the image point of interest detecting device performs downsampling processing on the original input image to obtain a plurality of sample images of different resolutions.
  • the downsampling process refers to that the image point of interest detecting device filters the original input image, and then performs sampling with a length of 2 to generate a multi-stage sampled image with a gradually decreasing resolution.
  • the filtering method used before sampling is generally Gaussian low-pass filtering or domain mean filtering.
  • the image point of interest detecting device may perform image point of interest detection on the sampled image of each resolution, and then combine the detected image points of interest of all the sampled images as points of interest of the original input image.
  • the image point of interest detecting device divides each sample image into a plurality of image patches.
  • the plurality of image patches obtained by the image point of interest detecting device for each sample image may be square or rectangular.
  • the image point of interest detecting device sequentially filters a plurality of image patches in each sampled image by using a Gaussian Laplacian filter to obtain a filtered image of the plurality of image patches in each of the sampled images.
  • the Gaussian Laplacian filter is generated according to a Gaussian filter and a Laplacian filter
  • the image point of interest detecting device can use a Gaussian Laplacian filter to block a plurality of images in each sampled image.
  • the image point of interest detecting device may further split the Gaussian Laplacian filter into a Gaussian filter and a Laplacian filter, using Gaussian The filter and the Laplacian filter sequentially filter a plurality of image patches in each sampled image.
  • the use of a Gaussian filter to filter a plurality of image patches in each sampled image is to construct a multi-scale layer of each image patch in the sampled image.
  • the Laplacian filter is used to calculate the response value of each pixel in the multi-scale layer of each image patch so that the image point of interest detecting device can be based on the multi-scale layer of each image patch.
  • the response value of each pixel is subjected to an extreme point search in the three-dimensional space composed of the above multi-scale layer to determine the position corresponding to the image interest point.
  • the response value corresponding to the Laplacian in the Laplacian filter is sensitive to the spot in the image, and the response value calculated by the Laplacian operator is in the three-dimensional space formed by the multi-scale layer.
  • An extreme point search is performed inside, and a good spot detection effect can be obtained.
  • Multi-scale layers are obtained by filtering the image patches using a Gaussian filter with the same resolution.
  • the image point of interest detecting device acquires an image point of interest of each sample image according to the filtered image of the plurality of image patches in each sampled image.
  • the image point of interest detecting device may acquire image interest points of each image patch according to the filtered image of the plurality of image patches in each sample image, according to image interest of each image patch Point to get the image points of interest for each sampled image.
  • the image point of interest detecting device may combine the filtered images of the plurality of image patches in each sample image to obtain a filtered image of each sample image, and obtain the filtered image according to each sample image. Points of interest for each sampled image.
  • the image point of interest detecting device may also combine the filtered images of the partial image patches in each sample image to obtain a filtered image of the image bulk, and obtain more images in each sample image.
  • the image is a large block of image points of interest, and then the image points of interest of each sample image are obtained according to the image points of interest of each image block in each sample image.
  • the image point of interest detecting device may first acquire one of a plurality of image patches in the sampled image, and sequentially filter the image small block by using a Gaussian Laplacian filter to obtain the image small block. Filtering the image, acquiring an image interest point of the filtered image of the image patch according to the filtered image of the image patch; and then the image interest point detecting device may acquire the next one of the plurality of image patches in the sampled image for processing.
  • the image point of interest detecting device may also perform parallel processing on a plurality of image patches in the sampled image. After obtaining the image interest points of the plurality of image patches in the sampled image, the image point of interest detection device may also acquire the feature points in the sampled image according to the image interest points, and identify the objects in the image according to the feature points.
  • each sample image is divided into a plurality of image patches, and a Gaussian Laplacian filter is used for filtering processing, and an image point of interest is searched for each image block.
  • the filtered image of each image patch is obtained, which reduces the memory resource occupancy and improves the detection speed.
  • FIG. 2 is a flowchart of an embodiment of dividing an image of each sample into a plurality of image patches in an image point of interest detection method provided by the present invention. As shown in FIG. 2, based on the embodiment shown in FIG. Dividing each sample image into a plurality of image patches specifically includes:
  • the image point of interest detecting device divides each sample image into a plurality of square image blocks of length X and width Y, wherein X and Y are positive integers, and if the length of the image block at the edge of the sampled image is smaller than X or the width is less than Y, and the image patches on the edge of the sampled image are pixel-filled.
  • the image point of interest detecting device may sequentially divide from the upper left corner of the sample image, and for a portion whose length is less than X or whose width is smaller than Y, a pixel value of 0 value is added to the edge thereof to make the size become X*Y square pixel. And dividing the obtained plurality of image patches for storage.
  • X and ⁇ are 128 pixels
  • the number of interest points detected by the image point of interest detection device may be more, and the obtained points of interest may have higher performance in image retrieval and matching. Therefore, preferably, both X and ⁇ are 128 pixels.
  • the sampled image can be divided into 5*4—a total of 20 small blocks according to the preceding and following columns, as shown in FIG. 3, wherein the first 15 small blocks can be Directly divided, the last 5 small blocks will be filled with 32 pixels below, and the last 5 small blocks will be filled with 128*128 square pixels.
  • you can press The sampled image is divided according to the manner of the first row and the back row.
  • the image point of interest detecting device performs pixel filling on each square image block of length X and width Y, so that the length of the filled square image block is X+Ml and the width is Y+M-1.
  • M is a positive integer.
  • the image point of interest detecting device may perform pixel filling on each square image block of length X and width Y.
  • M is an odd number
  • the pixels around (M-1) /2 are filled around each other; when M is even, the pixels of each square image of length X and width Y are filled with (M) /2 pixels.
  • the filled pixel value may be a mirror image of a square image block boundary pixel value of length X and width Y; in another case, the filled pixel value may also be 0; in another case, The image point of interest detecting device can determine whether the image patch is located at the position of the sampled image.
  • the side of the image patch located at the edge of the sampled image can be filled with 0, and there is no image in the small block.
  • the other edges at the edge of the sampled image are filled into a mirror image of the image.
  • M may be less than or equal to the value of X, and M is less than or equal to the value of Y to improve the accuracy of the filtering effect of the Gaussian Laplacian filter.
  • each sample image is divided into a plurality of square image blocks of length X and width Y, and each square image block of length X and width Y is performed.
  • Pixel padding and discrete Fourier transform to obtain small blocks of frequency domain images.
  • Gaussian Laplacian filter is used for filtering processing and image interest point searching, and filtering of each image small block is obtained.
  • the image reduces the memory resource usage, and places multiple image patches in the sampled image into the frequency domain for processing, which further improves the detection speed.
  • FIG. 4 is a flowchart of an embodiment of an image point detection method according to the present invention for sequentially filtering a plurality of image patches in each sample image by using a Gaussian Laplacian filter, as shown in FIG. 4, Based on the embodiment shown in FIG. 2, the following specifically includes:
  • the image point of interest detecting device uses a frequency domain Gaussian Laplacian filter to sequentially filter a plurality of image small blocks in each sample image to obtain a frequency domain Gaussian of a plurality of image patches in each sample image. Ras responds to the image.
  • the number of frequency domain Gaussian Laplacian filters is multiple, and the image interest point detecting device may sequentially filter each image small block by using multiple frequency domain Gaussian Laplacian filters to obtain each image.
  • the image patches mentioned here and below specifically refer to small blocks of frequency domain images.
  • the image point of interest detecting device may convert a plurality of frequency domain Gaussian Laplacian response images corresponding to each image patch into a spatial domain, and obtain a plurality of filtered images corresponding to each image patch, and then image interest
  • the point detecting means may perform local Gaussian Laplacian response value comparison in a three-dimensional space composed of a plurality of filtered images corresponding to each image patch according to a plurality of filtered images corresponding to each image patch. As shown in FIG. 5, for each pixel of the plurality of filtered images corresponding to each image patch, the response value is compared with the response value of the surrounding 26 pixel points, if the pixel point is smaller than the surrounding pixel points.
  • the pixel points are used as candidate image points of interest; and the Hessian matrix response values of the positions corresponding to the candidate image points of interest are calculated, and the points whose response values are greater than the preset threshold are regarded as boundaries. Points are removed and the remaining points of interest are image points of interest.
  • the Gaussian Laplacian filter is used to sequentially filter a plurality of image patches in each sample image to obtain a filtered image of a plurality of image patches in each sample image, and the image interest point detecting device
  • Multiple square spatial two-dimensional Gaussian filters and a square spatial two-dimensional Laplacian filter can be generated separately, and pixels of a plurality of square spatial two-dimensional Gaussian filters and one square spatial two-dimensional Laplacian filter are performed.
  • the length and width of the filled spatial two-dimensional Gaussian filter and the spatial two-dimensional Laplacian filter are respectively the same as the length and width of the filled square image patch, thereby making the frequency domain Gaussian filter and
  • the length and width of the frequency domain Laplacian filter are the same as the length and width of the frequency domain image block, respectively, reducing the computational complexity of the image point of interest detection device and increasing the detection speed.
  • the image point of interest detecting device may further generate a plurality of square spatial two-dimensional Gaussian filters according to the two-dimensional Gaussian kernel function and the preset Gaussian parameters, and the maximum width of the plurality of square spatial two-dimensional Gaussian filters is M;
  • a second-order Laplacian operator function generates a square-space two-dimensional Laplacian filter. If the width of the square-space two-dimensional Laplacian filter is less than M, the square-space two-dimensional Laplacian filter is performed. Pixel filling, so that the width of the filled square spatial two-dimensional Laplacian filter is M;
  • each The frequency domain Gaussian filter is used to perform filtering processing on multiple image patches in each sampled image; the filled square spatial two-dimensional Laplacian filter is converted into a frequency domain Laplacian filter.
  • the image blocks are processed by Gaussian filters of different widths, and multiple frequency domain images of different scales corresponding to the image patches are obtained.
  • the image point of interest detecting device is further configured to: for a square spatial two-dimensional Gaussian filter having a width smaller than ⁇ , symmetrically fill pixels around the pixel, and fill the pixel value to 0, so that the filled square spatial two-dimensional Gaussian filter
  • the width is ⁇ , for example, if ⁇ is 7, a square-space two-dimensional Gaussian filter with a width of 7 is filled with a square-space two-dimensional Gaussian filter of width 5 in Table 1, as shown in Table 2.
  • the width of a square spatial two-dimensional Laplacian filter has the same meaning as the width of a square spatial two-dimensional Gaussian filter.
  • the width of the square spatial two-dimensional Laplacian filter may be a pixel value of 3, 5 or 7, and preferably, the width of the square spatial two-dimensional Laplacian filter is 3.
  • the pixel filling method for the square spatial two-dimensional Laplacian filter may be: symmetrically filling the 0-value pixel around the square spatial two-dimensional Laplacian filter.
  • the Gaussian Laplacian filter is used to sequentially filter a plurality of image patches in each sample image to obtain a filtered image of a plurality of image patches in each sample image, and the image interest point detecting device It can also be used for squares with lengths of X+M-l and widths of Y+M-1.
  • the area of the upper left of the matrix is M, and the pixel values of the square spatial two-dimensional Gaussian filter of width M are replaced by one-to-one correspondence, and the square matrix is cyclically shifted to the left and upward by M/2 pixels, wherein, the square All pixel values of the matrix are zero.
  • the purpose of cyclically shifting the square matrix to the left and up by M/2 pixels is to correct the positional shift of the filtered image processed by the Gaussian Laplacian filter. If M is an odd number, the square matrix is cyclically shifted to the left and upward by (M-1)/2 pixels.
  • the image point of interest detecting device may further replace the pixel values of the square space two-dimensional Laplacian filter of width M by one-to-one, and cyclically shift the square matrix to the left and upward.
  • M/2 pixels where all pixel values of the square matrix are zero.
  • the purpose of shifting the square matrix to the left and up by M/2 pixels is to correct the positional shift of the filtered image processed by the Gaussian Laplacian filter. If M is an odd number, rotate the square matrix to the left and up by (M - 1 ) /2 pixels.
  • the image point of interest detecting device may generate a plurality of frequency domain Gaussian Laplacian filters according to a plurality of frequency domain Gaussian filters and a frequency domain Laplacian filter.
  • both the frequency domain Gaussian filter and the frequency domain Laplacian filter are generated offline, and the frequency domain Gaussian Laplacian filter is also generated offline, and can be written into the program code header file corresponding to the method, without online generation.
  • each pixel in the frequency domain image block is a complex pixel composed of a real part and an imaginary part
  • the above-mentioned frequency domain Gaussian Laplacian filter, frequency domain Gaussian filter and Each pixel in the frequency domain Laplacian filter is also a complex pixel composed of a real part and an imaginary part, so the frequency domain Gaussian Laplacian filter is used to filter the frequency domain image small block, actually The upper is the matrix complex point multiplication of the frequency domain image small block and the frequency domain Gaussian Laplacian filter.
  • the image point of interest detecting device performs inverse discrete Fourier transform on the frequency domain Gaussian Laplacian response image of the plurality of image patches in each sample image to obtain a filtered image of the plurality of image patches in each sample image.
  • each sample image is divided into a plurality of square image blocks of length X and width Y, and square image blocks of length X and width Y are respectively Perform pixel filling and discrete Fourier transform to obtain small blocks of frequency domain images.
  • the aforementioned program can be stored in a computer readable storage medium.
  • the program when executed, performs the steps including the above-described method embodiments; and the foregoing storage medium includes: various media that can store program codes, such as ROM, RAM, disk or optical disk.
  • FIG. 7 is a schematic structural diagram of an embodiment of an image point of interest detecting apparatus according to the present invention. As shown in FIG. 7, the method includes:
  • An obtaining module 71 configured to acquire an original input image
  • the processing module 72 is configured to perform a downsampling process on the original input image to obtain a plurality of sample images of different resolutions;
  • a dividing module 73 configured to divide each sample image into a plurality of image patches
  • the filtering module 74 is configured to sequentially filter a plurality of image patches in each sample image by using a Gaussian Laplacian filter to obtain a filtered image of a plurality of image patches in each sample image;
  • the obtaining module 71 is further configured to acquire an image interest point of each sample image according to the filtered image of the plurality of image patches in each sample image.
  • the obtaining module 71 may obtain image interest points of each image patch according to the filtered images of the plurality of image patches in each sample image, according to image interest points of each image patch. Get the image points of interest for each sampled image.
  • the obtaining module 71 may combine the filtered images of the plurality of image patches in each sample image to obtain a filtered image of each sample image, and acquire each image according to the filtered image of each sample image. The point of interest of the sampled image.
  • the obtaining module 71 may also combine the filtered images of the partial image patches in each sample image to obtain a filtered image of the image bulk, and obtain multiple images in each sample image. A large block of image points of interest, and then image points of interest for each sampled image are obtained based on image points of interest for each image in each sampled image.
  • the image point of interest detecting device may first acquire one of a plurality of image patches in the sampled image, and sequentially filter the image small block by using a Gaussian Laplacian filter to obtain the image small block. Filtering the image, acquiring an image interest point of the filtered image of the image patch according to the filtered image of the image patch; and then the image interest point detecting device may acquire the next one of the plurality of image patches in the sampled image for processing.
  • the image point of interest detecting device may also include a plurality of filtering modules 74 and a plurality of obtaining modules 71.
  • One filtering module 74 and one obtaining module 71 may be configured to process one of the plurality of image patches in the sampled image, and the plurality of filtering modules 74. And the plurality of acquisition modules 71 can be respectively used to process a plurality of image patches in the sampled image, so that the image point of interest detection device can perform parallel processing on the plurality of image patches in the sampled image to further improve the detection speed.
  • each sample image is divided into a plurality of image patches, and a Gaussian Laplacian filter is used for filtering processing, and an image point of interest is searched for each image block.
  • the filtered image of each image patch is obtained, which reduces the memory resource occupancy and improves the detection speed.
  • FIG. 8 is a schematic structural diagram of another embodiment of an image point of interest detecting apparatus according to the present invention. As shown in FIG. 8, based on the embodiment shown in FIG. 7, the method further includes: a filling module 75 and a converting module 76;
  • the dividing module 73 is specifically configured to divide each sample image into a plurality of lengths of X and a width of
  • a square image block of Y wherein X and Y are positive integers. If the length of the image patch at the edge of the sampled image is less than X or the width is less than Y, the filling module 75 performs pixel filling on the image patch of the edge of the sampled image;
  • the filling module 75 is further configured to perform pixel filling on each square image block of length X and width Y, so that the length of the filled square image patch is X+Ml and the width is Y+M-1.
  • M is a positive integer;
  • the conversion module 76 is configured to perform discrete Fourier transform on the filled square image block to obtain a frequency domain image small block.
  • each sample image is divided into a plurality of square image patches of length X and width Y, and squares each having a length of X and a width of Y.
  • Image blocks are filled with pixels and discrete Fourier transforms to obtain small blocks of frequency domain images.
  • Gaussian Laplacian filter is used for filtering processing, and image interest points are searched, and each is obtained.
  • the filtered image of the image block reduces the memory resource occupancy, and places multiple image blocks in the sampled image into the frequency domain for processing, which further improves the detection speed.
  • FIG. 9 is a schematic structural diagram of another embodiment of an image point of interest detecting apparatus according to the present invention. As shown in FIG. 9, on the basis of the embodiment shown in FIG. 8, the image point of interest detecting apparatus further includes: a generating module 77;
  • the filtering module 74 uses a frequency domain Gaussian Laplacian filter to sequentially filter a plurality of image patches in each sample image to obtain a filtered image of a plurality of image patches in each sample image, and then generate a module 77. Generating a plurality of square spatial two-dimensional Gaussian filters according to a two-dimensional Gaussian kernel function and a preset Gaussian parameter, wherein a maximum width of the plurality of square spatial two-dimensional Gaussian filters is M;
  • the generating module 77 is further configured to generate a square spatial two-dimensional Laplacian filter according to the second-order Laplacian function, if the width of the square spatial two-dimensional Laplacian filter is less than M, and the square airspace is two
  • the Verapella filter performs pixel filling so that the width of the filled square spatial two-dimensional Laplacian filter is M;
  • the conversion module 76 is further configured to convert the plurality of square spatial two-dimensional Gaussian filters into a plurality of frequency domain Gauss filters; convert the filled square spatial two-dimensional Laplacian filter into frequency domain Laplacian filtering Device
  • a plurality of frequency domain Gaussian Laplacian filters are generated according to a plurality of frequency domain Gaussian filters and a frequency domain Laplacian filter.
  • the filtering module 74 is specifically configured to sequentially filter a plurality of image patches in each sample image by using a frequency domain Gaussian Laplacian filter to obtain a plurality of image patches in each sample image.
  • Frequency domain Gaussian Laplacian response image
  • the conversion module 76 is further configured to perform inverse discrete Fourier transform on the frequency domain Gaussian Laplacian response image of the plurality of image patches in each sample image to obtain a filtered image of the plurality of image patches in each sample image.
  • the filling module 75 is further used for the length X+M-l, width
  • the area of the width M of the upper left of the square matrix of Y+Ml is replaced by the pixel values of the square spatial two-dimensional Gaussian filter of width M, and the square matrix is cyclically shifted to the left and upward by M/2. Pixels, wherein all pixel values of the square matrix are 0;
  • the filling module 75 is further used for a length of X+Ml and a width of Y+M-1.
  • the area of the upper left of the square matrix is M, and the pixel values of the square space two-dimensional Laplacian filter of width M are used for one-to-one correspondence replacement, and the square matrix is cyclically shifted to the left and upward by M/2. Pixels, where all pixel values of the square matrix are zero.
  • each sample image is divided into a plurality of square image patches of length X and width Y, and square image patches of length X and width Y are respectively Pixel filling and discrete Fourier transform are performed to obtain small frequency domain image blocks.
  • the frequency domain Gaussian Laplacian filter is used to sequentially filter multiple image small blocks in each sample image to obtain multiple images in each sample image.
  • the frequency domain Gaussian Laplacian response image of the image block is subjected to discrete Fourier inverse transform of the frequency domain Gaussian Laplacian response image of the plurality of image patches in each sample image to obtain each sample image. Filtering images of multiple image patches, performing image point of interest search on the filtered images of multiple image patches in each sample image, reducing memory resource occupancy, and placing multiple image patches in the sample image into the frequency domain Processing in the middle, further improving the detection speed.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Multimedia (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biodiversity & Conservation Biology (AREA)
  • Biomedical Technology (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Geometry (AREA)
  • Image Processing (AREA)
  • Image Analysis (AREA)

Abstract

本发明提供一种图像兴趣点检测方法和装置。其中,方法包括:获取原始输入图像;对原始输入图像进行降采样处理,得到多个不同分辨率的采样图像;将每个采样图像划分为多个图像小块;采用高斯拉普拉斯滤波器依次对每个采样图像中多个图像小块进行滤波处理,得到每个采样图像中多个图像小块的滤波图像;获取每个采样图像中多个图像小块的滤波图像的图像兴趣点。用于解决现有技术中内存资源占用较多,检测速度低的问题。

Description

图像兴趣点检测方法和装置
技术领域
本发明涉及图像处理技术, 尤其涉及一种图像兴趣点检测方法和装 置。 背景技术
移动视觉搜索是采用计算机视觉的方法对图像中的物体进行识别,将 图像中待识别的物体与相关的虚拟数字信息进行关联,使用户不需要输入 任何关键字, 就能检索到与之相关的各种资讯, 是一种非常人性化的信息 检索方式。 移动视觉搜索的关键技术之一是图像兴趣点检测技术, 图像兴 趣点检测技术指的是在图像中找到一定数量的标志,这些标志具有的特点 是:在包含同一物体的另一幅图像中也能在该物体的同一位置上检测得到 这些标志, 即便是两幅图像具有不同的尺寸或拍摄于不同的光照条件下或 角度下。
现有技术中, 图像兴趣点检测装置对原始图像进行降采样, 生成包括 一系列分辨率逐歩降低的图像的金字塔; 使用多个空域高斯滤波器获取图 像金字塔中每一层图像的多尺度图像; 使用空域拉普拉斯滤波器获取每一 层图像的多尺度高斯拉普拉斯响应图像,根据多尺度高斯拉普拉斯响应图 像, 获取每一层图像的兴趣点。
然而现有技术中,使用高斯滤波器获取图像金字塔中每一层图像的多 尺度图像后, 需要缓存图像金字塔中每一层图像的多尺度图像, 内存资源 占用较多, 导致检测速度低。 发明内容
本发明提供一种图像兴趣点检测方法和装置, 用于解决现有技术中内 存资源占用较多, 导致检测速度低的问题。
本发明的第一个方面是提供一种图像兴趣点检测方法, 包括: 获取原始输入图像; 对所述原始输入图像进行降采样处理, 得到多个不同分辨率的采样图 像;
将每个所述采样图像划分为多个图像小块;
采用高斯拉普拉斯滤波器依次对每个所述采样图像中多个图像小块 进行滤波处理, 得到每个所述采样图像中多个图像小块的滤波图像; 根据每个所述采样图像中多个图像小块的滤波图像, 获取每个所述采 样图像的图像兴趣点。
本发明的另一个方面提供一种图像兴趣点检测装置, 包括: 获取模块, 用于获取原始输入图像;
处理模块, 用于对所述原始输入图像进行降采样处理, 得到多个不同 分辨率的采样图像;
划分模块, 用于将每个所述采样图像划分为多个图像小块; 滤波模块, 用于采用高斯拉普拉斯滤波器依次对每个所述采样图像中 多个图像小块进行滤波处理, 得到每个所述采样图像中多个图像小块的滤 波图像;
所述获取模块, 还用于根据每个所述采样图像中多个图像小块的滤波 图像, 获取每个所述采样图像的图像兴趣点。
本发明通过将每个采样图像划分为多个图像小块, 针对每个图像小块 分别采用高斯拉普拉斯滤波器进行滤波处理, 以及图像兴趣点查找, 得到 每个图像小块的滤波图像, 减少了内存资源占用量, 提高了检测速度。 附图说明 图 1为本发明提供的图像兴趣点检测方法一个实施例的流程图; 图 2为本发明提供的图像兴趣点检测方法中将每个采样图像划分为多 个图像小块的一个实施例的流程图;
图 3为将采样图像划分为多个 128* 128平方像素的方形图像小块的示 意图;
图 4为本发明提供的图像兴趣点检测方法中采用高斯拉普拉斯滤波器 依次对每个采样图像中多个图像小块进行滤波处理的一个实施例的流程 图; 图 5为图像小块对应的多个滤波图像组成的三维空间中进行图像兴趣 点查找的示意图;
图 6为 σ =1时正方形空域二维高斯滤波器的示意图;
图 7为本发明提供的图像兴趣点检测装置一个实施例的结构示意图; 图 8为本发明提供的图像兴趣点检测装置又一个实施例的结构示意 图;
图 9为本发明提供的图像兴趣点检测装置另一个实施例的结构示意 图。 具体实施方式
为使本发明实施例的目的、 技术方案和优点更加清楚, 下面将结合本 发明实施例中的附图, 对本发明实施例中的技术方案进行清楚、 完整地描 述, 显然, 所描述的实施例是本发明一部分实施例, 而不是全部的实施例。 基于本发明中的实施例, 本领域普通技术人员在没有做出创造性劳动前提 下所获得的所有其他实施例, 都属于本发明保护的范围。
图 1为本发明提供的图像兴趣点检测方法一个实施例的流程图, 如图 1所示, 本发明涉及的图像兴趣点检测方法的执行主体可以为照相机、 摄 像机、 摄像头、 电脑或者手机等终端设备中具有处理功能的模块, 例如: CPU等。 该方法包括:
101、 图像兴趣点检测装置获取原始输入图像。
原始输入图像可以为照相机、 摄像机、 摄像头、 电脑或者手机等可以 进行拍照的设备获得的图像。
102、 图像兴趣点检测装置对原始输入图像进行降采样处理, 得到多 个不同分辨率的采样图像。
降采样处理指的是, 图像兴趣点检测装置对原始输入图像进行滤波, 然后进行歩长为 2的采样, 生成分辨率逐渐降低的多级采样图像。 采样前 采用的滤波方式一般是高斯低通滤波或领域均值滤波。 图像兴趣点检测装 置可以对每个分辨率的采样图像进行图像兴趣点检测, 然后将检测到的所 有采样图像的图像兴趣点集合到一起, 作为原始输入图像的兴趣点。
103、 图像兴趣点检测装置将每个采样图像划分为多个图像小块。 其中, 图像兴趣点检测装置对每个采样图像划分得到的多个图像小块 可以为正方形, 也可以为长方形。
104、 图像兴趣点检测装置采用高斯拉普拉斯滤波器依次对每个采样 图像中多个图像小块进行滤波处理, 得到每个采样图像中多个图像小块的 滤波图像。
其中, 高斯拉普拉斯滤波器是根据高斯滤波器和拉普拉斯滤波器生成 得到的, 图像兴趣点检测装置可以采用高斯拉普拉斯滤波器对每个采样图 像中多个图像小块进行滤波处理, 得到每个采样图像中多个图像小块的滤 波图像; 图像兴趣点检测装置还可以将高斯拉普拉斯滤波器拆分成高斯滤 波器和拉普拉斯滤波器, 采用高斯滤波器和拉普拉斯滤波器依次对每个采 样图像中多个图像小块进行滤波处理。采用高斯滤波器对每个采样图像中 多个图像小块进行滤波处理的作用是, 构建采样图像中每个图像小块的多 尺度图层。 采用拉普拉斯滤波器的作用是, 在每个图像小块的多尺度图层 中计算每个像素点的响应值, 以使图像兴趣点检测装置根据每个图像小块 的多尺度图层中每个像素点的响应值在上述多尺度图层所组成的三维空 间内进行极值点搜索, 以确定图像兴趣点对应的位置。 其中, 拉普拉斯滤 波器中拉普拉斯算子对应的响应值对于图像中的斑点比较敏感, 利用拉普 拉斯算子计算所得的响应值在上述多尺度图层所组成的三维空间内进行 极值点搜索, 能够获得较好的斑点检测效果。 多尺度图层是通过采用高斯 滤波器对图像小块滤波得到的, 具有相同的分辨率。
105、 图像兴趣点检测装置根据每个采样图像中多个图像小块的滤波 图像, 获取每个采样图像的图像兴趣点。
其中, 在一种实施场景下, 图像兴趣点检测装置可以根据每个采样图 像中多个图像小块的滤波图像, 获取每个图像小块的图像兴趣点, 根据每 个图像小块的图像兴趣点, 获取每个采样图像的图像兴趣点。
在另一种实施场景下, 图像兴趣点检测装置可以将每个采样图像中多 个图像小块的滤波图像进行合并, 得到每个采样图像的滤波图像, 根据每 个采样图像的滤波图像, 获取每个采样图像的兴趣点。
另外, 图像兴趣点检测装置也可以将每个采样图像中的部分图像小块 的滤波图像进行合并, 得到图像大块的滤波图像, 获取每个采样图像中多 个图像大块的图像兴趣点, 然后根据每个采样图像中每个图像大块的图像 兴趣点, 获取每个采样图像的图像兴趣点。
需要进行说明的是, 图像兴趣点检测装置可以先获取采样图像中多个 图像小块中的一个, 采用高斯拉普拉斯滤波器依次对该图像小块进行滤波 处理, 得到该图像小块的滤波图像, 根据该图像小块的滤波图像, 获取该 图像小块的滤波图像的图像兴趣点; 然后图像兴趣点检测装置可以获取采 样图像中多个图像小块中的下一个进行处理。 另外, 图像兴趣点检测装置 也可以对采样图像中多个图像小块进行并行处理。 图像兴趣点检测装置获 取采样图像中多个图像小块的所有图像兴趣点后, 还可以根据图像兴趣点 获取采样图像中的特征点, 根据特征点, 对图像中的物体进行识别。
本实施例提供的图像兴趣点检测方法中, 将每个采样图像划分为多个 图像小块, 针对每个图像小块分别采用高斯拉普拉斯滤波器进行滤波处 理, 以及图像兴趣点查找, 得到每个图像小块的滤波图像, 减少了内存资 源占用量, 提高了检测速度。
图 2为本发明提供的图像兴趣点检测方法中将每个采样图像划分为多 个图像小块的一个实施例的流程图, 如图 2所示, 在图 1所示实施例的基 础上, 将每个采样图像划分为多个图像小块具体包括:
1031、 图像兴趣点检测装置将每个采样图像划分为多个长度为 X、 宽 度为 Y的方形图像小块, 其中, X、 Y均为正整数, 若采样图像边缘的图像 小块的长度小于 X或者宽度小于 Y , 则对采样图像边缘的图像小块进行像 素填充。
图像兴趣点检测装置可以从采样图像的左上角开始依次进行划分, 对 于长度小于 X或者宽度小于 Y的部分, 在其边缘补上 0值的像素值, 使其 大小变成 X*Y平方像素, 并划分得到的多个图像小块进行存储。 当 X和 Υ 都为 128像素时, 图像兴趣点检测装置检测得到的图像兴趣点的数目可能 较多, 得到的兴趣点在图像检索和匹配中的性能可能较高。 因此, 优选的, X和 Υ均为 128像素。例如, 对于分辨率为 640*480平方像素的采样图像, 可以按照先行后列的方式将该采样图像分成 5*4—共 20个小块, 如图 3 所示, 其中前 15个小块可以直接划分得到, 后 5个小块将在其下方补上 32像素, 将后 5个小块的大小填充成 128*128平方像素。 另外, 也可以按 照先列后行的方式对采样图像进行划分。
1032、 图像兴趣点检测装置对每个长度为 X、 宽度为 Y的方形图像小 块进行像素填充, 以使填充后的方形图像小块的长度为 X+M-l、 宽度为 Y+M-1 , M为正整数。
图像兴趣点检测装置对每个长度为 X、 宽度为 Y的方形图像小块进行 像素填充的方式可以为, 当 M为奇数时, 对每个长度为 X、 宽度为 Y的方 形图像小块的四周分别填充 (M- 1 ) /2的像素; 当 M为偶数时, 对每个长 度为 X、 宽度为 Y的方形图像小块的四周分别填充 (M) /2的像素。 一种 情况下, 填充的像素值可以为长度为 X、 宽度为 Y的方形图像小块边界像 素值的镜像; 另一种情况下, 填充的像素值还可以为 0 ; 再一种情况下, 图像兴趣点检测装置可以判断图像小块是否位于采样图像的位置, 若图像 小块位于采样图像的边缘, 则可以将图像小块中位于采样图像边缘的一边 填充为 0, 将图像小块中没有位于采样图像边缘的其他边填充成图像小块 的镜像。 另外, 优选的, M可以小于等于 X的值, 且 M小于等于 Y的值, 以提高高斯拉普拉斯滤波器的滤波效果的准确度。
1033、 对填充后的方形图像小块进行离散傅里叶变换, 得到频域图像 小块。
本实施例提供的图像兴趣点检测方法中, 将每个采样图像划分为多个 长度为 X、 宽度为 Y的方形图像小块, 对每个长度为 X、 宽度为 Y的方形 图像小块进行像素填充以及离散傅里叶变换, 得到频域图像小块, 针对每 个频域图像小块分别采用高斯拉普拉斯滤波器进行滤波处理以及图像兴 趣点查找, 得到每个图像小块的滤波图像, 减少了内存资源占用量, 而将 采样图像中多个图像小块放置到频域中进行处理, 进一歩提高了检测速 度。
图 4为本发明提供的图像兴趣点检测方法中采用高斯拉普拉斯滤波器 依次对每个采样图像中多个图像小块进行滤波处理的一个实施例的流程 图, 如图 4所示, 在图 2所示实施例的基础上, 具体包括:
1041、 图像兴趣点检测装置采用频域高斯拉普拉斯滤波器依次对每个 采样图像中多个图像小块进行滤波处理, 得到每个采样图像中多个图像小 块的频域高斯拉普拉斯响应图像。 其中, 频域高斯拉普拉斯滤波器的数量为多个, 图像兴趣点检测装置 可以采用多个频域高斯拉普拉斯滤波器依次对每个图像小块进行滤波处 理, 得到每个图像小块对应的多个滤波图像。 此处以及下面提到的图像小 块, 具体指的是频域图像小块。
进一歩地, 图像兴趣点检测装置可以将每个图像小块对应的多个频域 高斯拉普拉斯响应图像, 转换成空域, 得到每个图像小块对应的多个滤波 图像, 然后图像兴趣点检测装置可以根据每个图像小块对应的多个滤波图 像, 在每个图像小块对应的多个滤波图像组成的三维空间中进行局部的高 斯拉普拉斯响应值比较。 如图 5所示, 对每个图像小块对应的多个滤波图 像中的每个像素点, 将其响应值与周围 26个像素点的响应值进行比较, 如果该像素点比周围像素点的响应值都大或者都小, 那么将该像素点作为 候选的图像兴趣点; 并计算这些候选的图像兴趣点对应位置的海森矩阵响 应值, 响应值大于预设阈值的点将被视为边界点并被剔除, 剩下的兴趣点 为图像兴趣点。
更进一歩地, 采用高斯拉普拉斯滤波器依次对每个采样图像中多个图 像小块进行滤波处理, 得到每个采样图像中多个图像小块的滤波图像之 前, 图像兴趣点检测装置可以分别生成多个正方形空域二维高斯滤波器以 及一个正方形空域二维拉普拉斯滤波器, 并对多个正方形空域二维高斯滤 波器以及一个正方形空域二维拉普拉斯滤波器进行像素填充, 以使填充后 的空域二维高斯滤波器和空域二维拉普拉斯滤波器的长度和宽度分别与 填充后的方形图像小块的长度和宽度相同, 从而使得频域高斯滤波器和频 域拉普拉斯滤波器的长度和宽度分别与频域图像小块的长度和宽度相同, 减少图像兴趣点检测装置的运算量, 提高检测速度。 例如, 图像兴趣点检 测装置还可以根据二维高斯核函数和预设的高斯参数, 生成多个正方形空 域二维高斯滤波器, 多个正方形空域二维高斯滤波器中的最大宽度为 M; 根据二阶拉普拉斯算子函数, 生成正方形空域二维拉普拉斯滤波器, 若正方形空域二维拉普拉斯滤波器的宽度小于 M, 对正方形空域二维拉普 拉斯滤波器进行像素填充, 以使填充后的正方形空域二维拉普拉斯滤波器 的宽度为 M;
将多个正方形空域二维高斯滤波器转换为多个频域高斯滤波器, 每个 频域高斯滤波器用于对每个采样图像中多个图像小块进行一次滤波处理; 将填充后的正方形空域二维拉普拉斯滤波器转换为频域拉普拉斯滤波器。
其中, σ = 1时正方形空域二维高斯滤波器的示意图, 如图 6所示, 对 应的坐标点的值如表 1所示, 正方形空域二维高斯滤波器的宽度指的是, 正方形空域二维高斯滤波器的示意图中取值不为零的区域的宽度。 如表 1 所示, 可知 σ = 1时, 正方形空域二维高斯滤波器的宽度为 5。 采用不同宽 度的高斯滤波器对图像小块进行处理, 会得到图像小块对应的多个尺度不 同的频域图像。
表 1
1/273
Figure imgf000010_0001
图像兴趣点检测装置还用于, 对宽度小于 Μ的正方形空域二维高斯滤 波器, 将其周围对称地进行像素填充, 填充的像素值为 0, 以使填充后的 正方形空域二维高斯滤波器的宽度为^ 例如, 若 Μ为 7, 对表 1中宽度 为 5的正方形空域二维高斯滤波器填充得到的宽度为 7的正方形空域二维 高斯滤波器, 如表 2所示。
表 2
1/273
Figure imgf000010_0002
正方形空域二维拉普拉斯滤波器的宽度含义与正方形空域二维高斯 滤波器的宽度含义类似。正方形空域二维拉普拉斯滤波器的宽度可以为 3、 5或 7等像素值, 优选的, 正方形空域二维拉普拉斯滤波器的宽度为 3。 对正方形空域二维拉普拉斯滤波器进行像素填充的方式可以为: 在正方形 空域二维拉普拉斯滤波器的周围对称地填充 0值像素点。
更进一歩地, 采用高斯拉普拉斯滤波器依次对每个采样图像中多个图 像小块进行滤波处理, 得到每个采样图像中多个图像小块的滤波图像之 前, 图像兴趣点检测装置还可以对于长度为 X+M- l、 宽度为 Y+M-1的方形 矩阵左上方的宽度为 M的区域, 使用宽度为 M的正方形空域二维高斯滤波 器的像素值进行一一对应替换, 将方形矩阵向左和向上循环移位 M/2个像 素, 其中, 方形矩阵的所有像素值为 0。 将方形矩阵向左和向上循环移位 M/2个像素的目的是, 对经过高斯拉普拉斯滤波器处理得到的滤波图像的 位置偏移进行校正。 若 M为奇数, 将方形矩阵向左和向上循环移位 (M-1 ) /2个像素。
再进一歩地, 将填充后的正方形空域二维拉普拉斯滤波器转换为频域 拉普拉斯滤波器之前, 对于长度为 X+M- l、 宽度为 Y+M- 1的方形矩阵左上 方的宽度为 M的区域, 图像兴趣点检测装置还可以使用宽度为 M的正方形 空域二维拉普拉斯滤波器的像素值进行一一对应替换, 将方形矩阵向左和 向上循环移位 M/2个像素, 其中, 方形矩阵的所有像素值为 0。 将方形矩 阵向左和向上循环移位 M/2个像素的目的是, 对经过高斯拉普拉斯滤波器 处理得到的滤波图像的位置偏移进行校正。 若 M为奇数, 将方形矩阵向左 和向上循环移位 (M- 1 ) /2个像素。
再进一歩地, 图像兴趣点检测装置可以根据多个频域高斯滤波器和频 域拉普拉斯滤波器, 生成多个频域高斯拉普拉斯滤波器。
另外, 频域高斯滤波器和频域拉普拉斯滤波器均为离线生成, 频域高 斯拉普拉斯滤波器也为离线生成, 可以写入方法对应的程序代码头文件 中, 无需在线生成。 需要说明的是, 上述频域图像小块中的每个像素点是 由实部和虚部构成的复数像素点, 上述提到的频域高斯拉普拉斯滤波器、 频域高斯滤波器和频域拉普拉斯滤波器中的每个像素点也是由实部和虚 部构成的复数像素点, 因此采用频域高斯拉普拉斯滤波器对频域图像小块 进行滤波的操作, 实际上是频域图像小块与频域高斯拉普拉斯滤波器的矩 阵复数点乘。
1042、 图像兴趣点检测装置对每个采样图像中多个图像小块的频域高 斯拉普拉斯响应图像进行离散傅里叶反变换, 得到每个采样图像中多个图 像小块的滤波图像。
本发明实施例提供的图像兴趣点检测方法中, 将每个采样图像划分为 多个长度为 X、 宽度为 Y的方形图像小块, 对每个长度为 X、 宽度为 Y的 方形图像小块进行像素填充以及离散傅里叶变换, 得到频域图像小块, 采 用频域高斯拉普拉斯滤波器对每个采样图像中多个图像小块进行滤波处 理, 得到每个采样图像中多个图像小块的频域高斯拉普拉斯响应图像, 对 每个采样图像中多个图像小块的频域高斯拉普拉斯响应图像进行离散傅 里叶反变换, 得到每个采样图像中多个图像小块的滤波图像, 对每个采样 图像中多个图像小块的滤波图像进行图像兴趣点查找, 而将采样图像中多 个图像小块放置到频域中进行处理, 进一歩提高了检测速度。
本领域普通技术人员可以理解: 实现上述各方法实施例的全部或部分 歩骤可以通过程序指令相关的硬件来完成。前述的程序可以存储于一计算 机可读取存储介质中。 该程序在执行时, 执行包括上述各方法实施例的歩 骤; 而前述的存储介质包括: R0M、 RAM, 磁碟或者光盘等各种可以存储程 序代码的介质。
图 7为本发明提供的图像兴趣点检测装置一个实施例的结构示意图, 如图 7所示, 包括:
获取模块 71, 用于获取原始输入图像;
处理模块 72, 用于对原始输入图像进行降采样处理, 得到多个不同分 辨率的采样图像;
划分模块 73, 用于将每个采样图像划分为多个图像小块;
滤波模块 74,用于采用高斯拉普拉斯滤波器依次对每个采样图像中多 个图像小块进行滤波处理, 得到每个采样图像中多个图像小块的滤波图 像;
获取模块 71, 还用于根据每个采样图像中多个图像小块的滤波图像, 获取每个采样图像的图像兴趣点。
其中, 在一种实施场景下, 获取模块 71可以根据每个采样图像中多 个图像小块的滤波图像, 获取每个图像小块的图像兴趣点, 根据每个图像 小块的图像兴趣点, 获取每个采样图像的图像兴趣点。
在另一种实施场景下, 获取模块 71可以将每个采样图像中多个图像 小块的滤波图像进行合并, 得到每个采样图像的滤波图像, 根据每个采样 图像的滤波图像, 获取每个采样图像的兴趣点。
另外, 获取模块 71也可以将每个采样图像中的部分图像小块的滤波 图像进行合并, 得到图像大块的滤波图像, 获取每个采样图像中多个图像 大块的图像兴趣点, 然后根据每个采样图像中每个图像大块的图像兴趣 点, 获取每个采样图像的图像兴趣点。
需要进行说明的是, 图像兴趣点检测装置可以先获取采样图像中多个 图像小块中的一个, 采用高斯拉普拉斯滤波器依次对该图像小块进行滤波 处理, 得到该图像小块的滤波图像, 根据该图像小块的滤波图像, 获取该 图像小块的滤波图像的图像兴趣点; 然后图像兴趣点检测装置可以获取采 样图像中多个图像小块中的下一个进行处理。 图像兴趣点检测装置也可以 包括多个滤波模块 74和多个获取模块 71,一个滤波模块 74和一个获取模 块 71可以用于处理采样图像中多个图像小块中的一个, 多个滤波模块 74 和多个获取模块 71可以分别用于处理采样图像中的多个图像小块, 从而 图像兴趣点检测装置可以对采样图像中的多个图像小块进行并行处理, 进 一歩提高检测速度。
本实施例提供的图像兴趣点检测装置中, 将每个采样图像划分为多个 图像小块, 针对每个图像小块分别采用高斯拉普拉斯滤波器进行滤波处 理, 以及图像兴趣点查找, 得到每个图像小块的滤波图像, 减少了内存资 源占用量, 提高了检测速度。
图 8为本发明提供的图像兴趣点检测装置又一个实施例的结构示意 图, 如图 8所示, 在图 7所示实施例的基础上, 还包括: 填充模块 75和 转换模块 76 ;
划分模块 73, 具体用于将每个采样图像划分为多个长度为 X、 宽度为
Y的方形图像小块, 其中, X、 Y均为正整数, 若采样图像边缘的图像小块 的长度小于 X或者宽度小于 Y ,则填充模块 75对采样图像边缘的图像小块 进行像素填充;
填充模块 75, 还用于对每个长度为 X、 宽度为 Y的方形图像小块进行 像素填充, 以使填充后的方形图像小块的长度为 X+M-l、 宽度为 Y+M- 1 , M 为正整数;
转换模块 76, 用于对填充后的方形图像小块进行离散傅里叶变换, 得 到频域图像小块。
本实施例提供的图像兴趣点检测装置中, 将每个采样图像划分为多个 长度为 X、 宽度为 Y的方形图像小块, 对每个长度为 X、 宽度为 Y的方形 图像小块进行像素填充以及离散傅里叶变换, 得到频域图像小块, 针对每 个频域图像小块分别采用高斯拉普拉斯滤波器进行滤波处理, 以及图像兴 趣点查找, 得到每个图像小块的滤波图像, 减少了内存资源占用量, 而将 采样图像中多个图像小块放置到频域中进行处理, 进一歩提高了检测速 度。
图 9为本发明提供的图像兴趣点检测装置另一个实施例的结构示意 图, 如图 9所示, 在图 8所示实施例的基础上, 图像兴趣点检测装置还包 括: 生成模块 77 ;
滤波模块 74采用频域高斯拉普拉斯滤波器依次对每个采样图像中多 个图像小块进行滤波处理, 得到每个采样图像中多个图像小块的滤波图像 之前, 生成模块 77, 用于根据二维高斯核函数和预设的高斯参数, 生成多 个正方形空域二维高斯滤波器, 多个正方形空域二维高斯滤波器中的最大 宽度为 M;
生成模块 77还用于, 根据二阶拉普拉斯算子函数, 生成正方形空域 二维拉普拉斯滤波器, 若正方形空域二维拉普拉斯滤波器的宽度小于 M, 对正方形空域二维拉普拉斯滤波器进行像素填充, 以使填充后的正方形空 域二维拉普拉斯滤波器的宽度为 M;
转换模块 76,还用于将多个正方形空域二维高斯滤波器转换为多个频 域高斯滤波器; 将填充后的正方形空域二维拉普拉斯滤波器转换为频域拉 普拉斯滤波器;
根据多个频域高斯滤波器和频域拉普拉斯滤波器, 生成多个频域高斯 拉普拉斯滤波器。
进一歩地, 滤波模块 74, 具体用于, 采用频域高斯拉普拉斯滤波器依 次对每个采样图像中多个图像小块进行滤波处理, 得到每个采样图像中多 个图像小块的频域高斯拉普拉斯响应图像;
转换模块 76,还用于对每个采样图像中多个图像小块的频域高斯拉普 拉斯响应图像进行离散傅里叶反变换, 得到每个采样图像中多个图像小块 的滤波图像。
更进一歩地, 转换模块 76将多个正方形空域二维高斯滤波器转换为 多个频域高斯滤波器之前, 填充模块 75还用于, 对于长度为 X+M- l、 宽度 为 Y+M-l的方形矩阵左上方的宽度为 M的区域, 使用宽度为 M的正方形空 域二维高斯滤波器的像素值进行一一对应替换, 将方形矩阵向左和向上循 环移位 M/2个像素, 其中, 方形矩阵的所有像素值为 0 ;
转换模块 76将填充后的正方形空域二维拉普拉斯滤波器转换为频域 拉普拉斯滤波器之前, 填充模块 75还用于, 对于长度为 X+M-l、 宽度为 Y+M-1的方形矩阵左上方的宽度为 M的区域, 使用宽度为 M的正方形空域 二维拉普拉斯滤波器的像素值进行一一对应替换, 将方形矩阵向左和向上 循环移位 M/2个像素, 其中, 方形矩阵的所有像素值为 0。
本发明实施例提供的图像兴趣点检测装置中, 将每个采样图像划分为 多个长度为 X、 宽度为 Y的方形图像小块, 对每个长度为 X、 宽度为 Y的 方形图像小块进行像素填充以及离散傅里叶变换, 得到频域图像小块, 采 用频域高斯拉普拉斯滤波器依次对每个采样图像中多个图像小块进行滤 波处理, 得到每个采样图像中多个图像小块的频域高斯拉普拉斯响应图 像, 对每个采样图像中多个图像小块的频域高斯拉普拉斯响应图像进行离 散傅里叶反变换, 得到每个采样图像中多个图像小块的滤波图像, 对每个 采样图像中多个图像小块的滤波图像进行图像兴趣点查找, 减少了内存资 源占用量, 而将采样图像中多个图像小块放置到频域中进行处理, 进一歩 提高了检测速度。
最后应说明的是: 以上各实施例仅用以说明本发明的技术方案, 而非对 其限制; 尽管参照前述各实施例对本发明进行了详细的说明, 本领域的普通 技术人员应当理解:其依然可以对前述各实施例所记载的技术方案进行修改, 或者对其中部分或者全部技术特征进行等同替换; 而这些修改或者替换, 并 不使相应技术方案的本质脱离本发明各实施例技术方案的范围。

Claims

权 利 要 求 书
1、 一种图像兴趣点检测方法, 其特征在于, 包括:
获取原始输入图像;
对所述原始输入图像进行降采样处理, 得到多个不同分辨率的采样图 像;
将每个所述采样图像划分为多个图像小块;
采用高斯拉普拉斯滤波器依次对每个所述采样图像中多个图像小块 进行滤波处理, 得到每个所述采样图像中多个图像小块的滤波图像;
根据每个所述采样图像中多个图像小块的滤波图像, 获取每个所述采 样图像的图像兴趣点。
2、 根据权利要求 1所述的方法, 其特征在于, 所述将每个所述采样 图像划分为多个图像小块, 包括:
将每个所述采样图像划分为多个长度为 X、宽度为 Y的方形图像小块, 其中, X、 Y均为正整数, 若所述采样图像边缘的图像小块的长度小于 X 或者宽度小于 Y , 则对所述采样图像边缘的图像小块进行像素填充;
对每个长度为 X、 宽度为 Y的方形图像小块进行像素填充, 以使填充 后的方形图像小块的长度为 X+M-l、 宽度为 Y+M- 1 , M为正整数;
对填充后的方形图像小块进行离散傅里叶变换, 得到频域图像小块。
3、 根据权利要求 2所述的方法, 其特征在于, 所述采用高斯拉普拉 斯滤波器依次对每个所述采样图像中多个图像小块进行滤波处理, 得到每 个所述采样图像中多个图像小块的滤波图像, 包括:
采用频域高斯拉普拉斯滤波器对每个所述采样图像中多个图像小块 进行多次滤波处理, 得到每个所述采样图像中多个图像小块的频域高斯拉 普拉斯响应图像;
对每个所述采样图像中多个图像小块的频域高斯拉普拉斯响应图像 进行离散傅里叶反变换, 得到每个所述采样图像中多个图像小块的滤波图 像。
4、 根据权利要求 3所述的方法, 其特征在于, 所述采用高斯拉普拉 斯滤波器依次对每个所述采样图像中多个图像小块进行滤波处理, 得到每 个所述采样图像中多个图像小块的滤波图像之前, 包括: 根据二维高斯核函数和预设的高斯参数, 生成多个正方形空域二维高 斯滤波器, 所述多个正方形空域二维高斯滤波器中的最大宽度为所述 M;
根据二阶拉普拉斯算子函数, 生成正方形空域二维拉普拉斯滤波器, 若所述正方形空域二维拉普拉斯滤波器的宽度小于所述 M , 对所述正方形 空域二维拉普拉斯滤波器进行像素填充, 以使填充后的所述正方形空域二 维拉普拉斯滤波器的宽度为所述 M;
将所述多个正方形空域二维高斯滤波器转换为多个所述频域高斯滤 波器; 将填充后的所述正方形空域二维拉普拉斯滤波器转换为所述频域拉 普拉斯滤波器;
根据多个所述频域高斯滤波器和所述频域拉普拉斯滤波器, 生成多个 所述频域高斯拉普拉斯滤波器;
所述将所述多个方形空域二维高斯滤波器转换为多个所述频域高斯 滤波器之前, 还包括: 对于长度为 X+M- l、 宽度为 Y+M- 1的方形矩阵左上 方的宽度为所述 M的区域, 使用宽度为所述 M的正方形空域二维高斯滤波 器的像素值进行一一对应替换, 将所述方形矩阵向左和向上循环移位 M/2 个像素, 其中, 所述方形矩阵的所有像素值为 0 ;
所述将填充后的所述正方形空域二维拉普拉斯滤波器转换为所述频 域拉普拉斯滤波器之前, 还包括:
对于长度为 X+M- 宽度为 Y+M- 1的方形矩阵左上方的宽度为所述 M 的区域, 使用宽度为所述 M的正方形空域二维拉普拉斯滤波器的像素值进 行一一对应替换,将所述方形矩阵向左和向上循环移位 M/2个像素,其中, 所述方形矩阵的所有像素值为 0。
5、 根据权利要求 1-4任一项所述的方法, 其特征在于, 所述根据每 个所述采样图像中多个图像小块的滤波图像, 获取每个所述采样图像的图 像兴趣点, 包括:
根据每个所述采样图像中多个图像小块的滤波图像, 获取每个图像小 块的图像兴趣点;
根据每个图像小块的图像兴趣点, 获取每个所述采样图像的图像兴趣 点;
或者, 将每个所述采样图像中多个图像小块的滤波图像进行合并, 得到每个 所述采样图像的滤波图像;
根据每个所述采样图像的滤波图像, 获取每个所述采样图像的兴趣 点。
6、 一种图像兴趣点检测装置, 其特征在于, 包括:
获取模块, 用于获取原始输入图像;
处理模块, 用于对所述原始输入图像进行降采样处理, 得到多个不同 分辨率的采样图像;
划分模块, 用于将每个所述采样图像划分为多个图像小块;
滤波模块, 用于采用高斯拉普拉斯滤波器依次对每个所述采样图像中 多个图像小块进行滤波处理, 得到每个所述采样图像中多个图像小块的滤 波图像;
所述获取模块, 还用于根据每个所述采样图像中多个图像小块的滤波 图像, 获取每个所述采样图像的图像兴趣点。
7、 根据权利要求 6所述的装置, 其特征在于, 还包括: 填充模块和 转换模块;
所述划分模块, 具体用于将每个所述采样图像划分为多个长度为 X、 宽度为 Y的方形图像小块, 其中, X、 Y均为正整数, 若所述采样图像边缘 的图像小块的长度小于 X或者宽度小于 Y , 则所述填充模块对所述采样图 像边缘的图像小块进行像素填充;
所述填充模块, 还用于对每个长度为 X、 宽度为 Y的方形图像小块进 行像素填充, 以使填充后的方形图像小块的长度为 X+M- l、 宽度为 Y+M-1 , M为正整数;
所述转换模块, 用于对填充后的方形图像小块进行离散傅里叶变换, 得到频域图像小块。
8、 根据权利要求 7所述的装置, 其特征在于, 所述滤波模块, 具体 用于, 采用频域高斯拉普拉斯滤波器对每个所述采样图像中多个图像小块 进行多次滤波处理, 得到每个所述采样图像中多个图像小块的频域高斯拉 普拉斯响应图像;
所述转换模块, 还用于对每个所述采样图像中多个图像小块的频域高 斯拉普拉斯响应图像进行离散傅里叶反变换, 得到每个所述采样图像中多 个图像小块的滤波图像。
9、 根据权利要求 8所述的装置, 其特征在于, 还包括: 生成模块; 所述滤波模块采用频域高斯拉普拉斯滤波器依次对每个所述采样图 像中多个图像小块进行滤波处理, 得到每个所述采样图像中多个图像小块 的滤波图像之前, 所述生成模块, 用于根据二维高斯核函数和预设的高斯 参数, 生成多个正方形空域二维高斯滤波器, 所述多个正方形空域二维高 斯滤波器中的最大宽度为所述 M;
所述生成模块还用于, 根据二阶拉普拉斯算子函数, 生成正方形空域 二维拉普拉斯滤波器, 若所述正方形空域二维拉普拉斯滤波器的宽度小于 所述 M , 对所述正方形空域二维拉普拉斯滤波器进行像素填充, 以使填充 后的所述正方形空域二维拉普拉斯滤波器的宽度为所述 M;
所述转换模块, 还用于将所述多个正方形空域二维高斯滤波器转换为 多个所述频域高斯滤波器; 将填充后的所述正方形空域二维拉普拉斯滤波 器转换为所述频域拉普拉斯滤波器;
根据多个所述频域高斯滤波器和所述频域拉普拉斯滤波器, 生成多个 所述频域高斯拉普拉斯滤波器;
所述转换模块将所述多个正方形空域二维高斯滤波器转换为多个所 述频域高斯滤波器之前, 所述填充模块还用于, 对于长度为 X+M- 宽度 为 Y+M-1的方形矩阵左上方的宽度为所述 M的区域, 使用宽度为所述 M的 正方形空域二维高斯滤波器的像素值进行一一对应替换, 将所述方形矩阵 向左和向上循环移位 M/2个像素, 其中, 所述方形矩阵的所有像素值为 0 ; 所述转换模块将填充后的所述正方形空域二维拉普拉斯滤波器转换 为所述频域拉普拉斯滤波器之前, 所述填充模块还用于, 对于长度为 X+M- 宽度为 Y+M- 1的方形矩阵左上方的宽度为所述 M的区域, 使用宽 度为所述 M的正方形空域二维拉普拉斯滤波器的像素值进行一一对应替 换, 将所述方形矩阵向左和向上循环移位 M/2个像素, 其中, 所述方形矩 阵的所有像素值为 0。
10、 根据权利要求 6-9任一项所述的装置, 其特征在于, 所述获取模 块根据每个所述采样图像中多个图像小块的滤波图像, 获取每个所述采样 图像的图像兴趣点中, 所述获取模块具体用于:
根据每个所述采样图像中多个图像小块的滤波图像, 获取每个图像小 块的图像兴趣点;
根据每个图像小块的图像兴趣点, 获取每个所述采样图像的图像兴趣 点;
或者, 所述获取模块具体用于:
将每个所述采样图像中多个图像小块的滤波图像进行合并, 得到每个 所述采样图像的滤波图像;
根据每个所述采样图像的滤波图像, 获取每个所述采样图像的兴趣 点。
PCT/CN2014/074927 2013-04-12 2014-04-08 图像兴趣点检测方法和装置 WO2014166377A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US14/880,981 US9779324B2 (en) 2013-04-12 2015-10-12 Method and device for detecting interest points in image

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201310126255.8A CN103235949B (zh) 2013-04-12 2013-04-12 图像兴趣点检测方法和装置
CN201310126255.8 2013-04-12

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US14/880,981 Continuation US9779324B2 (en) 2013-04-12 2015-10-12 Method and device for detecting interest points in image

Publications (1)

Publication Number Publication Date
WO2014166377A1 true WO2014166377A1 (zh) 2014-10-16

Family

ID=48883988

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2014/074927 WO2014166377A1 (zh) 2013-04-12 2014-04-08 图像兴趣点检测方法和装置

Country Status (3)

Country Link
US (1) US9779324B2 (zh)
CN (1) CN103235949B (zh)
WO (1) WO2014166377A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105608469A (zh) * 2015-12-21 2016-05-25 小米科技有限责任公司 图像分辨率的确定方法及装置

Families Citing this family (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103235949B (zh) 2013-04-12 2016-02-10 北京大学 图像兴趣点检测方法和装置
CN103793520A (zh) * 2014-02-14 2014-05-14 齐齐哈尔大学 一种嵌入图像处理软件的移动视觉搜索方法
EP3580690B1 (en) * 2017-05-24 2020-09-02 Google LLC Bayesian methodology for geospatial object/characteristic detection
CN108496202A (zh) * 2017-08-31 2018-09-04 深圳市大疆创新科技有限公司 斑点检测方法、系统、存储介质及计算机程序产品
CN107862316B (zh) * 2017-10-19 2021-06-04 珠海格力电器股份有限公司 卷积运算方法及装置
CN112204612A (zh) 2018-03-29 2021-01-08 罗伯特·博世有限公司 用于由车辆先验信息引导的基于视觉的车辆内部环境感测的方法和系统
US10846875B2 (en) * 2018-06-07 2020-11-24 Siemens Healthcare Gmbh Adaptive nonlinear optimization of shape parameters for object localization in 3D medical images
WO2020223313A1 (en) * 2019-05-02 2020-11-05 Crown Equipment Corporation Industrial vehicle with feature-based localization and navigation
TWI786409B (zh) 2020-06-01 2022-12-11 聚晶半導體股份有限公司 影像偵測裝置以及影像偵測方法
CN112149728B (zh) * 2020-09-22 2023-11-17 成都智遥云图信息技术有限公司 一种快速的多模态图像模板匹配方法
US11748865B2 (en) 2020-12-07 2023-09-05 International Business Machines Corporation Hierarchical image decomposition for defect detection
CN114677497A (zh) * 2022-05-18 2022-06-28 浙江大华技术股份有限公司 一种图像处理方法及装置

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101540046A (zh) * 2009-04-10 2009-09-23 凌阳电通科技股份有限公司 基于图像特征的全景图拼接方法和装置
CN102254144A (zh) * 2011-07-12 2011-11-23 四川大学 一种鲁棒的图像中二维码区域提取方法
CN103020630A (zh) * 2012-12-05 2013-04-03 曙光信息产业(北京)有限公司 图像特征的处理方法和装置
CN103235949A (zh) * 2013-04-12 2013-08-07 北京大学 图像兴趣点检测方法和装置

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
TWI371959B (en) * 2007-04-13 2012-09-01 Primax Electronics Ltd Image processing method and related partial psf estimation method thereof
US8406507B2 (en) * 2009-01-14 2013-03-26 A9.Com, Inc. Method and system for representing image patches
KR20110065997A (ko) * 2009-12-10 2011-06-16 삼성전자주식회사 영상처리장치 및 영상처리방법
CA2780595A1 (en) * 2011-06-22 2012-12-22 Roman Palenychka Method and multi-scale attention system for spatiotemporal change determination and object detection

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101540046A (zh) * 2009-04-10 2009-09-23 凌阳电通科技股份有限公司 基于图像特征的全景图拼接方法和装置
CN102254144A (zh) * 2011-07-12 2011-11-23 四川大学 一种鲁棒的图像中二维码区域提取方法
CN103020630A (zh) * 2012-12-05 2013-04-03 曙光信息产业(北京)有限公司 图像特征的处理方法和装置
CN103235949A (zh) * 2013-04-12 2013-08-07 北京大学 图像兴趣点检测方法和装置

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105608469A (zh) * 2015-12-21 2016-05-25 小米科技有限责任公司 图像分辨率的确定方法及装置

Also Published As

Publication number Publication date
CN103235949A (zh) 2013-08-07
US9779324B2 (en) 2017-10-03
US20160034780A1 (en) 2016-02-04
CN103235949B (zh) 2016-02-10

Similar Documents

Publication Publication Date Title
WO2014166377A1 (zh) 图像兴趣点检测方法和装置
JP5294343B2 (ja) 画像位置合わせ処理装置、領域拡張処理装置及び画質改善処理装置
EP3032494B1 (en) Using depth for recovering missing information in an image
JP2015225665A (ja) 画像ノイズ除去方法及び画像ノイズ除去装置
RU2631765C1 (ru) Способ и система исправления перспективных искажений в изображениях, занимающих двухстраничный разворот
US9489595B2 (en) Method, apparatus, and device for detecting convex polygon image block
CN107563974B (zh) 图像去噪方法、装置、电子设备及存储介质
WO2014146561A1 (zh) 缩略图生成方法及系统
WO2016009421A1 (en) Automatic image composition
WO2020186385A1 (zh) 图像处理方法、电子设备及计算机可读存储介质
Liu et al. Automatic blur-kernel-size estimation for motion deblurring
US9767533B2 (en) Image resolution enhancement based on data from related images
US9171227B2 (en) Apparatus and method extracting feature information of a source image
CN114758145A (zh) 一种图像脱敏方法、装置、电子设备及存储介质
JP5805001B2 (ja) 画像鮮鋭度評価装置
JP2019027890A (ja) 構造物検査システム、構造物検査装置及び構造物検査方法
CN113963009A (zh) 基于可形变划块的局部自注意力的图像处理方法和模型
JP5705611B2 (ja) 画像の正置からの回転角度検知装置および方法
KR101836238B1 (ko) 다운 스케일링과 비용함수를 이용한 이미지 연결선 추정방법
CN112785651B (zh) 用于确定相对位姿参数的方法和装置
WO2023131236A1 (zh) 一种图像处理方法、装置及电子设备
JP6080424B2 (ja) 対応点探索装置、そのプログラムおよびカメラパラメータ推定装置
CN116320792A (zh) 多帧图像中增强分辨率和降低噪声的方法
CN111028143B (zh) 一种对图像不同比例变换不变特征的设计方法
JP6056354B2 (ja) 画像処理装置、画像処理方法および画像処理プログラム

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 14783116

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205 DATED 15/01/2016)

122 Ep: pct application non-entry in european phase

Ref document number: 14783116

Country of ref document: EP

Kind code of ref document: A1