CN101719275B - And the image feature point extracting method for implementing an image copy detection method and system - Google Patents

And the image feature point extracting method for implementing an image copy detection method and system Download PDF

Info

Publication number
CN101719275B
CN101719275B CN 200910237737 CN200910237737A CN101719275B CN 101719275 B CN101719275 B CN 101719275B CN 200910237737 CN200910237737 CN 200910237737 CN 200910237737 A CN200910237737 A CN 200910237737A CN 101719275 B CN101719275 B CN 101719275B
Authority
CN
China
Prior art keywords
image
candidate
feature points
feature
feature point
Prior art date
Application number
CN 200910237737
Other languages
Chinese (zh)
Other versions
CN101719275A (en
Inventor
张勇东
李锦涛
谢洪涛
Original Assignee
中国科学院计算技术研究所
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 中国科学院计算技术研究所 filed Critical 中国科学院计算技术研究所
Priority to CN 200910237737 priority Critical patent/CN101719275B/en
Publication of CN101719275A publication Critical patent/CN101719275A/en
Application granted granted Critical
Publication of CN101719275B publication Critical patent/CN101719275B/en

Links

Abstract

本发明涉及图像特征点提取和实现方法、图像拷贝检测方法及其系统,所述图像特征点提取方法包括:步骤1,在被提取图像对应的低尺度图像上检测Harris角点,获得候选特征点集合;步骤2,对所述候选特征点集合中的每一个候选特征点,计算所述候选特征点在所述被提取图像对应的多个高尺度图像下Hessian矩阵的DET值,根据所述DET值去除所述候选特征点集合中的误检点,获得所述被提取图像的特征点。 The present invention relates to the image feature point extracting and implement methods, an image copy detection method and system, the image feature point extracting method includes: Step 1, the Harris corner detection on the low scale image corresponding to the image is extracted, the feature point candidate obtained set; step 2, for each candidate feature points in the candidate set, the value of DET calculated Hessian matrix at high scale image of the plurality of candidate feature points are extracted in the image corresponding to, according to the DET removing erroneous values ​​behave in the candidate feature point set, the feature point of the obtained image is extracted. 本发明能够提高检测速度,满足实时性要求。 The present invention can improve the detection speed, to meet real-time requirements.

Description

图像特征点提取和实现方法、图像拷贝检测方法及其系统 And the image feature point extracting method for implementing an image copy detection method and system

技术领域 FIELD

[0001] 本发明涉及图像拷贝检测领域,尤其涉及图像特征点提取方法、实现方法、图像拷贝检测方法和其系统。 [0001] The present invention relates to an image copy detection, and in particular relates to a method of extracting the image feature points, for a method, and a copy of the image detection system thereof.

背景技术 Background technique

[0002] 数字图像由于表现直观、内容生动等特点,成为一种被人们广泛使用的信息载体。 [0002] Because the performance of a digital visual image content vivid features, the information carrier to become a widely used by people. 随着数字图像获取工具、编辑工具和相关技术的普及,获取、制作和修改图像变得越来越简单。 With the popularity of digital image acquisition tools, editing tools and related technologies, access, create and edit images become more and more simple. 如今,越来越多的个人和组织出于不同的目的在网络上不停地传播自制图像,为数字图像内容管理、版权保护,特别是不良图像过滤带来了越来越严重的挑战。 Today, more and more individuals and organizations for different purposes constantly spread on the network self-image, digital image content management, copyright protection, especially poor image filtering has brought more and more serious challenge.

[0003] 基于内容的图像拷贝检测,简称图像拷贝检测,作为一项重要的数字图像管理技术,已成为当前的研究热点。 [0003] The image content-based copy detection, referred to as image copy detection, as an important digital image management technology, has become a hot topic. 图像拷贝检测的任务是给定被检测图像,在图像库中进行查找,检测是否存在内容相似或者相同的图像;如若存在,则该被检测图像为拷贝图像。 Copy of the image is given the task of detecting the detected image, in the image database lookup, similar or identical content detecting whether an image exists; Should present, the image is a copy of the image is detected. 拷贝图像是对源图像经过各种拷贝攻击而得,该些拷贝攻击包括裁剪、旋转、加入字幕或商标、 模糊、放缩、颜色、对比度变化等。 Copy image is a source image obtained through a variety of copy attack, including the plurality of copy attack crop, rotate, add captions or trademarks, blur, zoom, color, contrast changes. 虽然经过拷贝攻击,但是拷贝图像与原图像在内容上是一致的,图像拷贝检测技术需要能够应对该些拷贝攻击。 Even after copy attack, but the copy of the image on the original image content is the same, image copy detection technologies need to be able to cope with this copy of these attacks.

[0004] 与传统的数字水印不同,图像拷贝检测技术是根据图像的内容判断图像之间的相似性,而不必在图像发布前人工加入相关信息。 [0004] and different from conventional digital watermark, the copy image detection technique is to determine the similarity between images according to the contents of the image, without the artificial addition of the image information before release. 图像拷贝检测技术手段用于解决图像的版权保护,不良图像过滤、和图像库中冗余图像去除等技术问题。 Image copy detection technology to solve the technical problem of copyright protection of image, poor image filtering, image library and redundant image removal and so on.

[0005] 当前图像拷贝检测技术方案的关键技术手段为图像特征提取的方法,特征可分为全局特征和局部特征两类。 [0005] Current methods of image copy detection techniques key aspect of the image feature extraction, feature can be divided into two types of local features and global features. 全局特征是对图像内容的整体性描述,如颜色,形状和纹理等。 Wherein the overall description of the global image content, such as color, shape and texture. 全局特征计算简单检测速度快,但是对颜色、亮度和对比度变化比较敏感,也不能有效应对裁剪和遮挡等拷贝攻击。 Calculating global feature simple fast detection speed, but the color, brightness and contrast changes sensitive, and can not effectively deal with crop occlusion copy attack. 现有技术中基于局部特征的方法,把图像表征为若干局部区块的集合,通过查询组成目标图像的区块来检测相关图像。 Prior art methods based on local features, the image is characterized as a collection of several local block by block query component of the target image to detect the correlation image. 局部特征对常见的攻击方式具有一定的鲁棒性,因此得到了广泛的应用。 Local feature is robust against common attacks, so it has been widely used. 但是局部特征计算复杂度高,即使目前速度最快的SURF算法检测一幅640X480像素的图像平均也需要几百毫秒,不能满足大规模数据下实时应用的需求。 However, the localized feature calculation complexity is high, even the fastest of the SURF algorithm detects a 640X480 pixel image also requires several hundred milliseconds average, can not meet the needs of large-scale real-time application data. 如何减少局部特征计算时间复杂度为现有技术中需要解决的一个技术问题。 How to calculate time complexity is a technical problem to be solved in the prior art to reduce the local features. 现有技术中不同的解决方案如视觉关键词等已被提出,但是由于CPU计算能力有限,时间性能提升幅度不大。 In the prior art different solutions as a visual key words it has been proposed, but due to limited CPU power, time performance marginally.

[0006] GPU (Graphics Processing Units图形处理器),通常为显卡,自诞生以来,计算能力以超越摩尔定律的速度发展。 [0006] GPU (Graphics Processing Units graphics processor), usually for the graphics card, since the birth of computing power to go beyond Moore's Law pace. 当前GPU的浮点运算能力远远超过同代CPU,利用GPU进行加速计算成为新的研究热点。 The current GPU's floating point computing power far more than the same generation CPU, the use of GPU accelerated computing has become a new hotspot. GPU在天体计算和图像处理等通用计算领域得到广泛的应用, 与CPU实现相比可以获得1-2个数量级的加速比。 GPU is widely used in general purpose computing objects and image processing calculation, as compared with the CPU to achieve 1-2 orders of magnitude can be obtained speedup. 在图像拷贝检测领域,利用GPU加速已有的局部特征提取算法已被提出,获得时间性能一定的提升。 In the field of image copy detection, extraction algorithms have been proposed to give a certain time to enhance performance using existing GPU-accelerated local features. 已有算法虽然有潜在的并行特性,但是在图像的多尺度模拟过程中由于需要计算每个像素点的响应值而过于耗时;并且需要在不同存储器之间多次拷贝数据,因此不适于GPU。 Although existing algorithms potential parallelism, but the multi-scale process due to the analog image needs to be calculated response value of each pixel and too time consuming; and require multiple copy data between different memory, therefore not suitable GPU . 已有方法没有结合GPU的硬件特性的更高效的局部特征提取算法,因此不能充分利用GPU的计算能力来获得时间性能大幅提升。 The method has no binding properties of the GPU hardware more efficient local feature extraction algorithm, and therefore can not take full advantage of the GPU computing power to obtain the time performance boost. 发明内容 SUMMARY

[0007] 为解决上述问题,本发明提供了图像特征点提取方法、实现方法、图像拷贝检测方法和其系统,能够提高检测速度,满足实时性要求。 [0007] In order to solve the above problems, the present invention provides a method of extracting the image feature points, for a method, and a method for its copy of the image detecting system, the detection speed can be increased, to meet real-time requirements.

[0008] 本发明公开了一种图像特征点提取方法,包括: [0008] The present invention discloses an image feature point extracting method, comprising:

[0009] 步骤1,在被提取图像对应的低尺度图像上检测Harris角点,获得候选特征点集合; [0009] Step 1, the Harris corner detection on the low scale image corresponding to the image is extracted, obtaining a set of candidate feature points;

[0010] 步骤2,对所述候选特征点集合中的每一个候选特征点,计算所述候选特征点在所述被提取图像对应的多个高尺度图像下Hessian矩阵的DET值,根据如果所述候选特征点的DET值在所述多个高尺度图像构成的尺度空间中没有极大值或所述极大值小于预设阈值,则所述候选特征点为误检点,将所述候选特征点从所述候选特征点集合中除去;所述候选特征点集合中剩余的候选特征点为被提取图像的特征点。 [0010] Step 2, for each candidate feature points in the candidate set calculating the candidate feature points are extracted DET value of Hessian matrix at a plurality of high-scale image corresponding to the image, if according to the DET value of said feature point candidate is not the maximum value of the scale space image formed of a plurality of high-scale or maximum value is less than the predetermined threshold value, the feature point candidate behave error, the candidate feature point removed from the set of candidate feature points; the candidate feature point set in the remaining candidate feature points of the image feature points are extracted.

[0011] 所述步骤1进一步为, [0011] step 1 is further,

[0012] 步骤21,在被提取图像对应的低尺度图像上检测Harris角点,所述低尺度图像是由被提取图像和高斯函数做卷积获得,高斯函数的标准差为所述低尺度图像的尺度; [0012] Step 21, the Harris corner detection on the low scale image corresponding to the image is extracted, the low-scale image is obtained from the extracted image and a Gaussian function convolve, standard deviation of the Gaussian function is low image scale scale;

[0013] 步骤22,将所述Harris角点中在同一像素位置重复出现的角点去除,剩余角点组成所述候选特征点集合。 [0013] Step 22, the Harris corner at the same pixel position of the corner points of the repeated removal, the composition of the remaining candidate corners feature point set.

[0014] 所述步骤2进一步为, [0014] Step 2 is further said to,

[0015] 步骤31,对于所述每一个候选特征点,计算所述候选特征点在所述被提取图像对应的每个高尺度图像下Hessian矩阵的DET值; [0015] Step 31, for each of the candidate feature points, calculating the candidate feature points are extracted for each value DET high scale image corresponding to the image of the Hessian matrix;

[0016] 步骤32,对于所述每一个候选特征点,如果所述候选特征点的DET值在所述多个高尺度图像构成的尺度空间中没有极大值或极大值小于预设阈值,则所述候选特征点为误检点,将所述候选特征点从所述候选特征点集合中除去; [0016] Step 32, for each of the candidate feature points, if the DET value candidate feature point is not the maximum value of the plurality of scale space images consisting of high-scale or maximum value is less than a predetermined threshold value, then the candidate feature points behave as erroneous, the candidate feature point is removed from the set of candidate feature points;

[0017] 步骤33,所述候选特征点集合中未被除去的候选特征点为所述被提取图像的特征 [0017] Step 33, the candidate feature point set not removed the candidate feature points are extracted as feature image

点ο Point ο

[0018] 所述步骤2还包括, [0018] 2 further comprising the step of,

[0019] 步骤41,如果所述候选特征点的DET值在所述多个高尺度图像构成的尺度空间中有极大值,并且所述极大值大于所述预设阈值,则所述候选特征点为所述被提取图像的特征点,所述特征点的尺度为所述极大值对应的高尺度图像的尺度。 [0019] Step 41, if the DET value candidate feature point has a maximum value in the scale space image composed of the plurality of high standards, and the maximum value is greater than the preset threshold value, the candidate the feature points being extracted image feature points, the scale of the feature point is the maximum dimension of the high-scale image corresponding to the value.

[0020] 本发明还公开了一种在GPU上实现图像特征点提取的方法,包括: [0020] The present invention also discloses a method for implementing image feature points extracted on the GPU, comprising:

[0021] 步骤51,根据被提取图像的大小配置GPU计算内核,将所述被提取图像的数据拷贝到GPU的纹理存储器; [0021] Step 51, compute kernel GPU configuration according to the size of the image is extracted, the extracted image data is copied to the texture memory of the GPU;

[0022] 步骤52,GPU按预设低尺度处理被提取图像,获得所述被提取图像对应的低尺度图像,在所述低尺度图像上检测Harris角点,以所述Harris角点组成候选特征点集合,将所述候选特征点集合存储到GPU的全局存储器; [0022] Step 52, GPU processing according to a preset low-scale image is extracted, is extracted to obtain the low-scale image corresponding to the image, the Harris corner detection on the low scale image to the Harris corner point candidates composition wherein set point, the set of candidate feature points stored in the GPU global memory;

[0023] 步骤53,GPU按多个预设高尺度处理被提取图像,获得所述被提取图像对应的多个高尺度图像,对所述候选特征点集合中的每一个候选特征点,计算所述候选特征点在所述多个高尺度图像下Hessian矩阵的DET值,如果所述候选特征点的DET值在所述多个高尺度图像构成的尺度空间中没有极大值或所述极大值小于预设阈值,则所述候选特征点为误检点,将所述候选特征点从所述候选特征点集合中除去;所述候选特征点集合中剩余的候选特征点为被提取图像的特征点,将所述特征点存储到所述GPU的全局存储器。 [0023] Step 53, GPU processing by the plurality of preset high scales the extracted image to obtain a plurality of said extracted high-scale image corresponding to the image, for each of the candidate feature points in the candidate set, calculating DET said candidate feature points of the plurality of high value in the Hessian matrix scale image, if the feature point candidate value is not the maximum value DET scale space in the plurality of high-scale image consisting of said maximum or value is less than a predetermined threshold value, the feature point candidate behave error, the candidate feature point is removed from the set of candidate feature points; the candidate feature point set in the remaining candidate feature point is extracted as the image feature point, the feature point stored in global memory of the GPU.

[0024] 所述步骤51中还包括, [0024] The step 51 further comprises,

[0025] 步骤61,将用于计算卷积和差分的模板数据拷贝到GPU的常量存储器; [0025] Step 61, calculating the convolution for copying the template and difference data to GPU memory constant;

[0026] 所述步骤52进一步为, [0026] The step 52 is further,

[0027] 步骤62,GPU根据所述模板计算所述被提取图像和以所述预设低尺度为标准差的高斯函数的卷积,经卷积后的被提取图像存放于GPU的纹理存储器; [0027] Step 62 is, GPU computing and the extracted image to the predetermined standard deviation is low dimension convolution of the Gaussian function according to the template, is extracted after convolution of the image stored in the GPU texture memory;

[0028] 步骤63,GPU计算经卷积后的被提取图像在水平方向和垂直方向的差分,差分结果存放于GPU的纹理存储器; [0028] Step 63, the extracted image convolution GPU computing the difference in the horizontal direction and the vertical direction, the difference results stored in the GPU texture memory;

[0029] 步骤64,GPU将差分结果进行高斯模糊,经高斯模糊处理后的差分结果存放于GPU 的全局存储器; [0029] Step 64, the GPU Gaussian blur difference results, the differential result by a Gaussian blur processing is stored in the GPU global memory;

[0030] 步骤65,GPU根据差分结果计算Harris角点响应值,将Harris角点响应值存放于GPU的纹理存储器;根据所述Harris角点响应值判断所述被提取图像中的像素点是否为Harris角点,确定Harris角点的位置,将所述Harris角点的位置坐标存放于GPU的全局存储器。 [0030] Step 65, GPU Harris corner response value is calculated based on the difference result, the Harris corner response value stored in the texture memory of the GPU; the extracted image pixels according to whether the Harris corner response value determination Harris corner point, determining the position of the Harris corner, the position coordinates of the Harris corner point stored in the GPU global memory.

[0031] 所述步骤53进一步为, [0031] The step 53 is further,

[0032] 步骤71,GPU按多个预设高尺度对被提取图像中候选特征点和候选特征点的相邻区域进行高尺度模拟,获得候选特征点和相邻区域在多个高尺度图像对应区域,对于所述每一个候选特征点,计算所述候选特征点在所述被提取图像对应的每个高尺度图像下Hessian矩阵的DET值,将所述DET值存放于GPU的纹理存储器; [0032] Step 71, GPU according to a plurality of predetermined adjacent area of ​​high-scale image is extracted in the candidate feature points and feature point candidate high-scale simulation, obtaining a plurality of candidate feature points and the adjacent region of high scale image corresponding to region, for each of the candidate feature points, calculating the candidate feature points are extracted for each value DET high scale image corresponding to the image in the Hessian matrix, the value of DET is stored in the GPU texture memory;

[0033] 步骤72,对于每个候选特征点,GPU判断如果所述候选特征点的DET值在所述多个高尺度图像构成的尺度空间中没有极大值或极大值小于预设阈值,则所述候选特征点为误检点,将所述候选特征点从所述候选特征点集合中去除;将所述候选特征点集合中剩余的候选特征点作为所述被提取图像的特征点存放于GPU的全局存储器。 [0033] Step 72, feature points for each candidate, the GPU DET determines if the value of the feature point candidate is not the maximum value of the scale space image formed of a plurality of high-scale or maximum value is less than a predetermined threshold value, then the candidate feature points behave as erroneous, the candidate feature point is removed from the set of candidate feature points; the feature point set in the candidate remaining candidate as the feature points extracted image feature points stored in GPU global memory.

[0034] 本发明还公开了一种应用如上所述的图像特征点提取方法的图像拷贝检测方法, 包括: [0034] The present invention also discloses a method for extracting an image copying method for detecting image feature points in an application as described above, comprising:

[0035] 步骤81,按所述图像特征点提取方法提取被检测图像的图像特征点,计算每一个所述图像特征点的SIFT特征和OM顺序度量特征; [0035] Step 81, according to the image feature point extracting method for extracting the image feature points are detected in the image, and calculates OM SIFT features of the image sequence of each feature point feature measure;

[0036] 步骤82,对图像库中的每一幅图像,按所述图像特征点提取方法提取所有图像的图像特征点,计算所述图像特征点的SIFT特征和OM顺序度量特征; [0036] Step 82, the image library on each image, the image by the feature point extraction method of extracting the image feature points of all the images, the SIFT features and OM sequentially calculates the image feature point feature measure;

[0037] 步骤83,根据所述被检测图像和所述图像库中图像的SIFT特征和OM顺序度量特征进行拷贝检测。 [0037] Step 83, the copy detection metric based on the features detected SIFT and OM sequential images and the images in the image library.

[0038] 所述步骤82进一步为, [0038] The step 82 is further,

[0039] 步骤91,对图像库中的每一幅图像,按所述图像特征点提取方法提取所有图像的图像特征点,计算所述图像特征点的SIFT特征和OM顺序度量特征;将图像库中所有图像特征点的SIFT特征建立为ANN索引树,将图像库中所有图像特征点的OM顺序度量特征以文件的方式存放; [0039] Step 91, the image library on each image, the image by the feature point extraction method of extracting the image feature points of all the images, the SIFT features of the image feature points is calculated and the measurement sequence characteristic OM; image library All SIFT features of the image feature points established as ANN index tree, the order of all the metric OM wherein the image feature point in the image library stored in files;

[0040] 所述步骤83进一步为,[0041] 步骤92,以被检测图像的每个特征点为查询点,根据所述查询点的SIFT特征进行初始ANN检索,获得初始检测结果,初始检测结果为图像库中图像的图像特征点,所述初始检测结果的SIFT特征同查询点的SIFT特征的距离小于第二预设阈值; [0040] The step 83 is a further, [0041] Step 92, to each feature point is detected as a query image point, according to the initial ANN retrieves SIFT feature points of the query to obtain the initial test results, the initial results of the detection an image feature point image in the image database, the detection result of the initial query SIFT feature SIFT feature point with a distance smaller than a second predetermined threshold value;

[0042] 步骤93,去除掉所述初始检测结果中OM顺序度量特征同所述查询点的OM顺序度量特征不一致的特征点,所述初始检测结果中未被去除的特征点为同所述查询点最匹配的特征点,所述最匹配的特征点组成集合; [0042] step 93, to remove the initial detection result order OM metric measure characteristics inconsistent with the feature point feature OM order of the query point, the initial feature point detection result is not removed with the query point characteristic point matching, the feature points best match to form a set;

[0043] 步骤94,对所述集合中的特征点所属的所有图像,以投票的方式决定所述图像的排名,排名为前η的图像为所述被检测图像的相似图像,η为预设值。 [0043] Step 94, all the image feature point in the set belongs, by voting determines the ranking of the image, rank [eta] is a front image of the similar image is detected image, [eta] is a preset value.

[0044] 本发明还公开了一种图像特征点提取装置,包括: [0044] The present invention also discloses an image feature point extracting means, comprising:

[0045] 候选特征点集合获得模块,用于在被提取图像对应的低尺度图像上检测Harris 角点,获得候选特征点集合; [0045] The candidate feature point set acquisition module for detecting a Harris corner on the low scale image corresponding to the image is extracted, obtaining a set of candidate feature points;

[0046] 特征点提取模块,用于对所述候选特征点集合中的每一个候选特征点,计算所述候选特征点在所述被提取图像对应的多个高尺度图像下Hessian矩阵的DET值,如果所述候选特征点的DET值在所述多个高尺度图像构成的尺度空间中没有极大值或所述极大值小于预设阈值,则所述候选特征点为误检点,将所述候选特征点从所述候选特征点集合中除去;所述候选特征点集合中剩余的候选特征点为被提取图像的特征点。 DET value of Hessian matrix at a plurality of high-scale image [0046] The feature point extraction means for each candidate feature points in the candidate set calculating the candidate feature points are extracted in the image corresponding to , if the DET value candidate feature point is not the maximum value or the maximum value less than the preset threshold value of the scale space image formed of a plurality of high standards, then the candidate feature points behave as erroneous, the removing said candidate feature points from the feature point candidate set; the set of candidate feature points remaining candidate feature points of the image feature points are extracted.

[0047] 所述候选特征点集合获得模块进一步用于在被提取图像对应的低尺度图像上检测Harris角点,所述低尺度图像是由被提取图像和高斯函数做卷积获得,高斯函数的标准差为所述低尺度图像的尺度;将所述Harris角点中在同一像素位置重复出现的角点去除, 剩余角点组成所述候选特征点集合。 [0047] The set of candidate feature points on the obtained module is further configured to low-scale image corresponding to the image is extracted Harris corner detector, the low-scale image is obtained from the extracted image and a Gaussian convolution to do, a Gaussian function standard deviation of the mesoscale dimension image; Harris corner point of the corner points at the same pixel position recurring removed, the composition of the remaining candidate corners feature point set.

[0048] 所述特征点提取模块进一步用于对于所述每一个候选特征点,计算所述候选特征点在所述被提取图像对应的每个高尺度图像下Hessian矩阵的DET值;对于所述每一个候选特征点,如果所述候选特征点的DET值在所述多个高尺度图像构成的尺度空间中没有极大值或极大值小于预设阈值,则所述候选特征点为误检点,将所述候选特征点从所述候选特征点集合中除去;所述候选特征点集合中未被除去的候选特征点为所述被提取图像的特征点。 [0048] The feature extraction module is further configured to point to said each candidate feature points, calculating the candidate feature points are extracted DET value of Hessian matrix at each high-scale image corresponding to the image; to the each candidate feature point, if the DET value candidate feature point is not the maximum value of the plurality of scale space images consisting of high-scale or maximum value is less than a predetermined threshold value, the feature point candidate erroneous discreet , the candidate feature point is removed from the set of candidate feature points; the candidate feature point set not removed candidate feature points are extracted as the feature point image.

[0049] 所述特征点提取模块还用于如果所述候选特征点的DET值在所述多个高尺度图像构成的尺度空间中有极大值,并且所述极大值大于所述预设阈值,则确认所述候选特征点为所述被提取图像的特征点,所述特征点的尺度为所述极值对应的高尺度图像的尺度。 [0049] The feature point extraction module is further configured to, if the feature point candidate DET value has a maximum value in the scale space image composed of the plurality of high standards, and the maximum value is greater than the preset threshold, then confirming that the candidate feature points are extracted as the feature point of the image, the scale of the feature point of the high scale-scale image corresponding to the extreme value.

[0050] 本发明还公开了一种图像拷贝检测系统,包括: [0050] The present invention also discloses an image copy detection system, comprising:

[0051] 如前所述的图像特征点提取装置, Image feature points [0051] The extraction device as described above,

[0052] 被检测图像特征生成模块,用于使用所述图像特征点提取装置提取被检测图像的图像特征点,计算每一个所述图像特征点的SIFT特征和OM顺序度量特征; [0052] The generated image feature detection module, for using the image feature point extracting means extracts image feature points detected image, calculate and OM SIFT features of the image sequence of each feature point feature measure;

[0053] 图像库特征生成模块,用于对图像库中的每一幅图像,使用所述图像特征点提取装置提取所有图像的图像特征点,计算所述图像特征点的SIFT特征和OM顺序度量特征; [0053] Image generation module library feature, the image database for each image, using the image feature point extraction means extracts the image feature points of all the images, calculating the image feature points in order of metrics SIFT and OM feature;

[0054] 拷贝检测模块,用于根据所述SIFT特征和OM顺序度量特征进行拷贝检测。 [0054] The copy detection means for detecting copying metric features according to the order of SIFT and OM.

[0055] 所述图像库特征生成模块进一步用于对图像库中的每一幅图像,使用所述图像特征点提取装置提取所有图像的图像特征点,计算所述图像特征点的SIFT特征和OM顺序度量特征;将图像库中所有图像特征点的SIFT特征建立为ANN索引树,将图像库中所有图像特征点的OM顺序度量特征以文件的方式存放; [0055] The features of the image library module is further configured to generate the image database each image, using the image feature point extraction means extracts the image feature points of all the images, calculates the image feature point SIFT and OM characterized metric sequence; the image database SIFT features all established as the image feature points ANN index tree, the order of all the metric OM wherein the image feature point in the image library stored in files;

[0056] 所述拷贝检测模块进一步用于以被检测图像的每个特征点为查询点,根据所述查询点的SIFT特征进行初始ANN检索,获得初始检测结果,初始检测结果为图像库中图像的图像特征点,所述初始检测结果的SIFT特征同查询点的SIFT特征的距离小于第二预设阈值;去除掉所述初始检测结果中OM顺序度量特征同所述查询点的OM顺序度量特征不一致的特征点,所述初始检测结果中未被去除的特征点为同所述查询点最匹配的特征点,所述最匹配的特征点组成集合;对所述集合中的特征点所属的所有图像,以投票的方式决定所述图像的排名,排名为前η的图像为所述被检测图像的相似图像,η为预设值。 [0056] The image database copy detection module is further used for the image of each feature point in the image is detected query point, initial ANN retrieves The SIFT feature points of the query, obtaining the initial detection result, the detection result of the initial image feature points, from the query point with SIFT feature detection results of the initial SIFT feature is less than a second predetermined threshold value; to get rid of the initial detection result OM wherein the order of metrics the same metric feature point of the query sequence OM inconsistent feature points in the initial feature point detection result is not removed with the point of best match the query feature points, said feature points best match to form a set; all feature points in the set belongs image, by voting determines the ranking of the image, [eta] is a front rank of the detected images of the similar image, [eta] is a preset value.

[0057] 本发明的有益效果在于在提取图像特征点时,通过检测候选特征点和对候选特征点进行筛选,能够提高检测速度,满足实时性要求;通过避免了对整个图像进行高尺度模拟,减少了计算量提高了计算速度;GPU上采用本发明的提取图像特征点方法,能够避免了在GPU的不同层次的存储器间频繁交换数据,同现有技术相比具有更好的并行性,更适合于GPU加速实现和优化;在拷贝检测中使用ANN索引树提升了检测速度,采用OM顺序度量特征进行分层匹配保证了检测精度,在保持高检测精度的同时,实现了实时图像拷贝检测。 [0057] Advantageous effects of the present invention is that at the time of extracting the image feature points, and by detecting a candidate feature point pair candidate feature points screening possible to improve the detection speed to meet real-time requirements; by avoiding high scale simulation of the entire image, reduces the computation calculation speed is improved; extracting the image feature points using the method of the present invention on the GPU, can be avoided between the different levels of the memory in the GPU frequent exchange data with better parallelism as compared with the prior art, and more adapted to accelerate and optimize the GPU; ANN using the index tree to enhance detection speed in the copy detection using OM wherein stratified matching metric order to ensure the detection accuracy while maintaining a high detection precision, to achieve a real-time image copy detection.

附图说明 BRIEF DESCRIPTION

[0058] 图1是本发明图像特征点提取方法的流程图; [0058] FIG. 1 is a flowchart of a method of extracting the image feature points of the present invention;

[0059] 图2是本发明在GPU上实现图像特征点提取方法的流程图; [0059] FIG 2 is a flowchart of the present invention, image feature points extraction method implemented on the GPU;

[0060] 图3是本发明在GPU上提取图像特征点的数据划分与分配示例图; [0060] FIG. 3 of the present invention is to extract the data partitioning and allocation example of the image feature points on the GPU;

[0061] 图4是在GPU上提取具有尺度不变性的图像局部特征点的执行模式; [0061] FIG. 4 is performed to extract feature points of the pattern image having a local scale-invariant on the GPU;

[0062] 图5是应用本发明图像特征点提取方法的图像拷贝检测方法的流程图; [0062] FIG. 5 is a flowchart of a method of detecting an image copy image feature points extraction method of the present invention is applied;

[0063] 图6是本发明图像特征点提取装置的结构图; [0063] FIG. 6 is the image feature point extracting configuration diagram of the present invention apparatus;

[0064] 图7是本发明图像拷贝检测系统的结构图。 [0064] FIG. 7 is a copy of the image detection system configuration diagram of the present invention.

具体实施方式 Detailed ways

[0065] 下面结合附图,对本发明做进一步的详细描述。 [0065] DRAWINGS The present invention will be further described in detail.

[0066] 本发明图像特征点提取方法的流程如图1所示。 [0066] The present invention is the image feature point extracting process flow shown in Fig.

[0067] 步骤S101,在被提取图像对应的低尺度图像上检测Harris角点,获得候选特征点 [0067] step S101, the Harris corner detection on the low scale image corresponding to the image is extracted, the feature point candidate obtained

[0068] 所述步骤SlOl进一步为, [0068] The step of further SlOl,

[0069] 步骤1011,在被提取图像对应的低尺度图像上检测Harris角点,低尺度图像是由被提取图像和高斯函数做卷积获得,高斯函数的标准差为该低度图像的尺度 [0069] Step 1011, the Harris corner detection on the low scale image corresponding to the image is extracted, low-scale images is done by a convolution of the extracted image and a Gaussian function is obtained for the standard deviation of the Gaussian function of the image mesoscale

[0070] 步骤1012,将所述Harris角点中在同一像素位置重复出现的角点去除,剩余角点组成所述候选特征点集合。 [0070] Step 1012, the Harris corner at the same pixel position of the corner points of the repeated removal, the composition of the remaining candidate corners feature point set.

[0071] 实施例中,在低尺度,表示为σ C1,下检测被提取图像的Harris角点得到候选特征点集合,表示为Q。 [0071] In the embodiment, the low scale, expressed as σ C1, the Harris corner detection is extracted candidate image feature point set obtained, expressed as Q. %取0,则表示直接在原图上检测角点;%在一个的区间内取值,实施例中为[0,1.2]内取值,则表示在被提出图像的低尺度空间检测角点。 Set to 0%, then the corner points are detected directly on the original;% in value within a range of, for embodiment [0,1.2] value within the embodiment, the detector indicates a low corner scale space image is presented. 如果只在原图上检测,得到的结果为Ci ;若%取多个值,则将检测得到的结果中重复的Harris角点去除后, 得到最终的Ci。 If only on the original detection result is Ci of;% if taking a plurality of values, the detection result will be obtained in repeated Harris corner after the removal, to obtain the final Ci. [0072] 当直接在原图上检测角点时,采用公式一进行Harris角点检测。 [0072] When corner points are detected directly on the original, using a formula for a Harris corner detector.

[0073] R = detM-k (traceM)2 公式一 [0073] R = detM-k (traceM) 2 Formula a

“I2 II ' "I2 II '

[0074] 其中: [0074] wherein:

Figure CN101719275BD00111

和仁分别是图像1“7)在1和7方向的偏导数,1和 Kazuhito are 1 "7) the partial derivatives of the image 1 and 7 in the direction, and 1

y表示水平方向和垂直方向。 y represents the horizontal and vertical directions. k是经验值,实施例中取值范围是0.04至0.06。 k is an empirical value, the embodiment is in the range 0.04 to 0.06. 在得到的R 值矩阵里,R表示Harris算子响应值,若R中某一元素同时满足大于阈值T0,实施例中阈值TO预设为5000,并且是5X5邻域内的局部极大值,则认为该像素点是Harris角点。 R-values ​​obtained in the matrix, R represents a Harris operator response value, if an element R is greater than the threshold value satisfies T0, TO 5000 Example preset threshold embodiment, and is a local maxima 5X5 neighborhood, then the pixel is considered Harris corner.

[0075] 极大值为R中元素中的最大值,并且该最大值两侧都有比该元素小的邻近值。 [0075] R is the maximum value of the maximum element, and on both sides of the maximum value is smaller than the value of the adjacent element.

[0076] 当σ ^在一个区间内取值时,进行多尺度Harris角点检测。 [0076] When σ ^ values ​​within a range, multiscale Harris corner detection. 实施例中σ ^的取值为0和1. 2。 Example values ​​σ ^ 0 and 1.2.

[0077] [0077]

Figure CN101719275BD00112

[0078] Ix(X,oD)和Iy(X,oD)是将被提取图I(X),其中X= (x,y)X表示图中像素点,χ 和y表示水平和垂直方向坐标,与以低尺度ο D为标准差的高斯函数进行卷积,计算所述卷积后的被提取图像在χ和y方向的偏导数。 [0078] Ix (X, oD) and Iy (X, oD) is to be extracted in FIG. I (X), where X = (x, y) X represents a figure pixel, χ and y represent the horizontal and vertical coordinates , ο D scale with a low standard deviation of convolving a Gaussian function partial derivatives are calculated of the extracted image convolution χ and y directions. g(koD)是所述高斯函数,K为常数,实施例中取0. 7。 g (koD) is the Gaussian function, K is a constant, taken Example 0.7.

[0079] 步骤S102,对所述候选特征点集合中的每一个候选特征点,计算所述候选特征点在所述被提取图像对应的多个高尺度图像下Hessian矩阵的DET值,根据所述DET值去除所述候选特征点集合中的误检点,获得所述被提取图像的特征点。 [0079] step S102, a candidate for each of the feature points in the candidate set calculating the candidate feature points are extracted DET value of Hessian matrix at a plurality of high-scale image corresponding to the image, according to the DET error values ​​behave feature point set in the candidate removed, the obtained feature points are extracted image.

[0080] 所述步骤S102进一步为, [0080] The step S102 is further,

[0081] 步骤1021,对于每一个候选特征点,计算所述候选特征点在所述被提取图像对应的每个高尺度图像下Hessian矩阵的DET值。 [0081] Step 1021, for each candidate feature points, calculating the candidate feature points are extracted DET value of Hessian matrix at each high-scale image corresponding to the image in the.

[0082] 优选实施方式中,GPU按多个预设高尺度处理对被提取图像中候选特征点和候选特征点的邻近区域进行高尺度模拟,获得所述被提取图像对应的多个高尺度图像。 [0082] In a preferred embodiment, a plurality of preset high by the GPU scale processing on the extracted image feature points and the candidate feature point neighboring region candidate high-scale simulation, obtaining a plurality of said extracted high-scale image corresponding to the image .

[0083] 步骤1022,对于所述每一个候选特征点,如果所述候选特征点的DET值在所述多个高尺度图像构成的尺度空间中没有极大值或所述极大值小于预设阈值,则所述候选特征点为误检点,将所述候选特征点从所述候选特征点集合中除去。 [0083] Step 1022, for each of the candidate feature points, if the DET value candidate feature point is not the maximum value in the plurality of high scale space-scale image consisting of the maximum value or less than a preset threshold, then the candidate feature points behave as erroneous, the candidate feature point is removed from the set of feature point candidates.

[0084] 首先判断候选特征点的DET值在多个高尺度图像构成的尺度空间中是否有极大值,如果没有,则该候选特征点为误检点;如果有,则判断该极大值是否小于预设阈值,如果小于则该候选特征点为误检点。 [0084] DET is first determined whether the value of the feature point candidate has a maximum value in the scale space image formed of a plurality of high standards, if not, the feature point candidate erroneous discreet; if there is, it is determined whether the maximum value less than a preset threshold value, it is less than the candidate feature points behave as erroneous.

[0085] 步骤1023,所述候选特征点集合中剩余的候选特征点为被提取图像的特征点。 [0085] Step 1023, the set of candidate feature points remaining candidate feature points of the image feature points are extracted.

[0086] 实施例 [0086] Example

[0087] 按公式三,对候选特征点集合Ci中的每一个候选特征点计算该候选特征点在高尺度图像下Hessian矩阵的DET值,根据DET值去除误检点,并精确确定特征点的尺度。 [0087] according to formula III of candidate feature points of each candidate feature points Ci calculation DET value of the candidate feature points at a high scale image Hessian matrix set, removing erroneous behave according DET value, and precisely determine the scale of the feature point .

[0088] DET(H) = I IxxX Iyy-(0.9Ixy)21 公式三 [0088] DET (H) = I IxxX Iyy- (0.9Ixy) 21 with formula 3

[0089] Hessian 失巨阵:[0090] [0089] Hessian loss downline: [0090]

Figure CN101719275BD00121

公式四 Four formula

[0091] 其中,Ixx (X,o)、Iyy(X,σ)和Ixy(X,σ)分别为被提取图I(X)在像素点X = (χ, y)和标准差为σ的高斯函数G(O)做卷积运算,取x、y方向的二阶导数和xy方向的混合偏导数。 [0091] wherein, Ixx (X, o), Iyy (X, σ) and Ixy (X, σ), respectively, are extracted in FIG. I (X) at the pixel X = (χ, y) and the standard deviation [sigma] of Gaussian function G (O) convolve, taking x, mixing the second derivative y-direction and xy directions partial derivative.

[0092] 实施例中,误检点去除和尺度确定过程如下所述。 [0092] In an embodiment, erroneous determination process scale removal and behave as follows.

[0093] 对于Ci中的每一个候选特征点,计算该候选特征点在多个高尺度ση下的DET值。 [0093] For each candidate feature points Ci of calculating the candidate feature points DET high value at a plurality of scales ση. 高尺度选取ση = 1·2ησ。 High standards selected ση = 1 · 2ησ. (η = 1,2,···,Ν),σ。 (Η = 1,2, ···, Ν), σ. = 1·2,Ν 为8。 = 1 · 2, Ν 8.

[0094] 若候选特征点在多个高尺度ση构成的尺度空间内,DET值没有极大值,则认为该点是误检点加以去除。 [0094] If the candidate feature points in the scale space composed of a plurality of high standards ση, the DET value is not the maximum value, it is considered that point be removed behave in error.

[0095] 若候选特征点在高尺度为σn时DET值取得极大值并大于阈值Tl,Tl设为10,该候选特征点为特征点,对应尺度为σ η,若候选特征点的DET值小于Tl,则该候选特征点是误检点,要去除。 [0095] If the candidate feature points is high scale values ​​DET and achieved great value σn is greater than the threshold value Tl, Tl is set to 10, the feature point is the feature point candidate, for the corresponding dimension of σ η, DET when the value of the feature point candidate is less than Tl, then the candidate feature points behave in error, to be removed.

[0096] 如果所述候选特征点的DET值在所述多个高尺度图像构成的尺度空间中有极大值,并且所述极大值大于预设阈值Tl,则该候选特征点为被提取图像的特征点,所述特征点的尺度为所述极值对应的高尺度图像的尺度。 [0096] If the candidate feature point DET value has a maximum value in the scale space image composed of the plurality of high standards, and the maximum value is greater than a predetermined threshold value Tl, then the candidate feature point is extracted feature point of the image, the scale of the feature point of the high scale-scale image corresponding to the extreme value.

[0097] 极大值为离散值的拟合函数的峰值,极大值为离散值组成的集合中的最大值,并且该最大值两侧都有邻近值。 Set peak fit function [0097] The maximum value of discrete values, the maximum value of discrete values ​​consisting of the maximum value, and that on both sides adjacent to the maximum value. 比如有一系列尺度1、2、3、4、5构成的尺度空间,候选特征点的DET值分别为10、20、50、30、10,则50就是极大值,对应的尺度为3 ;如果候选特征点的DET值分别为10、20、30、10、50,则因为50只有一侧有相邻点,因而该候选特征点集合中没 There are a series of scale such as scale space 1,2,3,4,5 configuration, the DET value candidate feature points are 10,20,50,30,10, 50 is the maximum value, corresponding to the scale 3; if DET value candidate feature points are 10,20,30,10,50, since the point 50 adjacent to one side only, and thus the candidate feature point set not

有极大值。 Great value.

[0098] 图2是本发明在GPU中实现图像特征点提取方法的流程。 [0098] FIG 2 is a flowchart of the present invention, image feature points extraction method implemented in the GPU.

[0099] 步骤S201,根据被提取图像的大小配置GPU计算内核,将该被提取图像的数据拷贝到GPU的纹理存储器。 [0099] step S201, the GPU compute kernel configured according to the size of the image is extracted, the extracted image data is copied to the texture memory of the GPU.

[0100] 步骤S201中还包括,将用于计算卷积和差分的模板数据拷贝到GPU的常量存储 [0100] further comprising a step S201, the convolution calculation for copying the template and difference data stored in the GPU constants

ο ο

[0101] 内核设置包括,线程块(Block)数量、块内线程(Thread)数量。 [0101] provided comprising a core, (Block) blocks the number of threads, the thread blocks (Thread) number. 依据待处理问题和GPU硬件参数进行划分。 Divided according to the pending issues and GPU hardware parameters.

[0102] 本发明中特征点提取涉及基本的图像卷积和差分运算,因此均勻地划分图像数据在各个线程块之间分配。 [0102] The present invention relates to a basic feature point extraction image difference operation and convolution, so evenly divided image data blocks partitioned between threads. 如图3所示,每个线程块处理一块图像数据,每一个线程处理一个像素的相关计算,假设图像大小为widthXheight,其中width,height分别表示图像的宽和高,每个线程块内有mXn个线程,则需要的线程块个数按公式五计算。 3, each thread processing a block of image data, each pixel of a thread processing a correlation calculation, assuming that the image size is WIDTHxHEIGHT, wherein the width, height indicate the width and height of the image, each have a thread block mXn the number of threads in thread block is required by the calculation formula V.

[0103] Block_num = width/mXheight/n 公式五 [0103] Block_num = width / mXheight / n five formula

[0104] 实施例中使用的GPU是NVIDIA 9800GTX+显卡,每个线程块设定16 X 16个线程。 GPU [0104] Example embodiments are used NVIDIA 9800GTX + graphics, each block setting thread 16 X 16 threads.

[0105] 本发明以模板模拟高斯函数做卷积运算,当选定具体的尺度参数后模板相当于常量。 [0105] In the present invention, a template do analog convolution Gaussian function, when a specific scale parameter constant corresponding to the selected template. 因此在系统启动时把该些模板数据拷贝到GPU的常量存储器(constant memory)以避免重复拷贝数据。 Therefore, when the plurality of boot copies of the template data memory to the GPU constants (constant memory) in order to avoid duplicate copies of data. 在计算卷积时需要访问一个像素点的邻域数据,会产生边界问题。 In calculating the convolution needs to access a pixel neighborhood data, boundary problems occur. 本发明采用纹理存储器(texture memory)存储被提取图像数据,通过设置纹理的属性GPU在访问纹理存储器时能够自动高效地处理边界问题。 The present invention employs a texture memory (texture memory) stores the extracted image data, by setting the GPU texture properties can be efficiently handled automatically Boundary when accessing the texture memory. [0106] 步骤S202,GPU按预设低尺度处理被提取图像,获得所述被提取图像对应的低尺度图像,在所述低尺度图像上检测Harris角点,以所述Harris角点组成候选特征点集合, 将所述候选特征点集合存储到GPU的全局存储器。 [0106] Step S202, GPU processing according to a preset low-scale image is extracted, obtaining the low-scale image corresponding to the image is extracted, the Harris corner detection on the low scale image to the Harris corner point candidates composition wherein set point, the set of candidate feature points stored in the global memory to the GPU.

[0107] 所述步骤S202进一步为: [0107] The step S202 further as:

[0108] 步骤S2021,GPU根据所述模板计算所述被提取图像和以所述预设低尺度为标准差的高斯函数的卷积,经卷积后的被提取图像存放于GPU的纹理存储器; [0108] Step S2021, the extracted image is calculated GPU and dimension to the predetermined low standard deviation for the Gaussian convolution function according to the template, it is extracted after convolution of the image stored in the GPU texture memory;

[0109] 步骤S2022,GPU计算经卷积后的被提取图像在水平方向和垂直方向的差分,差分结果存放于GPU的纹理存储器; [0109] Step S2022, the extracted differential image in the horizontal direction and the vertical direction after GPU computing the convolution, differential results stored in the GPU texture memory;

[0110] 步骤S2023,GPU将差分结果进行高斯模糊,经高斯模糊处理后的差分结果存放于GPU的全局存储器; [0110] Step S2023, the differential results GPU Gaussian blur, the differential result by a Gaussian blur processing is stored in the GPU global memory;

[0111] 步骤S2024,GPU根据差分结果计算Harris角点响应值,将Harris角点响应值存放于GPU的纹理存储器;根据所述Harris角点响应值判断所述被提取图像中的像素点是否为Harris角点,确定Harris角点的位置,将所述Harris角点的位置存放于GPU的全局存储器。 [0111] Step S2024, the differential calculation result GPU Harris corner response value, the Harris corner response value stored in the texture memory of the GPU; is extracted according to the Harris corner response value of the image pixels is determined whether the Harris corner point, determining the position of the Harris corner point, the position of the Harris corner point stored in the global memory of the GPU.

[0112] 实施例 [0112] Example

[0113] 步骤al,计算图像和高斯函数的卷积,计算结果存放于GPU的纹理存储器;为了加速计算卷积,先在原图数据I (χ,y)上计算χ方向的卷积得到结果gx(x,y),然后在gx(x,y) 的基础上计算y方向的卷积。 [0113] Step Al, is calculated convolved image and a Gaussian function, the calculation result stored in the GPU texture memory; To accelerate the convolution calculation, [chi] direction is first calculated on the original data I (χ, y) gx convolution results (x, y), and the y-direction is calculated on the basis of gx (x, y) of the convolution.

[0114] 步骤bl,计算高斯卷积后的被提取图像在水平方向和垂直方向的差分,计算结果存放于GPU的纹理存储器。 [0114] Step BL, the difference image is extracted in the horizontal and vertical directions is calculated after the Gaussian convolution calculation results stored in the GPU texture memory.

[0115] 步骤Cl,对差分结果进行高斯模糊,计算结果存放于GPU的全局存储器。 [0115] Step Cl, the results of the differential Gaussian blur, the calculation result stored in the GPU global memory.

[0116] 步骤dl,计算Harris角点响应值,计算结果存放于GPU的纹理存储器。 [0116] Step DL, Harris corner response value is calculated, the calculation result stored in the GPU texture memory.

[0117] 步骤el,确定Harris角点的位置,计算结果存放于GPU的全局存储器。 [0117] Step EL, determining the position of the Harris corner calculated results stored in the GPU global memory.

[0118] 步骤S203,GPU按多个预设高尺度处理被提取图像,获得所述被提取图像对应的多个高尺度图像,对所述候选特征点集合中的每一个候选特征点,计算所述候选特征点在所述多个高尺度图像下Hessian矩阵的DET值,根据所述DET值去除所述候选特征点集合中的误检点,获得所述被提取图像的特征点,将所述特征点存储到所述GPU的全局存储器。 [0118] Step S203, GPU processing by the plurality of preset high scales the extracted image to obtain a plurality of said extracted high-scale image corresponding to the image, for each of the candidate feature points in the candidate set, calculating DET value of said candidate feature points in the Hessian matrix of the plurality of high-scale image, removing erroneous behave in the candidate feature point set based on the value DET, to obtain the feature point image is extracted, the feature point of the GPU memory to global memory.

[0119] 所述步骤S203进一步为, [0119] The step S203 is further,

[0120] 步骤2031,GPU按多个预设高尺度对被提取图像中候选特征点和候选特征点的邻近区域进行高尺度模拟,获得所述被提取图像对应的多个高尺度图像,对于所述每一个候选特征点,计算所述候选特征点在所述被提取图像对应的每个高尺度图像下Hessian矩阵的DET值,将所述DET值存放于GPU的纹理存储器。 [0120] Step 2031, GPU according to a plurality of preset high scales the extracted image feature point candidate and the candidate feature point neighboring region of high scale simulation, obtaining a plurality of said extracted high-scale image corresponding to the image, for the said each candidate feature points, calculating the candidate feature points are extracted for each value DET high scale image corresponding to the image in the Hessian matrix, the value of DET is stored in the GPU texture memory.

[0121] 步骤2032,GPU判断如果所述候选特征点的DET值在所述多个高尺度图像的构成的尺度空间中没有极大值或在所述多个高尺度图像的构成的尺度空间中极大值小于预设阈值,则所述候选特征点为误检点,将所述候选特征点从所述候选特征点集合中除去;将所述候选特征点集合中剩余的候选特征点作为所述被提取图像的特征点存放于GPU的全局存储器。 [0121] Step 2032, GPU DET determines if the value of the feature point candidate is not the maximum value of the plurality of scale space formed in a high image scale or the scale-space constituting the plurality of high-scale image maximum value less than a preset threshold value, the feature point candidate behave error, the candidate feature point is removed from the set of candidate feature points; the feature point set as the candidate feature points remaining candidate extracting the image feature points to be stored in the GPU global memory.

[0122] 实施例 [0122] Example

[0123] 步骤a2,重新设置GPU计算内核。 [0123] Step a2, reset the GPU compute kernel. [0124] 针对候选特征点计算DET值,需要重新启动不同大小的内核。 [0124] DET calculated for the candidate value of the feature points need to restart kernels of different sizes. 假设有k个候选特征点,则需要启动k/64个线程块,每块有64个线程。 Consider k candidate feature points, it is necessary to start k / 64 thread blocks, each block having 64 threads.

[0125] 步骤1^2,对每个候选特征点,计算其在多个高尺度图像下Hessian矩阵的DET值, 计算结果存放于GPU的纹理存储器。 [0125] Step 1 ^ 2 for each candidate feature points, DET values ​​calculated at the plurality of high-scale image of the Hessian matrix, the calculation result stored in the GPU texture memory.

[0126] 步骤c2,根据Hessian矩阵的DET值,去除误检点并精确确定特征点的尺度,计算结果存放于GPU的全局存储器,用于计算特征点的SIFT特征和OM顺序度量特征。 [0126] Step c2, the value of Hessian matrix according DET, and removing erroneous behave precisely determining the dimension of the feature points, calculation results stored in the GPU global memory, and SIFT features for the feature point is calculated sequentially OM metric features.

[0127] 基于GPU的具有尺度不变性的图像局部特征点提取算法在GPU上的执行模式如图4所示,首先主程序把图像数据拷入纹理存储器,把模板数据拷入常量存储器(shared memory);然后各个线程按照编号读取相应的数据处理执行,中间结果放在共享存储器,各个线程依此通信协调;之后各线程继续执行直到执行结束把结果写入全局存储器(global memory);最后结果传入内存(DRAM),主程序继续执行。 [0127] Algorithm execution mode based on the GPU extracting local feature point scale invariant image GPU shown in Figure 4, first, the main image data into a texture memory, the template data into a constant memory (shared memory ); then reads the corresponding number of each thread in accordance with the process execution data, intermediate results in shared memory, each thread so communication coordinator; after each thread to continue until the end of the execution result is written to global memory (global memory); final outcome incoming memory (DRAM), the main program to continue.

[0128] 本发明应用所述的图像特征点提取方法的图像拷贝检测方法如图5所示。 [0128] The method of detecting copy of the image feature points extraction method applied to the image according to the present invention shown in FIG.

[0129] 步骤S501,按所述图像特征点提取方法提取被检测图像的图像特征点,计算每一个所述图像特征点的SIFTGcale invariant feature transform,尺度不变特征变换)特征和OM(Ordinal measure,顺序量度)顺序度量特征。 [0129] step S501, the image by the feature point extraction method of extracting the image feature points are detected in the image, calculates SIFTGcale invariant feature transform each of the image feature points, scale invariant feature transform) features and OM (Ordinal measure, order metric) measurement sequence features.

[0130] 步骤S502,对图像库中的每一幅图像,按所述图像特征点提取方法提取所有图像的图像特征点,计算每一个所述图像特征点的SIFT特征和OM顺序度量特征。 [0130] step S502, the image library on each image, the image by the feature point extraction method of extracting the image feature points of all the images, and calculates OM SIFT features of the image sequence of each feature point feature measure.

[0131] 步骤S503,根据所述被检测图像和图像库中图像的SIFT特征和OM顺序度量特征进行拷贝检测。 [0131] step S503, the copy detection metric based on the features detected SIFT and OM sequential images and images in the image library.

[0132] 所述步骤S502进一步为, [0132] The step S502 is further,

[0133] 步骤5021,对图像库中的每一幅图像,按所述图像特征点提取方法提取所有图像的图像特征点,计算所述图像特征点的SIFT特征和OM顺序度量特征;将图像库中所有图像特征点的SIFT特征建立为ANN索引树,将图像库中所有图像特征点的OM顺序度量特征以文件的方式存放。 [0133] Step 5021, the image of each image in the library, according to the image feature point extracting method for extracting the image feature points of all the images, the SIFT features and OM sequentially calculates the image feature point feature measure; image library SIFT features all image feature points ANN index tree is established, the order will measure all characteristics of OM image feature point in the image library by way of file storage.

[0134] 所述步骤S503进一步为, [0134] The step S503 is further,

[0135] 步骤5031,以被检测图像的每个特征点为查询点,根据所述查询点的SIFT特征进行初始ANN检索,获得初始检测结果,初始检测结果为图像库中特征点,所述初始检测结果的SIFT特征同查询点的SIFT特征的距离小于第二预设阈值。 [0135] Step 5031, each feature point is detected in the image as a query point, initial ANN retrieves The SIFT feature points of the query to obtain the initial test results, the initial detection result of the feature point image database, the initial SIFT feature detection results from the query point with SIFT feature is less than a second preset threshold.

[0136] 步骤5032,去除掉所述初始检测结果中OM顺序度量特征同所述查询点的OM顺序度量特征不一致的特征点,所述初始检测结果中未被去除的特征点为同所述查询点最匹配的特征点,所述最匹配的特征点组成集合。 [0136] Step 5032, to remove the initial detection result order OM metric measure characteristics inconsistent with the feature point feature OM order of the query point, the initial feature point detection result is not removed with the query point characteristic point matching, the composition of the set of best matches feature points.

[0137] 步骤5033,对所述集合中的特征点所属的所有图像,以投票的方式决定所述图像的排名,排名为前η的图像为所述被检测图像的相似图像,η为预设值。 [0137] Step 5033, all of the image feature points in the set belongs, by voting determines the ranking of the image, rank [eta] is a front image of the similar image is detected image, [eta] is a preset value.

[0138] 本发明图像特征点提取装置,如图6所示。 [0138] The present invention, image feature extraction means shown in Fig.

[0139] 候选特征点集合获得模块601,用于在被提取图像对应的低尺度图像上检测Harris角点,获得候选特征点集合。 [0139] candidate feature point set acquisition module 601, a Harris corner point detection on the low scale image corresponding to the image is extracted, the candidate feature point set is obtained.

[0140] 特征点提取模块602,用于对所述候选特征点集合中的每一个候选特征点,计算所述候选特征点在所述被提取图像对应的多个高尺度图像下Hessian矩阵的DET值,根据所述DET值去除所述候选特征点集合中的误检点,获得所述被提取图像的特征点。 DET Hessian matrix of the plurality of high-scale image [0140] feature point extraction module 602, a candidate for each of the feature points in the candidate set calculating the candidate feature points are extracted in the image corresponding to value according to the value of DET behave removing the bad set of candidate feature points, said feature points are extracted to obtain an image. [0141] 一个优选实施方式中,候选特征点集合获得模块601进一步用于在被提取图像对应的低尺度图像上检测Harris角点,所述低尺度图像是由被提取图像和高斯函数做卷积获得,高斯函数的标准差为所述低尺度图像的尺度;将所述Harris角点中在同一像素位置重复出现的角点去除,剩余角点组成所述候选特征点集合。 [0141] In a preferred embodiment, the candidate feature point set acquisition module 601 is further configured to extract a low-scale image corresponding to the image is the Harris corner detector, the low-scale image is done by a convolution of the extracted image and a Gaussian function obtained scale-scale image to the low standard deviation of the Gaussian function; Harris corner point of the corner points at the same pixel position recurring removed, the composition of the remaining candidate corners feature point set.

[0142] 一个优选实施方式中,特征点提取模块602进一步用于对于所述每一个候选特征点,计算所述候选特征点在所述被提取图像对应的每个高尺度图像下Hessian矩阵的DET 值;对于所述每一个候选特征点,如果所述候选特征点的DET值在所述多个高尺度图像构成的尺度空间中没有极大值或极大值小于预设阈值,则所述候选特征点为误检点,将所述候选特征点从所述候选特征点集合中除去;所述候选特征点集合中未被除去的候选特征点为所述被提取图像的特征点。 [0142] In a preferred embodiment, the feature point extraction module 602 is further used for each of the feature point candidate, the candidate is calculated at each of the feature points are extracted image corresponding to a high-scale image of the Hessian matrix of DET value; for each of the candidate feature points, the feature point if the candidate is not the maximum value of DET scale space in the plurality of high-scale image consisting of the maximum value or less than a preset threshold value, the candidate erroneous discreet feature point, the feature point candidate is removed from the set of candidate feature points; the candidate feature point set not removed candidate feature points are extracted as the feature point image.

[0143] 一个优选实施方式中,特征点提取模块602,还用于如果所述候选特征点的DET值在所述多个高尺度图像构成的尺度空间中有极大值,并且所述极大值大于所述预设阈值, 则确认所述候选特征点为所述被提取图像的特征点,所述特征点的尺度为所述极值对应的高尺度图像的尺度。 [0143] In a preferred embodiment, the feature point extraction module 602, for further if the candidate feature point DET value has a maximum value in the scale space image composed of the plurality of high standards, and the great value is greater than the preset threshold value, confirming that the candidate feature points are extracted as characteristic points of the image, the scale of the feature point of the high scale-scale image corresponding to the extreme value.

[0144] 一种图像拷贝检测系统如图7所示,包括: [0144] A copy of the image detecting system shown in Figure 7, comprising:

[0145] 所述图像特征点提取装置701, [0145] The image feature extraction means 701,

[0146] 被检测图像特征生成模块702,用于使用所述图像特征点提取装置701提取被检测图像的图像特征点,对每一个图像特征点,计算每一个所述图像特征点的SIFT特征和OM 顺序度量特征; [0146] wherein the detected image generation module 702, 701 for extracting image feature point detection image using the image feature point extracting means, for each image feature points to calculate each of SIFT features of the image feature point and OM order of metric characteristics;

[0147] 图像库特征生成模块703,用于对图像库中的每一幅图像,使用所述图像特征点提取装置701提取所有图像的图像特征点; [0147] The image feature database generating module 703, the image database for each image, using the image feature point extracting device 701 extracts the image feature points of all the images;

[0148] 拷贝检测模块704,用于根据所述SIFT特征和OM顺序度量特征进行拷贝检测。 [0148] a copy detection module 704 for detecting copy feature measure based on the order of SIFT and OM.

[0149] 一个优选实施方式中,图像库特征生成模块703进一步用于对图像库中的每一幅图像,使用所述图像特征点提取装置701提取所有图像的图像特征点,计算所述图像特征点的SIFT特征和OM顺序度量特征;将图像库中所有图像特征点的SIFT特征建立为ANN索引树,将图像库中所有图像特征点的OM顺序度量特征以文件的方式存放。 [0149] In a preferred embodiment, the image feature database generating module 703 is further configured to image library each image, using the image feature point extracting device 701 extracts the image feature points of all the images, calculating the image feature OM and SIFT features characteristic measure point sequence; SIFT feature all the image feature points in the image database is established ANN index tree, the order of all the metric OM wherein the image feature point in the image library stored as files.

[0150] 拷贝检测模块704进一步用于以被检测图像的每个特征点为查询点,根据所述查询点的SIFT特征进行初始ANN检索,获得初始检测结果,初始检测结果为图像库中特征点, 所述初始检测结果的SIFT特征同查询点的SIFT特征的距离小于第二预设阈值;去除掉所述初始检测结果中OM顺序度量特征同所述查询点的OM顺序度量特征不一致的特征点,所述初始检测结果中未被去除的特征点为同所述查询点最匹配的特征点,所述最匹配的特征点组成集合;对所述集合中的特征点所属的所有图像,以投票的方式决定所述图像的排名, 排名为前η的图像为所述被检测图像的相似图像,η为预设值。 [0150] module 704 is further configured to copy detection feature points are detected in each image as a query point, initial ANN retrieves The SIFT feature points of the query to obtain the initial test results, the initial detection result of the image feature points in the library , from the query point with SIFT feature detection results of the initial SIFT feature is less than a second predetermined threshold value; to get rid of the initial order of the detection result OM metric measure characteristics inconsistent with the characteristic feature points of the query point in order OM , the initial feature point detection result is not removed with the point of best match the query feature points, said feature points best match to form a set; all images in the set of feature points belong to vote the image of the decided rank, rank [eta] is a front image as the similar image is detected image, [eta] is a preset value.

[0151] 拷贝检测模块704首先根据查询点q的SIFT特征进行ANN检索,得到初始结果集合S' ;然后去除S'中那些OM特征和q不一致的点,即得到与q最相似的点的集合,对被检测图像I的每一个特征点进行特征点匹配,得到与该点最相似的点集S',I的所有特征点的匹配结果集合为S = { US' }。 [0151] a copy detection module 704 is first carried out according to the SIFT feature query point q ANN search to obtain an initial result set S '; then removing S' those OM features and q inconsistent point, i.e., to obtain the q most similar to the set of points , for each detected feature points are feature points of the image I matching point to obtain the most similar to the set of points S ', all the feature points of the matching result set is I S = {US'}. 对于S中的点所属的所有图像,以投票的方式决定图像的排名,排名最靠前的图像就是和I最相似的图像。 For all images S in the point belongs, determined by voting ranking images, the highest ranked images and I was most similar image.

[0152] 本领域的技术人员在不脱离权利要求书确定的本发明的精神和范围的条件下,还可以对以上内容进行各种各样的修改。 [0152] Those skilled in the art without departing from the spirit and scope of the claims of the invention being determined, various modifications may also be about the above. 因此本发明的范围并不仅限于以上的说明,而是由权利要求书的范围来确定的。 The scope of the present invention is not limited to the above description, but rather determined by the scope of the claimed requirements.

Claims (15)

1. 一种图像特征点提取方法,其特征在于,包括:步骤1,在被提取图像对应的低尺度图像上检测Harris角点,获得候选特征点集合; 步骤2,对所述候选特征点集合中的每一个候选特征点,计算所述候选特征点在所述被提取图像对应的多个高尺度图像下Hessian矩阵的DET值,如果所述候选特征点的DET值在所述多个高尺度图像构成的尺度空间中没有极大值或所述极大值小于预设阈值,则所述候选特征点为误检点,将所述候选特征点从所述候选特征点集合中除去;所述候选特征点集合中剩余的候选特征点为被提取图像的特征点。 An image feature point extracting method, comprising the steps of: 1, is detected on the low scale image corresponding to the image is extracted Harris corner, to obtain a set of candidate feature points; Step 2, the set of candidate feature points each candidate feature points, calculating the candidate feature points are extracted DET plurality of high value scale image corresponding to the image of the Hessian matrix, if the candidate feature point in the plurality of high value DET scale scale space image formed no maximum value or the maximum value less than a preset threshold value, the feature point candidate behave error, the candidate feature point is removed from the set of candidate feature points; the candidate feature point set in the remaining candidate feature points of the image feature points are extracted.
2.如权利要求1所述的图像特征点提取方法,其特征在于, 所述步骤1进一步为,步骤21,在被提取图像对应的低尺度图像上检测Harris角点,所述低尺度图像是由被提取图像和高斯函数做卷积获得,高斯函数的标准差为所述低尺度图像的尺度;步骤22,将所述Harris角点中在同一像素位置重复出现的角点去除,剩余角点组成所述候选特征点集合。 The image feature points extraction method according to claim 1, wherein said step is a further step 21, the Harris corner detection on the low scale image corresponding to the extracted image, the image is a low-scale is extracted from the image and a Gaussian function convolve obtained, the standard deviation for the Gaussian function mesoscale dimension of the image; step 22, the Harris corner removed at the same pixel position corner repeated, the remaining corner the composition of the candidate feature point set.
3.如权利要求1所述的图像特征点提取方法,其特征在于, 所述步骤2进一步为,步骤31,对于所述每一个候选特征点,计算所述候选特征点在所述被提取图像对应的每个高尺度图像下Hessian矩阵的DET值;步骤32,对于所述每一个候选特征点,如果所述候选特征点的DET值在所述多个高尺度图像构成的尺度空间中没有极大值或极大值小于预设阈值,则所述候选特征点为误检点,将所述候选特征点从所述候选特征点集合中除去;步骤33,所述候选特征点集合中未被除去的候选特征点为所述被提取图像的特征点。 The image feature points extraction method according to claim 1, wherein said step of further 2, step 31, for each of the candidate feature points, calculating the candidate feature points are extracted in the image DET value of the Hessian matrix at each corresponding to a high-scale image; step 32, for each of the feature point candidate, if the candidate feature point DET value is not very high in the scale space-scale image of the plurality of configuration large or maximum value is less than a predetermined threshold, then the candidate feature points behave as erroneous, the candidate feature point is removed from the set of candidate feature points; step 33, the candidate feature point set not removed the feature point candidate is extracted as the feature point of the image.
4.如权利要求3所述的图像特征点提取方法,其特征在于, 所述步骤2还包括,步骤41,如果所述候选特征点的DET值在所述多个高尺度图像构成的尺度空间中有极大值,并且所述极大值大于所述预设阈值,则所述候选特征点为所述被提取图像的特征点, 所述特征点的尺度为所述极大值对应的高尺度图像的尺度。 The image feature points extraction method according to claim 3, wherein said step 2 further comprises the step 41, if the candidate feature points in scale space DET value of the plurality of high-scale image composed of has maximum value and the maximum value is greater than the preset threshold value, the feature point candidate is extracted as the feature point of the image, the scale of the feature point corresponding to the maximum value of the high scale scale image.
5. 一种在GPU上实现图像特征点提取的方法,其特征在于,包括:步骤51,根据被提取图像的大小配置GPU计算内核,将所述被提取图像的数据拷贝到GPU的纹理存储器;步骤52,GPU按预设低尺度处理被提取图像,获得所述被提取图像对应的低尺度图像, 在所述低尺度图像上检测Harris角点,以所述Harris角点组成候选特征点集合,将所述候选特征点集合存储到GPU的全局存储器;步骤53,GPU按多个预设高尺度处理被提取图像,获得所述被提取图像对应的多个高尺度图像,对所述候选特征点集合中的每一个候选特征点,计算所述候选特征点在所述多个高尺度图像下Hessian矩阵的DET值,如果所述候选特征点的DET值在所述多个高尺度图像构成的尺度空间中没有极大值或所述极大值小于预设阈值,则所述候选特征点为误检点,将所述候选特征点从所述候选特征点集合中 An image feature point extraction method implemented in GPU, which is characterized in that, comprising: Step 51, compute kernel GPU configuration according to the size of the image is extracted, the extracted image data copied to the GPU texture memory; step 52, GPU processing according to a preset low-scale image is extracted, obtaining the low-scale image corresponding to the image is extracted, the Harris corner detection on the low scale image to the composition of the Harris corner candidate feature point set, the candidate feature point set stored in the GPU global memory; step 53 is, GPU processing by the plurality of preset high scales the extracted image to obtain a plurality of said extracted high-scale image corresponding to the image, the feature point candidate each set of candidate feature points, the calculated value of DET candidate feature points in said plurality of high-scale image of the Hessian matrix, if the scale of the feature point candidate value of the plurality of high-DET-scale image composed of space is no maximum value or the maximum value less than a preset threshold value, the feature point candidate erroneous discreet, the feature point candidate from the candidate set of feature points 去;所述候选特征点集合中剩余的候选特征点为被提取图像的特征点,将所述特征点存储到所述GPU的全局存储器。 To; the candidate feature point set in the remaining candidate feature points are extracted image feature points, the feature point stored to the GPU global memory.
6.如权利要求5所述的在GPU上实现图像特征点提取的方法,其特征在于,所述步骤51中还包括,步骤61,将用于计算卷积和差分的模板数据拷贝到GPU的常量存储器; 所述步骤52进一步为,步骤62,GPU根据所述模板计算所述被提取图像和以所述预设低尺度为标准差的高斯函数的卷积,经卷积后的被提取图像存放于GPU的纹理存储器;步骤63,GPU计算经卷积后的被提取图像在水平方向和垂直方向的差分,差分结果存放于GPU的纹理存储器;步骤64,GPU将差分结果进行高斯模糊,经高斯模糊处理后的差分结果存放于GPU的全局存储器;步骤65,GPU根据差分结果计算Harris角点响应值,将Harris角点响应值存放于GPU的纹理存储器;根据所述Harris角点响应值判断所述被提取图像中的像素点是否为Harris角点,确定Harris角点的位置,将所述Harris角点的位置坐标存放于GPU的全局存储 6. A method as claimed in the image feature points extracted implemented on the GPU. 5, characterized in that, further comprising the step of 51, step 61, calculating the convolution for copying the template and difference data to the GPU constant memory; further step 52 is step 62, GPU is calculated based on the extracted image and the template at the predetermined low standard deviation for the dimension convolution of a Gaussian function, it is extracted by the image convolved stored in the GPU texture memory; step 63, the extracted image convolved GPU computing a difference, the difference results in a horizontal direction and the vertical direction stored in the GPU texture memory; step 64, the GPU Gaussian blur difference results by Gaussian blur difference processing result stored in the GPU global memory; step 65, GPU Harris corner response value is calculated based on the difference result, the Harris corner response value stored in the GPU texture memory; according to the Harris corner response value determination the extracted image pixels whether the Harris corner point, determining the position of the Harris corner, the position coordinates of the Harris corner point stored in global memory GPU .
7.如权利要求5所述的在GPU上实现图像特征点提取的方法,其特征在于, 所述步骤53进一步为,步骤71,GPU按多个预设高尺度对被提取图像中候选特征点和候选特征点的相邻区域进行高尺度模拟,获得候选特征点和相邻区域在多个高尺度图像对应区域,对于所述每一个候选特征点,计算所述候选特征点在所述被提取图像对应的每个高尺度图像下Hessian 矩阵的DET值,将所述DET值存放于GPU的纹理存储器;步骤72,对于每个候选特征点,GPU判断如果所述候选特征点的DET值在所述多个高尺度图像构成的尺度空间中没有极大值或极大值小于预设阈值,则所述候选特征点为误检点,将所述候选特征点从所述候选特征点集合中去除;将所述候选特征点集合中剩余的候选特征点作为所述被提取图像的特征点存放于GPU的全局存储器。 7. The method as claimed in the image feature points extracted implemented on the GPU 5, characterized in that, further to the step 53, step 71, GPU according to a plurality of preset high scales the extracted image feature point candidate and adjacent regions of the candidate feature points higher scale simulation, to obtain the candidate feature points and the plurality of adjacent regions in a corresponding region of high scale image, for each of the candidate feature points, calculating the candidate feature points are extracted in the DET value of Hessian matrix at each high-scale image corresponding to the image, the value of DET is stored in the GPU texture memory; step 72, for each candidate feature points, determining if the GPU DET value candidate feature point in the scale space of said plurality of high-scale image composed of no less than a maximum value or a preset maximum threshold value, then the candidate feature points behave as erroneous, the candidate feature point is removed from the set of candidate feature points; the feature point set of candidate feature points remaining candidate is extracted as the image feature points stored in the GPU global memory.
8. 一种应用如权利要求1所述的图像特征点提取方法的图像拷贝检测方法,其特征在于,包括:步骤81,按所述图像特征点提取方法提取被检测图像的图像特征点,计算每一个所述图像特征点的SIFT特征和OM顺序度量特征;步骤82,对图像库中的每一幅图像,按所述图像特征点提取方法提取所有图像的图像特征点,计算所述图像特征点的SIFT特征和OM顺序度量特征;步骤83,根据所述被检测图像和所述图像库中图像的SIFT特征和OM顺序度量特征进行拷贝检测。 Copy of the image detection image feature points extraction method A use as claimed in claim 1, characterized by comprising: a step 81, according to the image feature point extracting method for extracting an image feature point detection image is calculated OM and SIFT features of the image sequence of each feature point feature measure; step 82, the image library on each image, the image by the feature point extraction method of extracting the image feature points of all the images, calculating the image feature OM and SIFT features characteristic measurement point sequence; step 83, a measure for the copy detection feature is based on the detection order SIFT and OM image and the image in the image library.
9.如权利要求8所述的图像拷贝检测方法,其特征在于, 所述步骤82进一步为,步骤91,对图像库中的每一幅图像,按所述图像特征点提取方法提取所有图像的图像特征点,计算所述图像特征点的SIFT特征和OM顺序度量特征;将图像库中所有图像特征点的SIFT特征建立为ANN索引树,将图像库中所有图像特征点的OM顺序度量特征以文件的方式存放;所述步骤83进一步为,步骤92,以被检测图像的每个特征点为查询点,根据所述查询点的SIFT特征进行初始ANN检索,获得初始检测结果,初始检测结果为图像库中图像的图像特征点,所述初始检测结果的SIFT特征同查询点的SIFT特征的距离小于第二预设阈值;步骤93,去除掉所述初始检测结果中OM顺序度量特征同所述查询点的OM顺序度量特征不一致的特征点,所述初始检测结果中未被去除的特征点为同所述查询点最匹配的 The image copy detection method according to claim 8, wherein said step of further 82, step 91, the image library on each image, all the images extracted by the image feature point extraction method image feature points to calculate the image feature points in order of metrics SIFT and OM wherein; SIFT feature all the image feature points in the image database index tree established as ANN, the OM order to measure all of the image feature points in the image feature database file storage mode; the step 83 is a further step 92, for each feature point to be detected in the query image point, according to the initial ANN retrieves SIFT feature points of the query, obtaining the initial detection result, the detection result of the initial image database the image feature points of the image, the distance with the query point SIFT feature detection results of the initial SIFT feature is less than the second preset threshold; step 93, to remove the initial detection result with OM wherein said metric sequence OM order query feature points feature points inconsistent measure, the initial feature point detection result is not removed with the best match of the query point 征点,所述最匹配的特征点组成集合;步骤94,对所述集合中的特征点所属的所有图像,以投票的方式决定所述图像的排名, 排名为前η的图像为所述被检测图像的相似图像,η为预设值。 Of feature points, said feature points best match to form a set; step 94, all the image feature point in the set belongs, by voting determines the ranking of the image, for the first ranking to the image to be η similar image detection image, η is a preset value.
10. 一种图像特征点提取装置,其特征在于,包括:候选特征点集合获得模块,用于在被提取图像对应的低尺度图像上检测Harris角点, 获得候选特征点集合;特征点提取模块,用于对所述候选特征点集合中的每一个候选特征点,计算所述候选特征点在所述被提取图像对应的多个高尺度图像下Hessian矩阵的DET值,如果所述候选特征点的DET值在所述多个高尺度图像构成的尺度空间中没有极大值或所述极大值小于预设阈值,则所述候选特征点为误检点,将所述候选特征点从所述候选特征点集合中除去; 所述候选特征点集合中剩余的候选特征点为被提取图像的特征点。 An image feature point extracting means, wherein, comprising: a candidate feature point set acquisition module for detecting a Harris corner on the low scale image corresponding to the image is extracted, obtaining the set of candidate feature points; feature point extraction module for each candidate feature points in the candidate set, the candidate feature points calculated values ​​DET Hessian matrix at a plurality of high-scale image is extracted in the image corresponding to, if the candidate feature points the maximum value is not DET scale space in the plurality of high-scale image consisting of the maximum value or less than a preset threshold value, the feature point candidate behave error, the candidate feature points from the removing the set of candidate feature points; the candidate feature points remaining candidate set of feature points of the image feature points are extracted.
11.如权利要求10所述的图像特征点提取装置,其特征在于,所述候选特征点集合获得模块进一步用于在被提取图像对应的低尺度图像上检测Harris角点,所述低尺度图像是由被提取图像和高斯函数做卷积获得,高斯函数的标准差为所述低尺度图像的尺度;将所述Harris角点中在同一像素位置重复出现的角点去除,剩余角点组成所述候选特征点集合。 The image of the feature point extraction device of claim 10, wherein the candidate feature point set acquisition module is further for detecting a Harris corner on the low scale image corresponding to the extracted image, the low image scale It is obtained from the extracted image and a Gaussian convolution to do, for the standard deviation of the Gaussian function mesoscale dimension of the image; Harris corner point of the corner points at the same pixel position recurring removed, the composition of the remaining corners said candidate feature point set.
12.如权利要求10所述的图像特征点提取装置,其特征在于,所述特征点提取模块进一步用于对于所述每一个候选特征点,计算所述候选特征点在所述被提取图像对应的每个高尺度图像下Hessian矩阵的DET值;对于所述每一个候选特征点,如果所述候选特征点的DET值在所述多个高尺度图像构成的尺度空间中没有极大值或极大值小于预设阈值,则所述候选特征点为误检点,将所述候选特征点从所述候选特征点集合中除去;所述候选特征点集合中未被除去的候选特征点为所述被提取图像的特征点。 The image of the feature point extraction device of claim 10, wherein said feature extraction module is further configured to point to the each candidate feature points, calculating the candidate feature points are extracted in the image corresponding DET value of the Hessian matrix at each higher scale images; for each of the candidate feature points, the feature point if the candidate value of DET no maximum value in scale space or very high-scale image of the plurality of configuration is greater than the preset threshold value, the feature point candidate behave error, the candidate feature point is removed from the set of candidate feature points; the candidate feature point set not removed as the candidate feature points the extracted feature points of the image.
13.如权利要求12所述的图像特征点提取装置,其特征在于,所述特征点提取模块还用于如果所述候选特征点的DET值在所述多个高尺度图像构成的尺度空间中有极大值,并且所述极大值大于所述预设阈值,则确认所述候选特征点为所述被提取图像的特征点,所述特征点的尺度为所述极大值对应的高尺度图像的尺度。 The image feature points 12 if the candidate feature point scale space values ​​DET-scale image of the plurality of high-configured extraction device in claims, characterized in that the feature point extraction module is further configured to has a maximum value, and the maximum value is greater than the preset threshold value, it is confirmed that the candidate feature point the feature point extracted image, the scale of the feature point corresponding to the maximum value of the high scale scale image.
14. 一种图像拷贝检测系统,其特征在于,包括:如权利要求10所述的图像特征点提取装置,被检测图像特征生成模块,用于使用所述图像特征点提取装置提取被检测图像的图像特征点,计算每一个所述图像特征点的SIFT特征和OM顺序度量特征;图像库特征生成模块,用于对图像库中的每一幅图像,使用所述图像特征点提取装置提取所有图像的图像特征点,计算所述图像特征点的SIFT特征和OM顺序度量特征;拷贝检测模块,用于根据所述SIFT特征和OM顺序度量特征进行拷贝检测。 14. A copy of the image detection system comprising: the image feature points extraction device as claimed in claim 10, wherein the image detection module is generated, using the image feature point extracting means extracts the detected image image feature points, and calculates OM SIFT features of the image sequence of each feature point feature measure; image feature database generating module, for each image in an image database, using the image feature point extraction means extracts all images image feature points to calculate the image feature points and OM SIFT features characteristic measurement sequence; copy detection means for detecting copying metric features according to the order of SIFT and OM.
15.如权利要求14所述的图像拷贝检测系统,其特征在于,所述图像库特征生成模块进一步用于对图像库中的每一幅图像,使用所述图像特征点提取装置提取所有图像的图像特征点,计算所述图像特征点的SIFT特征和OM顺序度量特征;将图像库中所有图像特征点的SIFT特征建立为ANN索引树,将图像库中所有图像特征点的OM顺序度量特征以文件的方式存放; 所述拷贝检测模块进一步用于以被检测图像的每个特征点为查询点,根据所述查询点的SIFT特征进行初始ANN检索,获得初始检测结果,初始检测结果为图像库中图像的图像特征点,所述初始检测结果的SIFT特征同查询点的SIFT特征的距离小于第二预设阈值;去除掉所述初始检测结果中OM顺序度量特征同所述查询点的OM顺序度量特征不一致的特征点,所述初始检测结果中未被去除的特征点为同所述查询点最匹配的 15. A copy of the image detection system according to claim 14, wherein the image feature database generating module is further configured to image library each image, using the image feature point extraction means extracts all images image feature points to calculate the image feature points in order of metrics SIFT and OM wherein; SIFT feature all the image feature points in the image database index tree established as ANN, the OM order to measure all of the image feature points in the image feature database file storage mode; the detection module is further configured to copy each feature point is detected as a query image point, according to the initial ANN retrieves SIFT feature points of the query to obtain the initial test results, the initial detection result of the image library image feature points in the image, the detection result of the initial SIFT features SIFT feature from the query point with less than a second predetermined threshold; order to remove the measure OM OM wherein the detection result with the initial order of the query point measure inconsistent feature point feature, the initial feature point detection result is not removed with the best match of the query point 特征点,所述最匹配的特征点组成集合;对所述集合中的特征点所属的所有图像,以投票的方式决定所述图像的排名,排名为前η的图像为所述被检测图像的相似图像,η为预设值。 Feature points to form a set of the best matches feature points; all images in the set of feature points belong, the ranking determined by voting of the image, rank the front η image is detected according to the image similar image, η is a preset value.
CN 200910237737 2009-11-23 2009-11-23 And the image feature point extracting method for implementing an image copy detection method and system CN101719275B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN 200910237737 CN101719275B (en) 2009-11-23 2009-11-23 And the image feature point extracting method for implementing an image copy detection method and system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN 200910237737 CN101719275B (en) 2009-11-23 2009-11-23 And the image feature point extracting method for implementing an image copy detection method and system

Publications (2)

Publication Number Publication Date
CN101719275A CN101719275A (en) 2010-06-02
CN101719275B true CN101719275B (en) 2011-12-14

Family

ID=42433845

Family Applications (1)

Application Number Title Priority Date Filing Date
CN 200910237737 CN101719275B (en) 2009-11-23 2009-11-23 And the image feature point extracting method for implementing an image copy detection method and system

Country Status (1)

Country Link
CN (1) CN101719275B (en)

Families Citing this family (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101944183B (en) * 2010-09-02 2013-12-25 北京航空航天大学 Method for identifying object by utilizing SIFT tree
CN101996325B (en) * 2010-09-08 2013-09-04 北京航空航天大学 Improved method for extracting characteristic point from image
CN102034243B (en) * 2010-12-29 2012-09-26 北京智安邦科技有限公司 Method and device for acquiring crowd density map from video image
CN102629325B (en) * 2012-03-13 2014-06-04 深圳大学 Image characteristic extraction method, device thereof, image copy detection method and system thereof
CN103530224A (en) * 2013-06-26 2014-01-22 郑州大学 Harris corner detecting software system based on GPU
CN103927709B (en) * 2014-03-19 2017-06-06 西安电子科技大学昆山创新研究院 Watermarking embedding and extracting methods reversible robust feature region based on optimized geometry
CN104091977B (en) * 2014-05-06 2016-06-15 无锡日联科技股份有限公司 The method of detecting the wound lithium ion battery
CN103996043A (en) * 2014-05-29 2014-08-20 重庆大学 Detection method of image invariant feature area
CN104331696B (en) * 2014-09-28 2017-10-31 苏州新视线文化科技发展有限公司 A rapid extraction method based on the image feature graphics processing unit
CN104484869B (en) * 2014-11-14 2017-02-22 中国科学院计算技术研究所 Measure feature for sorting image matching method and system
CN105550974A (en) * 2015-12-13 2016-05-04 复旦大学 GPU-based acceleration method of image feature extraction algorithm

Also Published As

Publication number Publication date
CN101719275A (en) 2010-06-02

Similar Documents

Publication Publication Date Title
Alcantarilla et al. KAZE features
Amerini et al. A sift-based forensic method for copy–move attack detection and transformation recovery
Aiger et al. 4-points congruent sets for robust pairwise surface registration
Al-Qershi et al. Passive detection of copy-move forgery in digital images: State-of-the-art
CN1278280C (en) Method and computer system for detecting image copy of contents
CN101133429B (en) Document/image searching method, and document/image recording and searching device
Yang et al. Local difference binary for ultrafast and distinctive feature description
CN1139039C (en) Method and apparatus for extraction and recognition and abnormality judge of pattern in image
Pan et al. Region duplication detection using image feature matching
Zheng et al. A robust method for eye features extraction on color image
Perfetti et al. Cellular neural networks with virtual template expansion for retinal vessel segmentation
Salti et al. A performance evaluation of 3d keypoint detectors
Tao et al. Manifold ranking-based matrix factorization for saliency detection
Farid Exposing digital forgeries in scientific images
Li et al. Automatic image stitching using SIFT
CN101697232B (en) SIFT characteristic reducing method facing close repeated image matching
CN102663391A (en) Image multifeature extraction and fusion method and system
JP6334735B2 (en) Image feature detection using edge vector
CN103116763B (en) Alive human face detection method based on statistical characteristics of the color space hsv
CN105160361A (en) Image identification method and apparatus
Yang et al. Copy-move forgery detection based on hybrid features
Su et al. Region segmentation in histopathological breast cancer images using deep convolutional neural network
Mohamadian et al. Detection of duplication forgery in digital images in uniform and non-uniform regions
CN101847163B (en) Design patent image retrieval method with multi-characteristics fusion
EP2534612B1 (en) Efficient scale-space extraction and description of interest points

Legal Events

Date Code Title Description
C06 Publication
C10 Request of examination as to substance
C14 Granted