CN112906710A - Visual image feature extraction method based on BAKAZE-MAGSAC - Google Patents

Visual image feature extraction method based on BAKAZE-MAGSAC Download PDF

Info

Publication number
CN112906710A
CN112906710A CN202110327237.0A CN202110327237A CN112906710A CN 112906710 A CN112906710 A CN 112906710A CN 202110327237 A CN202110327237 A CN 202110327237A CN 112906710 A CN112906710 A CN 112906710A
Authority
CN
China
Prior art keywords
image
point
determining
model
points
Prior art date
Legal status (The legal status 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 status listed.)
Pending
Application number
CN202110327237.0A
Other languages
Chinese (zh)
Inventor
胡燕祝
王松
贺琬婧
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing University of Posts and Telecommunications
Original Assignee
Beijing University of Posts and Telecommunications
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 Beijing University of Posts and Telecommunications filed Critical Beijing University of Posts and Telecommunications
Priority to CN202110327237.0A priority Critical patent/CN112906710A/en
Publication of CN112906710A publication Critical patent/CN112906710A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/30—Noise filtering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

The invention relates to a visual image feature extraction method based on BAKAZE-MAGSAC, which is a method for extracting features of disaster visual images, belongs to the field of image processing and machine learning, and is characterized by comprising the following steps: (1) determining a nonlinear diffusion filter; (2) determining a non-linear scale space of the image; (3) determining the main direction of the feature points; (4) matching the characteristic points; (5) determining a model quality function; (6) the probability of each point being an inlier is determined. The method effectively solves the problem that the number of extracted feature points is reduced because the feature points cannot be effectively extracted when the image has scale and fuzzy changes, greatly improves the influence of interference noise on the extraction of the feature points, and simultaneously improves the matching time efficiency and the matching precision. The method with high classification accuracy is provided for the field of feature extraction of disaster vision images.

Description

Visual image feature extraction method based on BAKAZE-MAGSAC
Technical Field
The invention relates to the field of image processing and machine learning, in particular to a method for extracting features of disaster vision images.
Background
At present, the extraction of image features mainly adopts fuzzy extraction of image contours, filters interference noise of images and then screens out key feature values. For feature extraction of an image, the conventional methods include image feature matching based on an AKAZE algorithm and image feature identification based on a Relief algorithm. For the image feature matching based on the AKAZE algorithm, because the binary descriptor is adopted to describe the feature points of the image, the extraction quantity of the feature points is unstable and sharply reduced when the scale and the blur of the image change, the accuracy of the feature points is reduced at the same time, and the deviation is easily generated. And by adopting the Relief algorithm, because the feature detection is carried out by adopting a differential frame scanning technology, the edge contour feature is marked and the ambiguity is identified, the nonlinear interference noise is difficult to inhibit, and the change of the image brightness in different scale spaces is difficult to describe. Due to the change of image dimensions and the uncertainty of interference noise, the problems of failure in feature point extraction and interference noise shielding are easily caused in the practical application process of the technology. Thereby affecting the accurate extraction of image features.
For the feature extraction of the complex image, the key points are to suppress and filter the interference noise of the image and improve the extraction accuracy rate when the image has scale and fuzzy change. Aiming at the problems, an improved visual image feature extraction algorithm based on BAKAZE-MAGSAC is provided, by constructing a nonlinear scale space, the edge information of an image is greatly reserved, the interference noise in the image is effectively filtered, and meanwhile, the error matching is eliminated by respectively utilizing the feature point scale information constraint and the root mean square error constraint in the stages of coarse matching and fine matching of features, so that the matching time efficiency and the matching precision are improved.
Disclosure of Invention
In view of the problems existing in the prior art, the technical problem to be solved by the present invention is to provide a visual image feature extraction method based on BAKAZE-magsa, and the specific flow of the method is shown in fig. 1.
The technical scheme comprises the following implementation steps:
(1) determining nonlinear diffusion filtering
Figure BDA0002995111220000011
Figure BDA0002995111220000012
Where M represents image brightness, t represents evolution time, and c (x, y, t) represents propagationDerivative functions, (x, y) represent pixel coordinates, x represents pixel abscissa, and y represents pixel ordinate. The scale parameter σ to be in units of pixelsiConversion to evolution time ti：
Figure BDA0002995111220000021
In the formula, i is a dimensional space.
(2) Determining a non-linear scale space L of an imagei+1：
Li+1＝(I+τA(Li))Li
Wherein L represents an image parameter, I represents an identity matrix, τ represents a time step, and A (L)i) Expressed as a matrix of the image in dimension i, A is expressed as a matrix, LiRepresented as the result of the image L in dimension i. In a nonlinear scale space, a non-maximum value inhibition method is used for calculating a Hessian matrix value of each pixel point in an image pyramid, and meanwhile, the Hessian matrix value is compared with p pixel points on the same layer and upper and lower adjacent layers, and the Hessian matrix maximum value after normalization of different scales is searched.
(3) Determining the main direction of the feature points:
taking a first-order differential L of the characteristic point field on the gradient image by taking the characteristic point as a center and taking the radius r as a statistical rangexAnd LyAnd performing Gaussian weighting calculation, then selecting a sector area with the size of theta to rotate around the origin, calculating the vector sum in the area, and taking the longest direction as the main direction of the characteristic point. And describing the feature points in the image by adopting the BRISK feature descriptor.
(4) Matching the characteristic points:
Figure BDA0002995111220000022
wherein a and b represent corresponding feature point descriptors, miRepresenting the first bit, n, in descriptor aiDenotes the first bit in descriptor b, j denotes the number of bits in the a and b descriptors,
Figure BDA0002995111220000023
the representation descriptor is exclusive-ored.
(5) Determining a model quality function Q*(α,P)：
Selecting s in the input data point set to calculate the characteristic points to obtain a model W, obtaining a homography matrix parameter alpha, and determining a model quality function:
Figure BDA0002995111220000024
where P is a set of data points, l represents the degree of image blur, σmaxRepresenting the maximum value of the projection error, k representing the number of feature points, σiExpressed as the projection error of the characteristic point, d is the counting parameter, pdAnd the probability of the corresponding characteristic point of the image when the dimension is D is expressed, and D represents residual error. And judging whether the current model is the optimal model according to the model quality function.
(6) Determine the probability L (p | α) that each point is an inlier:
Figure BDA0002995111220000031
in the formula, p is the probability that each point is an inlier. And (4) regarding the probability of each point as the weight of each point, and fitting the optimization model by using a weighted least square method according to the weight. Marginalizing the sigma, determining iteration times h:
Figure BDA0002995111220000032
and when the model is the optimal model, updating the iteration times h, otherwise, stopping the iteration. And outputting the optimal model before the output as the optimal model of the data set, calculating correct matching point pairs according to the model, and eliminating error points.
Compared with the prior art, the invention has the advantages that:
(1) the problem that the number of extracted feature points is reduced because the feature points cannot be effectively extracted when the image has scale and fuzzy changes is effectively solved.
(2) The influence of interference noise on feature point extraction is greatly improved, and meanwhile, the matching time efficiency and the matching precision are improved.
Drawings
For a better understanding of the present invention, reference is made to the following further description taken in conjunction with the accompanying drawings.
FIG. 1 is a flow chart of the steps to establish a BAKAZE-MAGSAC based visual image feature extraction method;
FIG. 2 is a flow chart of a method for establishing a BAKAZE-MAGSAC-based visual image feature extraction method;
FIG. 3 is the result of feature extraction for four sets of visual images using the present invention;
detailed description of the preferred embodiments
The present invention will be described in further detail below with reference to examples.
The data selected by the implementation case come from a typical demonstration area of an underground shielding space, 1000 groups of samples are shared in total, wherein 5 scenes including an underground tunnel, a railway tunnel, an underground mall, an underground parking lot and a subway station are provided, each scene has 200 groups of data, 140 groups of samples are extracted from each data in the 5 groups of scenes by adopting a random sampling method to serve as a training set, and the rest are used as a testing set. Finally, the total number of samples used as the training set is 700 and the total number of samples used as the test set is 300.
The overall flow of the visual image feature extraction method provided by the invention is shown in fig. 1, and the specific steps are as follows:
(1) determining nonlinear diffusion filtering
Figure BDA0002995111220000041
Figure BDA0002995111220000042
Where M represents image brightness, t represents evolution time, c (x, y, t) represents a transfer function, (x, y) represents pixel coordinates,x represents the abscissa of the pixel and y represents the ordinate of the pixel. The scale parameter σ to be in units of pixelsiConversion to evolution time ti：
Figure BDA0002995111220000043
In the formula, i is a dimensional space.
(2) Determining a non-linear scale space L of an imagei+1：
Li+1＝(I+3A(Li))Li
In the formula, L represents image parameters, I represents an identity matrix, tau represents a time step, and the value is 3s, A (L)i) Expressed as a matrix of the image in dimension i, A is expressed as a matrix, LiRepresented as the result of the image L in dimension i. In a nonlinear scale space, a non-maximum value inhibition method is used for calculating a Hessian matrix value of each pixel point in an image pyramid, and meanwhile, the Hessian matrix value is compared with 26 pixel points on the same layer, the upper layer and the lower layer, and the Hessian matrix maximum value after normalization of different scales is searched.
(3) Determining the main direction of the feature points:
taking a first-order differential L of the characteristic point field in a statistical range of radius 6 sigma by taking the characteristic point as a center on the gradient imagexAnd LyAnd performing Gaussian weighting calculation, then selecting a 60-degree fan-shaped area to rotate around the origin, calculating the vector sum in the area, and taking the longest direction as the main direction of the characteristic point. And describing the feature points in the image by adopting the BRISK feature descriptor.
(4) Matching the characteristic points:
Figure BDA0002995111220000044
wherein a and b represent corresponding feature point descriptors, miRepresenting the first bit, n, in descriptor aiDenotes the first bit in descriptor b, j denotes the number of bits in the a and b descriptors,
Figure BDA0002995111220000045
the representation descriptor is exclusive-ored.
(5) Determining a model quality function Q*(α,P)：
Selecting 4 pairs of characteristic points in the input data point set for calculation to obtain a model W, obtaining a homography matrix parameter alpha, and determining a model quality function:
Figure BDA0002995111220000051
where P is a set of data points, l represents the degree of image blur, σmaxRepresenting the maximum value of the projection error, k representing the number of feature points, σiExpressed as the projection error of the characteristic point, d is the counting parameter, pdAnd the probability of the corresponding characteristic point of the image when the dimension is D is expressed, and D represents residual error. And judging whether the current model is the optimal model according to the model quality function.
(6) Determine the probability L (p | α) that each point is an inlier:
Figure BDA0002995111220000052
in the formula, p is the probability that each point is an inlier. And (4) regarding the probability of each point as the weight of each point, and fitting the optimization model by using a weighted least square method according to the weight. Marginalizing the sigma, determining iteration times h:
Figure BDA0002995111220000053
and when the model is the optimal model, updating the iteration times h, otherwise, stopping the iteration. And outputting the optimal model before the output as the optimal model of the data set, calculating correct matching point pairs according to the model, and eliminating error points.
In order to verify the accuracy of extracting the visual image features, four groups of visual image feature extraction experiments are performed on the invention, and the experimental results are shown in fig. 3. As can be seen from FIG. 3, the accuracy of feature extraction of the visual image established by the invention is kept above 98%, and higher accuracy can be achieved on the basis of ensuring stability, and the feature extraction effect is good. The visual image feature extraction method established by the invention is effective, provides a better method for establishing accurate image feature extraction, and has certain practicability.

Claims (1)

1. The invention is characterized in that: (1) determining a nonlinear diffusion filter; (2) determining a non-linear scale space of the image; (3) determining the main direction of the feature points; (4) matching the characteristic points; (5) determining a model quality function; (6) determining the probability that each point is an interior point; the method specifically comprises the following six steps:
the method comprises the following steps: determining nonlinear diffusion filtering
Figure FDA0002995111210000011
Figure FDA0002995111210000012
In the formula, M represents image brightness, t represents evolution time, c (x, y, t) represents a transfer function, (x, y) represents pixel coordinates, x represents pixel abscissa, and y represents pixel ordinate; the scale parameter σ to be in units of pixelsiConversion to evolution time ti：
Figure FDA0002995111210000013
In the formula, i is a dimensional space;
step two: determining a non-linear scale space L of an imagei+1：
Li+1＝(I+τA(Li))Li
Wherein L represents an image parameter, I represents an identity matrix, τ represents a time step, and A (L)i) Expressed as an image in dimension iMatrix, A is expressed as matrix, LiExpressed as the result of the image L in dimension i; in a nonlinear scale space, calculating a Hessian matrix value of each pixel point in an image pyramid by using a non-maximum value inhibition method, simultaneously comparing the Hessian matrix value with p pixel points on the same layer and upper and lower adjacent layers, and searching the Hessian matrix maximum value after normalization of different scales;
step three: determining the main direction of the feature points:
taking a first-order differential L of the characteristic point field on the gradient image by taking the characteristic point as a center and taking the radius r as a statistical rangexAnd LyPerforming Gaussian weighting calculation, then selecting a sector area with the size of theta to rotate around an original point, calculating the vector sum in the area, and taking the longest direction as the main direction of the characteristic point; describing the feature points in the image by adopting a BRISK feature descriptor;
step four: matching the characteristic points:
Figure FDA0002995111210000014
wherein a and b represent corresponding feature point descriptors, miRepresenting the first bit, n, in descriptor aiDenotes the first bit in descriptor b, j denotes the number of bits in the a and b descriptors,
Figure FDA0002995111210000015
performing exclusive-or operation on the representation descriptor;
step five: determining a model quality function Q*(α,P)：
Selecting s in the input data point set to calculate the characteristic points to obtain a model W, obtaining a homography matrix parameter alpha, and determining a model quality function:
Figure FDA0002995111210000021
where P is a set of data points, l represents the degree of image blur, σmaxK table representing maximum projection errorNumber of characteristic points, σiExpressed as the projection error of the characteristic point, d is the counting parameter, pdRepresenting the probability of the corresponding characteristic point of the image when the dimension is D, and D represents a residual error; judging whether the current model is the optimal model according to the model quality function;
step six: determine the probability L (p | α) that each point is an inlier:
Figure FDA0002995111210000022
in the formula, p is the probability that each point is an interior point; taking the probability of each point as the weight of each point, and fitting an optimization model by using a weighted least square method according to the weight; marginalizing the sigma, determining iteration times h:
Figure FDA0002995111210000023
when the model is the optimal model, updating the iteration times h, otherwise, stopping the iteration; and outputting the optimal model before the output as the optimal model of the data set, calculating correct matching point pairs according to the model, and eliminating error points.
CN202110327237.0A 2021-03-26 2021-03-26 Visual image feature extraction method based on BAKAZE-MAGSAC Pending CN112906710A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110327237.0A CN112906710A (en) 2021-03-26 2021-03-26 Visual image feature extraction method based on BAKAZE-MAGSAC

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110327237.0A CN112906710A (en) 2021-03-26 2021-03-26 Visual image feature extraction method based on BAKAZE-MAGSAC

Publications (1)

Publication Number Publication Date
CN112906710A true CN112906710A (en) 2021-06-04

Family

ID=76108835

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110327237.0A Pending CN112906710A (en) 2021-03-26 2021-03-26 Visual image feature extraction method based on BAKAZE-MAGSAC

Country Status (1)

Country Link
CN (1) CN112906710A (en)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106022342A (en) * 2016-05-05 2016-10-12 南京邮电大学 Image feature extraction method based on KAZE algorithm
US20160350618A1 (en) * 2015-04-01 2016-12-01 Take-Two Interactive Software, Inc. System and method for image capture and modeling
CN109146833A (en) * 2018-08-02 2019-01-04 广州市鑫广飞信息科技有限公司 A kind of joining method of video image, device, terminal device and storage medium
CN111476737A (en) * 2020-04-15 2020-07-31 腾讯科技（深圳）有限公司 Image processing method, intelligent device and computer readable storage medium

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160350618A1 (en) * 2015-04-01 2016-12-01 Take-Two Interactive Software, Inc. System and method for image capture and modeling
CN106022342A (en) * 2016-05-05 2016-10-12 南京邮电大学 Image feature extraction method based on KAZE algorithm
CN109146833A (en) * 2018-08-02 2019-01-04 广州市鑫广飞信息科技有限公司 A kind of joining method of video image, device, terminal device and storage medium
CN111476737A (en) * 2020-04-15 2020-07-31 腾讯科技（深圳）有限公司 Image processing method, intelligent device and computer readable storage medium

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
吴含前 等: "一种改进的A-KAZE算法在图像配准中的应用", 《东南大学学报（自然科学版）》 *
宋伟等: "基于改进AKAZE算法的无人机影像匹配", 《电子测量与仪器学报》 *
张宇: "基于AKAZE特征的图像配准算法研究", 《中国优秀博硕士学位论文全文数据库(硕士) 信息科技辑》 *
程禹 等: "基于改进AKAZE算法的图像特征匹配方法", 《西安工程大学学报》 *

Similar Documents

Publication Publication Date Title
CN109740603B (en) Vehicle character recognition method based on CNN convolutional neural network
CN108230278B (en) Image raindrop removing method based on generation countermeasure network
CN106548169B (en) Fuzzy literal Enhancement Method and device based on deep neural network
CN113808180B (en) Heterologous image registration method, system and device
CN109344845B (en) Feature matching method based on triple deep neural network structure
CN111242026B (en) Remote sensing image target detection method based on spatial hierarchy perception module and metric learning
CN105184225B (en) A kind of multinational banknote image recognition methods and device
CN111209858A (en) Real-time license plate detection method based on deep convolutional neural network
CN114091606A (en) Tunnel blasting blast hole half-hole mark identification and damage flatness evaluation classification method
CN111401358A (en) Instrument dial plate correction method based on neural network
CN117036737A (en) Feature extraction and matching method based on information entropy, GMS and LC significant detection
CN113610024B (en) Multi-strategy deep learning remote sensing image small target detection method
CN113421223B (en) Industrial product surface defect detection method based on deep learning and Gaussian mixture
CN114627424A (en) Gait recognition method and system based on visual angle transformation
CN113673515A (en) Computer vision target detection algorithm
CN114065798A (en) Visual identification method and device based on machine identification
CN110991501B (en) Improved ORB feature point matching method based on Hessian matrix
CN115880683B (en) Urban waterlogging ponding intelligent water level detection method based on deep learning
CN112070714A (en) Method for detecting copied image based on local ternary counting characteristics
CN116778164A (en) Semantic segmentation method for improving deep V < 3+ > network based on multi-scale structure
CN113850748A (en) Point cloud quality evaluation system and method
CN108960285B (en) Classification model generation method, tongue image classification method and tongue image classification device
CN112906710A (en) Visual image feature extraction method based on BAKAZE-MAGSAC
CN117079272A (en) Bullet bottom socket mark feature identification method combining manual features and learning features
CN115205155A (en) Distorted image correction method and device and terminal equipment

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
WD01 Invention patent application deemed withdrawn after publication
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20210604