CN102819747B - Method for automatically classifying forestry service images - Google Patents
Method for automatically classifying forestry service images Download PDFInfo
- Publication number
- CN102819747B CN102819747B CN201210251256.0A CN201210251256A CN102819747B CN 102819747 B CN102819747 B CN 102819747B CN 201210251256 A CN201210251256 A CN 201210251256A CN 102819747 B CN102819747 B CN 102819747B
- Authority
- CN
- China
- Prior art keywords
- image
- key point
- images
- forestry
- classification
- 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.)
- Active
Links
Landscapes
- Image Analysis (AREA)
Abstract
The invention relates to a method for automatically classifying forestry service images, comprising the main steps of training and classifying. The step of training is as follows: converting images, calculating the set of key points on a gray-scale image, describing the key points by determining the main direction of the key points and generating eigenvectors, clustering, and producing histograms to express the images. The step of classifying is as follows: expressing the classified images with the histograms, and classifying by a classifier. Therefore, the classification of the forestry service images is finished. The numerous forestry service images collected by the forest rangers are used for constructing a reasonable visual vocabulary book according to the characteristics and the color information of the data of the forestry service images, the forestry service images are divided accurately into seven categories including forest fires, illegal use of forest land, illegal logging, illegal hunting and the like, and the forestry service images of different categories are respectively transferred to the functional management departments to realize the fast, effective and timely management of forest and the information modernization of the management of the forest.
Description
Technical field
The present invention relates to a kind of forestry business classification of images method, belong to forest resource monitoring and technical field of information processing.
Background technology
Forestry has huge ecology, economy and social function, is the effective way of ecocrisis and the weather variation issue caused in Coping with Economic Globalization evolution.So become the important construction content of governments at all levels to the forest reserves and ecological protection.And forest ranger is as the backbone of work on forestry, after forestry field data can being taken by mobile phone, be transferred to server.These view data are concentrated and can be classified rapidly according to forestry business demand on the server, and classification results are sent to relevant Management offorestry department, thus can carry out processing in time, effectively to dependent event.
This forest reserves regulatory format, abandon traditional ladder of management and accurately and timely cannot understand its present situation and dynamically, but make each administrative authority of forestry, comprise Law-enforcement in Forestry mechanism and the law-enforcing ranks can coordinate comprehensively, mutual coordination, strengthen decision support and Ability of emergency management, its core is the classification realizing forestry business image.Realize the classification of forestry business image, its theoretical foundation is that to be based upon scene image classificatory.And scene image classification starts the new research field of rising in later 1990s, within 2006, hold scene first at MIT and understood symposial, specify that scene classification will be a new promising study hotspot.Before 2005, scene image classification mainly adopts based on the method for low-level image feature (low level features) and the method based on scene structure; And from 2005 so far, the main method adopted based on image vision vocabulary of scene image classification.
Initial scene classification method mostly based on image low-level image feature (feature such as the overall situation of image or the texture of piecemeal, color), and combines with measure of supervision.But the method based on low-level features utilizes spatial information few, making to there is larger semantic gap between the middle low-level feature of image and high-level semantic, has not been the study hotspot of scene classification at present.
In order to describe each ingredient content and mutual relationship thereof in scene image, the method that scholars propose based on Local Structure of Image or structure Intermediate semantic layer carrys out classified image, thus makes up the semantic gap existed between them.Such as, the scene configuration model (scene configuration model) that the people such as Lipson propose, the combination zone template (composite region templates) that the people such as Smith propose.The method of Lipson and Smith describes scene type by certain contexture, therefore when classified image, the structure of Water demand test pattern forms, and without the need to comparing training sample one by one, makes up low-level features method like this in the deficiency representing image, semantic message context.But its shortcoming is, the design of model may not necessarily the feature of semanteme of accurate description image, and usually needs Image Segmentation Using, and the problem of an Iamge Segmentation inherently more complicated.In addition, the people such as Oliva uses the main contents structure of the visually-perceptible attribute description scenes such as roughness, broad degree, range of extension; The people such as Vogel define one group of local semantic concept, and the frequency of occurrences that generation local semantic conceptual model calculates corresponding semanteme carries out scene classification.Then the method for this two people needs manually to mark a large amount of data, because which limit their range of application.
The people such as Sivic propose the concept of visual vocabulary the earliest in video scene classification and retrieval, and the word bag method (bag of words, BOW) in text classification is applied to Images Classification.
After obtaining the visual vocabulary of image, directly can calculate BOW to represent and carry out scene classification, also probability conventional in text classification can be adopted on BOW represents to generate topic model and carry out modeling, this people's probability implicit semantic analytical model such as latent dirichlet allocation (LDA) model and Bosch etc. (pLSA) comprising the people such as Li carries out unsupervised scene classification.The unsupervised segmentation algorithm of pLSA and LDA, makes training data not need artificial mark.
Scene image classification, through the development of more than ten years, achieves abundant achievement, but often have ignored the colouring information of image, very responsive to the convergent-divergent of image in addition.Particularly the current research to scene image classification all rests on some general natural land images, and image is different classes of to differ greatly.And carry out automated Classification for the image of forestry business and have not been reported.
Summary of the invention
The technical problem to be solved in the present invention is to provide a kind of forestry business classification of images method.
The technical solution adopted in the present invention is divided into training and classification two large divisions, and concrete steps are as follows:
One, train
(1) by each width forestry business image in database
hsv color model image is converted respectively to from RGB color model image
with 256 grades of gray level images
wherein i=1,2 ..., N, N are the number of training image in database;
(2) at gray level image
on, adopt SURF to calculate set of keypoints KP
i(k
1, k
2..., k
ni), wherein Ni represents key point number, key point k
jin (x, y, s), j=1,2 ..., Ni, x, y represent the position of key point, and s is dimensional information;
(3) key point describes: the SURF key point obtained by step one (2), asks corresponding pixel points in HSV space
in H color component on position, then divide 2 steps to realize the description of these key points:
A) key point principal direction is determined:
on image, with each key point for the center of circle, take s as the Haar small echo of point in x, y direction in scale-value to calculate radius be 6s neighborhood, the Haar small echo length of side gets 4s response, and be weighted to the σ=2s Gaussian window of these responses centered by key point, response within the scope of 60 ° be added and form new vector, then travel through whole border circular areas, the direction selecting mould the longest long vector is the principal direction of this key point;
B) proper vector generates:
on image, centered by key point, by X-axis rotate to principal direction, choose according to principal direction the square area that the length of side is 20s, and this region is divided into the subregion of 4 × 4, in each subregion, sample according to the size of 5 × 5, calculate respectively relative to the little wave response of Haar on the horizontal and vertical direction of principal direction, be designated as dx and dy, give weight coefficient equally; Then to the response of every sub regions and the absolute value summation of response, 4 dimensional vectors are obtained in every sub regions
l=1 ..., 16, then 16 sub regions form the description vectors of one 64 dimension
(4) cluster: utilize Hierarchical K-means clustering algorithm to carry out cluster to the feature interpretation vector V that all training images in image data base extract, wherein clusters number K ∈ [3000,8000], these cluster centres are formed visual vocabulary book;
(5) histogram table diagram picture is generated: represent all training images in image data base with the visual vocabulary book that step one (4) obtains, namely in statistical picture, all key point features calculate the number of times occurred in visual vocabulary book by arest neighbors, and the histogram that the frequency that the key point of last image occurs at visual vocabulary book is formed is as iamge description feature;
Two, classify
(1) by step one (1), (2) and (3) same procedure, obtain the H color component in the corresponding HSV space of all SURF key points of image to be classified, and be expressed as the proper vector of 64 dimensions, then the visual vocabulary book using step one (4) to obtain, is expressed as histogram by this image;
(2) SVM classifier is adopted to classify: the histogram of all training images histogram of forestry business image to be sorted and step one (5) obtained is classified as the input of SVM, and the classification results obtained is the classification of the forestry business image of this image.
The invention has the beneficial effects as follows the feature for forestry business view data, the colouring information that make use of in image builds rational visual vocabulary book, thus describes well forestry business image.Image, according to the demand of forestry service management, is carried out high-precision classification by the present invention, various information is passed to each functional management department respectively, meets the needs of forest department's management, realizes quick, timely, accurate and effective management.
Accompanying drawing explanation
Accompanying drawing is the basic procedure schematic diagram of this method
Embodiment
The present invention is also described in further detail with reference to accompanying drawing below in conjunction with embodiment:
Hardware environment for implementing is: Intel Core 2 Duo CPU P8400 2.26G computing machine, 2GB internal memory, 256M video card, the software environment of operation is: Windows XP sp3, Visual C++6.0 and OpenCV.Visual C++6.0 is used to realize the method for the present invention's proposition in conjunction with OpenCV.View data have employed all kinds of forestry business image 2063 width that forest ranger gathers.According to current Management offorestry, business image is divided into: forest fire, illegal use forest land, non-causative fault, forest disease and pest, and animal anomaly is dead, Illegal capture, purchase, transports, peddles wild animal, disorderly adopts and disorderly digs 7 classes such as rare wild plant.
The present invention is divided into training and classification two large divisions, and concrete steps are as follows:
One, train
(1) by each width forestry business image in database
hsv color model image is converted respectively to from RGB color model image
with 256 grades of gray level images
wherein i=1,2 ..., N, N are the number of training image in database, as shown in drawings;
(2) at gray level image
on, adopt SURF to calculate set of keypoints KP
i(k
1, k
2..., k
ni), wherein Ni represents key point number, key point k
jin (x, y, s), j=1,2 ..., Ni, x, y represent the position of key point, and s is dimensional information;
The detection of SURF key point is based on Scale-space theory, adopts approximate Hessian matrix to detect key point.Wherein the calculating of Hessian matrix replaces gaussian filtering second order to lead by being similar to frame-shaped wave filter, accelerates convolution, to improve computing velocity with integral image.In order to make algorithm have scale invariability, detective operators possesses the ability that can find the key point representing same physical location under different scale, and SURF adopts the frame-shaped wave filter of different size to reach metric space layering.From 9 × 9 wave filters, Gauss's second order local derviation of approximate σ=1.2, represent the yardstick of approximate template with s, initial gauges is s=σ=1.2.By the approximate template of initial gauges, the ground floor that convolution algorithm obtains metric space is done to image, then along with yardstick increases, the bank of filters of different scale size will carry out filtering to same image, and the approximate Hessian determinant of a matrix response diagram that the wave filter that can obtain each yardstick is formed, and be made up of the metric space of pyramid structure these figure.
For a certain pixel, after obtaining extreme value with approximate Hessian matrix, on it, a yardstick, next yardstick and this yardstick can construct the three-dimensional neighborhood of 3 × 3 × 3.In three dimension scale space (x, y, s), carry out non-maximal value suppression, only have the point all larger than the response of 26 points closed on just to be chosen as key point.The method of three-dimensional quadratic function matching is finally adopted accurately to locate candidate key point.
The reflection of this feature be the multiple dimensioned local invariant feature of gray level image, it provide a kind of picture material had in statistical significance and describe, thus accurately can reflect the essential attribute of gray level image.
(3) key point describes: the SURF key point obtained by step one (2), asks corresponding pixel points in HSV space
in H color component on position, be ensure rotational invariance, then divide 2 steps to realize the description of these key points:
A) key point principal direction is determined:
on image, with each key point for the center of circle, take s as the Haar small echo of point in x, y direction in scale-value to calculate radius be 6s neighborhood, the Haar small echo length of side gets 4s response, and be weighted to the σ=2s Gaussian window of these responses centered by key point, response within the scope of 60 ° be added and form new vector, then travel through whole border circular areas, the direction selecting mould the longest long vector is the principal direction of this key point;
B) proper vector generates:
on image, centered by key point, by X-axis rotate to principal direction, choose according to principal direction the square area that the length of side is 20s, and this region is divided into the subregion of 4 × 4, in each subregion, sample according to the size of 5 × 5, calculate respectively relative to the little wave response of Haar on the horizontal and vertical direction of principal direction, be designated as dx and dy, give weight coefficient equally; Then to the response of every sub regions and the absolute value summation of response, 4 dimensional vectors are obtained in every sub regions
l=1 ..., 16, then 16 sub regions form the description vectors of one 64 dimension
see characteristic extraction part in accompanying drawing;
(4) cluster: utilize Hierarchical K-means clustering algorithm to carry out cluster to the feature interpretation vector V that all training images in image data base extract, wherein clusters number K ∈ [3000,8000], gets K=5000 in the present embodiment.Be formed centrally visual vocabulary book by these clusters, see cluster in accompanying drawing and form visual vocabulary this part;
(5) histogram table diagram picture is generated: represent all training images in image data base with the visual vocabulary book that step one (4) obtains, namely in statistical picture, all key point features calculate the number of times occurred in visual vocabulary book by arest neighbors, the histogram that the frequency that the key point of last image occurs at visual vocabulary book is formed, as iamge description feature, is shown in the histogram example that in accompanying drawing, the latter half is formed;
Two, classify
(1) by step one (1), (2) and (3) same procedure, obtain the H color component in the corresponding HSV space of all SURF key points of image to be classified, and be expressed as the proper vector of 64 dimensions, then the visual vocabulary book that step one (4) obtains is used, this image is expressed as histogram, sees shown in accompanying drawing;
(2) SVM classifier is adopted to classify: the histogram of all training images histogram of forestry business image to be sorted and step one (5) obtained is classified as the input of SVM, and the classification results obtained is the classification of the forestry business image of this image.
Due to the complicacy of forestry business, obviously not that single type feature can accurate description.Innovative point of the present invention is the feature for forestry business view data, by Scale Model, incorporate colouring information, extract the characteristics of image under different resolution, thus building efficient visual vocabulary book, the effective information making full use of forestry image carries out forestry business Images Classification.The advantage that the present invention brings is the high-accuracy that can reach classification according to forestry business demand, thus the aid decision making situation that can realize relevant data information and formation is sent to relevant Management offorestry department carries out processing in time, effectively.
Claims (1)
1. a forestry business classification of images method, is divided into training and classification two large divisions, it is characterized in that carrying out as follows:
One, train
(1) by each width forestry business image in database
hsv color model image is converted respectively to from RGB color model image
with 256 grades of gray level images
wherein i=1,2 ..., N, N are the number of training image in database;
(2) at gray level image
on, adopt SURF to calculate set of keypoints KP
i(k
1, k
2..., k
ni), wherein Ni represents key point number, key point k
jin (x, y, s), j=1,2 ..., Ni, x, y represent the position of key point, and s is dimensional information;
(3) key point describes: the SURF key point obtained by step one (2), asks corresponding pixel points in HSV space
in H color component on position, then divide 2 steps to realize the description of these key points:
A) key point principal direction is determined:
on image, with each key point for the center of circle, take s as the Haar small echo of point in x, y direction in scale-value to calculate radius be 6s neighborhood, the Haar small echo length of side gets 4s response, and be weighted to the σ=2s Gaussian window of these responses centered by key point, response within the scope of 60 ° be added and form new vector, then travel through whole border circular areas, the direction selecting mould the longest long vector is the principal direction of this key point;
B) proper vector generates:
on image, centered by key point, by X-axis rotate to principal direction, choose according to principal direction the square area that the length of side is 20s, and this region is divided into the subregion of 4 × 4, in each subregion, sample according to the size of 5 × 5, calculate respectively relative to the little wave response of Haar on the horizontal and vertical direction of principal direction, be designated as dx and dy, give weight coefficient equally; Then to the response of every sub regions and the absolute value summation of response, 4 dimensional vectors are obtained in every sub regions
l=1 ..., 16, then 16 sub regions form the description vectors of one 64 dimension
(4) cluster: utilize Hierarchical K-means clustering algorithm to carry out cluster to the feature interpretation vector V that all training images in image data base extract, wherein clusters number K ∈ [3000,8000], these cluster centres are formed visual vocabulary book;
(5) histogram table diagram picture is generated: represent all training images in image data base with the visual vocabulary book that step one (4) obtains, namely in statistical picture, all key point features calculate the number of times occurred in visual vocabulary book by arest neighbors, and the histogram that the frequency that the key point of last image occurs at visual vocabulary book is formed is as iamge description feature;
Two, classify
(1) by step one (1), (2) and (3) same procedure, obtain the H color component in the corresponding HSV space of all SURF key points of image to be classified, and be expressed as the proper vector of 64 dimensions, then the visual vocabulary book using step one (4) to obtain, is expressed as histogram by this image;
(2) SVM classifier is adopted to classify: the histogram of all training images histogram of forestry business image to be sorted and step one (5) obtained is classified as the input of SVM, and the classification results obtained is the classification of the forestry business image of this image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201210251256.0A CN102819747B (en) | 2012-07-18 | 2012-07-18 | Method for automatically classifying forestry service images |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201210251256.0A CN102819747B (en) | 2012-07-18 | 2012-07-18 | Method for automatically classifying forestry service images |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN102819747A CN102819747A (en) | 2012-12-12 |
| CN102819747B true CN102819747B (en) | 2015-07-08 |
Family
ID=47303855
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201210251256.0A Active CN102819747B (en) | 2012-07-18 | 2012-07-18 | Method for automatically classifying forestry service images |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102819747B (en) |
Families Citing this family (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103413142B (en) * | 2013-07-22 | 2017-02-08 | 中国科学院遥感与数字地球研究所 | Remote sensing image land utilization scene classification method based on two-dimension wavelet decomposition and visual sense bag-of-word model |
| CN105488509A (en) * | 2015-11-19 | 2016-04-13 | Tcl集团股份有限公司 | Image clustering method and system based on local chromatic features |
| CN108734130A (en) * | 2018-05-21 | 2018-11-02 | 西南交通大学 | A kind of road detection system, method and storage medium |
| CN109726724B (en) * | 2018-12-21 | 2023-04-18 | 浙江农林大学暨阳学院 | Water gauge image feature weighted learning identification method under shielding condition |
| CN109829372A (en) * | 2018-12-26 | 2019-05-31 | 北京木业邦科技有限公司 | Trees intermediate cutting method, apparatus, electronic equipment and computer readable storage medium |
| CN110675588B (en) * | 2019-09-30 | 2021-06-01 | 北方民族大学 | Forest fire detection device and method |
| CN111860547B (en) * | 2020-08-10 | 2023-04-18 | 华侨大学 | Image segmentation method, device and equipment based on sparse representation and storage medium |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101930547A (en) * | 2010-06-24 | 2010-12-29 | 北京师范大学 | Method for automatically classifying remote sensing image based on object-oriented unsupervised classification |
| CN102496034A (en) * | 2011-11-29 | 2012-06-13 | 南京师范大学 | High-spatial resolution remote-sensing image bag-of-word classification method based on linear words |
-
2012
- 2012-07-18 CN CN201210251256.0A patent/CN102819747B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101930547A (en) * | 2010-06-24 | 2010-12-29 | 北京师范大学 | Method for automatically classifying remote sensing image based on object-oriented unsupervised classification |
| CN102496034A (en) * | 2011-11-29 | 2012-06-13 | 南京师范大学 | High-spatial resolution remote-sensing image bag-of-word classification method based on linear words |
Non-Patent Citations (1)
| Title |
|---|
| 动态成像条件下基于SURF和Mean shift的运动目标高精度检测;胡光龙等;《智能系统学报》;20120215;第7卷(第01期);第61-67页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN102819747A (en) | 2012-12-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN102819747B (en) | Method for automatically classifying forestry service images | |
| CN109271895B (en) | Pedestrian re-identification method based on multi-scale feature learning and feature segmentation | |
| CN102622607B (en) | Remote sensing image classification method based on multi-feature fusion | |
| CN103218831B (en) | A kind of video frequency motion target classifying identification method based on profile constraint | |
| CN102646200B (en) | Image classifying method and system for self-adaption weight fusion of multiple classifiers | |
| CN109508710A (en) | Based on the unmanned vehicle night-environment cognitive method for improving YOLOv3 network | |
| CN102496034B (en) | High-spatial resolution remote-sensing image bag-of-word classification method based on linear words | |
| CN104166841A (en) | Rapid detection identification method for specified pedestrian or vehicle in video monitoring network | |
| CN104616316B (en) | Personage's Activity recognition method based on threshold matrix and Fusion Features vision word | |
| Liu et al. | Enhancing spectral unmixing by local neighborhood weights | |
| CN109948593A (en) | Based on the MCNN people counting method for combining global density feature | |
| CN106354735A (en) | Image target searching method and device | |
| CN108319693A (en) | A kind of geomorphic feature clustering method based on three-dimensional Remote Sensing Database | |
| CN111783589B (en) | Complex scene crowd counting method based on scene classification and multi-scale feature fusion | |
| CN104182763A (en) | Plant type identification system based on flower characteristics | |
| CN107480620A (en) | Remote sensing images automatic target recognition method based on heterogeneous characteristic fusion | |
| CN105184298A (en) | Image classification method through fast and locality-constrained low-rank coding process | |
| CN105224937A (en) | Based on the semantic color pedestrian of the fine granularity heavily recognition methods of human part position constraint | |
| CN105631892A (en) | Aviation image building damage detection method based on shadow and texture characteristics | |
| CN105574545B (en) | The semantic cutting method of street environment image various visual angles and device | |
| CN112101189B (en) | SAR image target detection method and test platform based on attention mechanism | |
| CN110287798A (en) | Vector network pedestrian detection method based on characteristic module and context fusion | |
| Varish | A modified similarity measurement for image retrieval scheme using fusion of color, texture and shape moments | |
| CN103093243A (en) | High resolution panchromatic remote sensing image cloud discriminating method | |
| Shangzheng | A traffic sign image recognition and classification approach based on convolutional neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant |