CN102855497B - Obtain the method for description information of image and device and sorter training method - Google Patents
Obtain the method for description information of image and device and sorter training method Download PDFInfo
- Publication number
- CN102855497B CN102855497B CN201110188933.4A CN201110188933A CN102855497B CN 102855497 B CN102855497 B CN 102855497B CN 201110188933 A CN201110188933 A CN 201110188933A CN 102855497 B CN102855497 B CN 102855497B
- Authority
- CN
- China
- Prior art keywords
- image
- subimage
- similarity
- similarity factor
- feature
- 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 embodiment of the invention discloses a kind of obtain description information of image method and device and sorter training method.The method obtaining description information of image comprises: split described multiple image, obtain the subimage of described multiple image; Based on described subimage characteristic similarity each other, obtain the similarity factor of described subimage; The similarity factor based on described subimage calculates the descriptor of the image corresponding to described subimage.The embodiment of the present invention is when obtaining the descriptor of image, can calculate according to the similarity factor of the feature of this image and correspondence thereof, thus make the feature in the large region of importance degree in image occupy larger proportion, and the feature in the little region of importance degree occupies less proportion, foreground object in image is such as made to occupy larger proportion in final feature, and background content occupies less proportion, effectively reduce the impact of ground unrest, obtain description information of image more accurately.
Description
Technical field
Relate generally to technical field of image processing of the present invention, especially a kind of method and device obtaining description information of image, sorter training method, and image-recognizing method and device.
Background technology
Along with developing rapidly of computer technology and internet, in a lot of fields, the quantity of digital picture and complexity all present volatile rising tendency.Therefore, how the challenge problem naturally becoming us and face is managed fast and effectively to these large nuber of images, mainly comprise the access to these images, access, tissue, retrieval etc.In order to meet this demand, since eighties of last century the nineties, the researchist of every field and scholar have dropped into a lot of energy and have studied image recognition technology, and develop some effective technology and systems.
In image identification system, according to the feature in image, as color, texture and shape facility, the distance between computed image, namely determines the similarity between image, completes the function of image recognition.Such as, 10 class object C1...C10 are identified, so first, need to prepare some images for each classification (from C1 to C10), i.e. training plan image set (training sample), comprise the training plan image set of the image composition C1 of object C1, comprise the training plan image set of the image composition C2 of object C2 ...; Then, extraction characteristics of image is concentrated from these training images; Finally, model is built according to these characteristics of image.At cognitive phase, when a given image to be identified, first from this image, extract characteristics of image, then, the model built before utilization judges the classification of this image, completes recognition function.
But, piece image often comprises complicated content, as comprised foreground object and background parts, and when image recognition often just for foreground object, if when identifying extracts characteristics of image, difference is had no for the foreground object in whole image and background parts, the accuracy of image recognition can be caused to decline.Therefore, before extraction characteristics of image, how to reduce the impact of background parts as much as possible, it is very important for obtaining description information of image more accurately.
Summary of the invention
In view of this, embodiments provide a kind of acquisition methods and device of description information of image, the impact of background parts in image can be reduced, obtain description information of image more accurately.
According to an aspect of the embodiment of the present invention, a kind of method obtaining description information of image based on multiple image is provided, comprises:
Split described multiple image, obtain the subimage of described multiple image;
Based on described subimage characteristic similarity each other, obtain the similarity factor of described subimage;
The similarity factor based on described subimage calculates the descriptor of the image corresponding to described subimage.
According to another aspect of the embodiment of the present invention, a kind of sorter training method is provided, comprises:
Utilize the above-mentioned method obtaining description information of image based on multiple image, using training image as described multiple image, calculate the description information of image of described training image;
Based on the description information of image training classifier of described training image.
According to another aspect of the embodiment of the present invention, a kind of image-recognizing method is provided, comprises:
Utilize and above-mentionedly obtain the method for description information of image based on multiple image, using image to be identified and training image jointly as described multiple image, calculate the description information of image of described image to be identified;
Based on the description information of image of described image to be identified, utilize the sorter based on above-mentioned sorter training method training, discriminator is carried out to described image to be identified.
According to another aspect of the embodiment of the present invention, a kind of device obtaining description information of image based on multiple image is provided, comprises:
Image segmentation unit, is configured to split described multiple image, obtains the subimage of described multiple image;
Factor acquirement unit, is configured to, based on described subimage characteristic similarity each other, obtain the similarity factor of described subimage;
Information acquisition unit, is configured to the descriptor calculating the image corresponding to described subimage based on the similarity factor of described subimage.
According to another aspect of the embodiment of the present invention, a kind of pattern recognition device is provided, comprises:
Description information of image acquisition device as above, for obtaining the description information of image of image to be identified;
Sorter, is configured to the description information of image based on described image to be identified, carries out discriminator to described image to be identified.
In addition, according to a further aspect in the invention, a kind of storage medium is additionally provided.Described storage medium comprises machine-readable program code, when performing described program code on messaging device, described program code makes described messaging device perform according to the above-mentioned method obtaining description information of image based on multiple image of the present invention.
In addition, in accordance with a further aspect of the present invention, a kind of program product is additionally provided.Described program product comprises the executable instruction of machine, and when performing described instruction on messaging device, described instruction makes described messaging device perform according to the above-mentioned method obtaining description information of image based on multiple image of the present invention.
According to above-mentioned a kind of method of the embodiment of the present invention, by calculating the similarity factor of the subimage of zones of different in every piece image, designate the importance degree in each region in this image, thus when obtaining the descriptor of image, can calculate according to the similarity factor of the feature of this image and correspondence thereof, thus make the feature in the large region of importance degree in image occupy larger proportion, and the feature in the little region of importance degree occupies less proportion, foreground object in image is such as made to occupy larger proportion in final feature, and background content occupies less proportion, effectively reduce the impact of ground unrest, obtain description information of image more accurately.
Provide other aspects of the embodiment of the present invention in instructions part below, wherein, describe the preferred embodiment being used for the openly embodiment of the present invention fully in detail, and do not apply to limit to it.
Accompanying drawing explanation
Below in conjunction with specific embodiment, and with reference to accompanying drawing, the above-mentioned of the embodiment of the present invention and other object and advantage are further described.In the accompanying drawings, the identical or corresponding Reference numeral of employing represents by the technical characteristic of identical or correspondence or parts.
Fig. 1 illustrates the process flow diagram providing a kind of method based on multiple image acquisition description information of image as the embodiment of the present invention;
Fig. 2 is the method flow diagram that the segmentation image provided as the embodiment of the present invention is shown;
Fig. 3 is the schematic diagram of the image obtained under a kind of partitioning scheme of providing as the embodiment of the present invention is shown;
Fig. 4 is the schematic diagram of the image obtained under the another kind of partitioning scheme that provides as the embodiment of the present invention is shown;
Fig. 5 is the preparation method process flow diagram of the similarity factor that each subimage provided as the embodiment of the present invention is shown;
Fig. 6 is the schematic diagram of the image weight matrix obtained under a kind of partitioning scheme of providing as the embodiment of the present invention is shown;
Fig. 7 is the schematic diagram of the image weight matrix obtained under the another kind of partitioning scheme that provides as the embodiment of the present invention is shown;
Fig. 8 illustrates the method flow diagram providing the descriptor based on the image corresponding to the similarity factor calculating subimage of subimage as the embodiment of the present invention;
Fig. 9 is the schematic diagram of the final image weight matrix obtained under two kinds of partitioning schemes providing as the embodiment of the present invention are shown;
Figure 10 is the schematic diagram that the image characteristic point provided as the embodiment of the present invention is shown;
Figure 11 illustrates a kind of schematic diagram obtaining the device of description information of image based on multiple image provided as the embodiment of the present invention;
Figure 12 is the structural representation that a kind of pattern recognition device provided as the embodiment of the present invention is shown;
Figure 13 is the block diagram of the example arrangement of the personal computer illustrated as the messaging device adopted in embodiments of the invention.
Embodiment
With reference to the accompanying drawings embodiments of the invention are described.
See Fig. 1, a kind of method based on multiple image acquisition description information of image that the embodiment of the present invention provides comprises:
S101: segmentation multiple image, obtains the subimage of multiple image.
In the embodiment of the present invention, this multiple image can be a training plan image set.When obtaining the descriptor of each image, first carry out splitting the subimage obtaining multiple image to this multiple image, this cutting procedure can adopt a kind of partitioning scheme once to split every width image, and multiple different partitioning scheme also can be adopted to carry out repeated segmentation to every width image.During concrete segmentation, the method for stress and strain model can be adopted to split, Iamge Segmentation is become the rectangular area of N × M, or triangle or polygonal mesh, and grid not want Seeking Truth uniform; In addition, also can adopt the various existing or image segmentation algorithm developed future, split by the content of image and/or visual signature and/or other features.Wherein, when adopting multiple partitioning scheme to split image, as far as possible different to the partitioning scheme of same image, in the same once segmentation of multiple image, different images adopts identical partitioning scheme as far as possible.
S102: based on subimage characteristic similarity each other, obtains the similarity factor of subimage.
After the subimage obtaining each image, feature can be extracted to each subimage, such as, color, shape and textural characteristics, then the similarity between the feature calculating each subimage in this multiple image, obtain the similarity factor of each subimage, the similarity of a certain subimage and other subimage is higher, and the similarity factor values of this subimage is also larger.For each image that same training image is concentrated, the subimage that the similarity factor is high is more likely the critical area in its image, such as, foreground object in image.
Wherein, method for calculating the subimage similarity factor has multiple, the method of sequence (ranking) is such as adopted to process all subimages in this multiple image, as manifold ranking or pagerank technology etc., wherein, manifold ranking technology can be Dengyong Zhou see author, Jason Weston, Arthur Gretton, name is called Ranking on Data Manifold, be published in the list of references in Proceeding of the Advances in Neural Information Processing Systems 2003, pagerank technology can see United States Patent (USP) Method for node ranking in a linked database (publication number: 6, 285, 999, the applying date: September 4 calendar year 2001), United States Patent (USP) Method for scoring documents in a linked database (publication number: 6, 799, 176, the applying date: on September 28th, 2004), United States Patent (USP) Method for node ranking in a linked database (publication number: 7, 058, 628, the applying date: on June 6th, 2006), and United States Patent (USP) Scoring documents in a linked database (publication number: 7, 269, 587, the applying date: on September 11st, 2007).The full content of document is by quoting and being all attached in the application herein.According to the similarity between each subimage feature, each subimage is sorted, obtain rank corresponding to each subimage (rank) value, using the similarity factor of this rank value as corresponding subimage.Wherein, different similarity factor computing method can be had according to the partitioning scheme of multiple image, specifically please refer to the description of subsequent embodiment.
S103: the similarity factor based on subimage calculates the descriptor of the image corresponding to subimage.
When obtaining the descriptor of image, can to the feature of each position of image, the similarity factor corresponding according to this position is weighted.The similarity factor that in each image, diverse location is corresponding is relevant to the partitioning scheme of this image, for the situation adopting a kind of partitioning scheme to split described multiple image, the similarity factor that each image diverse location is corresponding is the similarity factor of the subimage corresponding to this position, for adopting multiple partitioning scheme to split the situation of described multiple image, the similarity factor corresponding to each image diverse location can the similarity factor of multiple subimages corresponding to this position obtain.Specifically refer to the description of subsequent embodiment.
The embodiment of the present invention is by calculating the similarity factor of the subimage of zones of different in every piece image, designate the importance degree in each region in this image, thus when obtaining the descriptor of image, can calculate according to the similarity factor (i.e. importance degree) of the feature of this image and correspondence thereof, thus make the feature in the large region of importance degree in image occupy larger proportion, and the feature in the little region of importance degree occupies less proportion, foreground object in image is such as made to occupy larger proportion in final feature, and background content occupies less proportion, effectively reduce the impact of ground unrest, obtain description information of image more accurately.
In another embodiment of the invention, to adopt two kinds of different partitioning scheme segmentation multiple images to be described.As shown in Figure 2, the process of this segmentation image can comprise the following steps:
S201, becomes the rectangular area of 2 × 2 (or other numerical value) by the every width Iamge Segmentation in multiple image.
As shown in Figure 3, for wherein piece image A, the image after segmentation comprises subimage a, b, c, d.
Every width Iamge Segmentation in above-mentioned multiple image is the rectangular area of 3 × 3 (or other numerical value) by S202.
As shown in Figure 4, still for image A, the image after segmentation comprises subimage e, f, g, h, i, j, k, l, m.
Above two steps are irrelevant mutually, and its order can adjust as required, and in other embodiments, partitioning scheme can have more kinds of, repeats no longer one by one herein.
For above-mentioned partitioning scheme, the similarity factor of each subimage can obtain by the following method, and as shown in Figure 5, the method can comprise:
S501, extracts the feature of all subimages.
This feature extracting method can adopt any image feature extraction techniques of the prior art, repeats no more herein.All subimages in the present embodiment are all subimages obtained under above-mentioned multiple partitioning scheme.
S502, according to the similarity between all subimage features, sorts to all subimages, obtains the rank value of each subimage.
In the present embodiment, adopt manifold ranking method to sort, above-mentioned twice segmentation is carried out to multiple image and obtains N number of subimage altogether afterwards, be designated as A
1..., A
n.After feature extraction, each subimage generates a proper vector, and these proper vectors are designated as f
1... f
n.
First initialization: with vectorial R=[r
1..., r
n] represent the rank value of each subimage, and be initialized as 1, wherein r
nrepresent the rank value (rank value) of N number of subimage;
Calculate any two proper vector f
i, f
jbetween distance, be designated as d
ij.Citing calculating can adopt the methods such as Euclidean distance; According to the distance d between subimage proper vector
ij, for each subimage determines its k neighbour subimage.That is: with certain subimage apart from the k neighbour subimage of k subimage before minimum as this subimage.Wherein, k is empirical parameter.
Calculate similarity matrix W (affinity matrix); Similarity matrix obtains according to the distance between subimage proper vector, and its size is NxN matrix, all corresponding subimage of every a line of matrix, each row.W
ijthe element of the i-th row, jth row in representing matrix, σ is empirical parameter, desirable all d
ijmean value, or other empirical values.
The normalization of matrix W.S=D
-1/2wD
-1/2, wherein, D is a diagonal matrix, on diagonal line the value of element equal the value of element in corresponding row in W and;
Loop iteration according to the following formula, until convergence:
R(t+1)＝a＊S＊R(t)+(1-a)＊1
Wherein, a is an empirical parameter between [0,1], generally gets the value close to 1.The rank value of subimage when R (t+1) and R (t) is illustrated respectively in t+1 and the t time iteration.S be above-mentioned W is normalized after the matrix that obtains.
Namely the rank value (Rank value) of each subimage obtained can be used as the similarity factor of each subimage.
In another embodiment, the similarity factor of each subimage can also be obtained by other method, such as, still for two kinds of partitioning schemes shown in above-mentioned Fig. 3,4, under acquisition two kinds of partitioning schemes multiple image subimage after, extract the feature of each subimage, then to each partitioning scheme, based on all subimage features similarity each other that this partitioning scheme obtains, obtain the similarity factor of each subimage.Concrete, not that all subimages under two kinds of partitioning schemes are sorted according to characteristic similarity, but for different partitioning schemes, all subimages obtained under each partitioning scheme are sorted, also namely carry out two minor sorts, obtain the similarity factor of each subimage respectively.
For under more kinds of partitioning scheme, the preparation method of the subimage similarity factor also can adopt similar above-mentioned two kinds of modes, respectively: based on the feature similarity each other of all subimages that all partitioning schemes obtain, obtain the similarity factor of each subimage; Or, to each partitioning scheme, based on the feature similarity each other of all subimages that this partitioning scheme obtains, obtain the similarity factor of each subimage.
Certainly, in another embodiment, if what adopt above-mentioned multiple image is a kind of partitioning scheme, then only feature extraction need be carried out, then according to the similarity factor of each subimage of the Similarity measures between subimage feature to all subimages obtained.
After the similarity factor obtaining each subimage, in embodiments of the present invention, as shown in Figure 8, the similarity factor based on subimage calculates the descriptor of the image corresponding to subimage, can comprise:
S801, according to the similarity factor of each subimage, obtains the weight matrix of image.
After the similarity factor obtaining each subimage, according to each subimage position in the picture, the weight matrix of this image for token image each region importance degree can be obtained.
According to multiple partitioning scheme, multiple image is split, then can obtain corresponding weight matrix under different partitioning scheme, as shown in Figure 6,7, weight matrix wherein shown in Fig. 6 corresponds to the similarity factor of each subimage that this image obtains under partitioning scheme as shown in Figure 3, and the weight matrix shown in Fig. 7 corresponds to the similarity factor of each subimage that same image obtains under partitioning scheme as shown in Figure 4.The weight matrix that then this image is final can obtain based on the weight matrix shown in Fig. 6,7, the concrete mode of two weight matrix superpositions that can adopt obtains, this superposition can be the addition of the similarity factor of overlapping region in Fig. 6,7, or be multiplied, or the mathematic(al) manipulation of other any embodiment Overlay, such as this image weight matrix final as shown in Figure 9, for the superposition to the weight matrix shown in Fig. 6,7, also namely to the addition of the similarity factor of overlapping region in Fig. 6,7, namely the weight matrix shown in this Fig. 9 indicates the importance of this image zones of different.
Split multiple image according to a kind of partitioning scheme, as only adopted the partitioning scheme shown in Fig. 3 to split, then the weight matrix according to Fig. 6 of each subimage position acquisition in the picture is the final weight matrix of this image.
After the weight matrix obtaining image, the follow-up weight matrix of this image that can utilize, to the characteristic weighing in image, forms the descriptor of image.
S802, extracts the feature of image.
The extraction of this feature can use conventional methods, such as: color histogram, edge orientation histogram, local feature etc.This step 802 and step 801 can be carried out simultaneously or adjustment sequentially, is not construed as limiting herein as required.
S803, according to the weight matrix of image, determines the similarity factor corresponding to the position residing for characteristics of image extracted.
After the weight matrix obtaining image, namely obtaining the similarity factor that this image diverse location place is corresponding, is also weights, and then according to the position at the characteristics of image place of extracting, can determine the similarity factor of this position.
For the weight matrix shown in Fig. 9, the similarity factor corresponding to the characteristics of image of extraction is the weights corresponding to this feature position in fig .9.
S804, the similarity factor corresponding to this position, is weighted the feature extracted, to obtain the descriptor of image.
Illustrate for edge orientation histogram, suppose that the edge orientation histogram that will generate comprises 180 bin, namely each bin corresponds to 1 degree in 0 degree to 180 degree edge direction interval, and this histogram is designated as H.As the unique point a of 4 in Figure 10, b, c, d.Suppose that the edge direction of these four some correspondences is respectively 30 degree, 45 degree, 120 degree, 30 degree.
During edge calculation direction histogram, need each unique point in traversing graph picture, and the number of the unique point of statistics in each edge direction.For the unique point of 4 in Figure 10:
When running into first unique point a, the edge direction of this unique point is 30 degree, therefore, needs the weights 0.4 adding this Feature point correspondence on the histogram bin corresponding with 30 degree, obtains H (30)=0.4;
Similar with said process, for unique point b and c, need the weights adding that on the histogram bin corresponding with 45 degree and 120 b and c point is corresponding, obtain H (45)=0.6, H (120)=0.4;
When running into the 4th unique point d, the edge direction of this unique point is 30 degree, therefore, needs the weights 0.4 adding this Feature point correspondence on the histogram bin corresponding with 30 degree.In process before, obtain H (30)=0.4, so after statistical nature point d, new value is H (30)=0.4+0.4=0.8.
H (30)=0.4 a point
H (45)=0.6 b point
H (120)=0.4 c point
H (30)=0.4+0.4=0.8 d point
By the acquisition of the characteristics of image corresponding position similarity factor, and according to the weighting of similarity factor pair feature, make foreground object occupy larger proportion in final feature, and background content occupies less proportion, reach the object reducing ground unrest.
In another embodiment of the invention, before the weights weighting that the feature of each position of image is obtained according to the similarity factor of the multiple subimages corresponding to this position, can first be normalized the similarity factor of the subimage obtained under different partitioning scheme, and then to the weights weighting that the feature of each position of image obtains according to the similarity factor after the normalization of the multiple subimages corresponding to this position.
Based on the method for above-mentioned acquisition description information of image, the embodiment of the present invention also provides a kind of sorter training method, the training image that first training image is concentrated by the method is as the multiple image in above-described embodiment, then said method is utilized to obtain the description information of image of training image, then based on the description information of image training classifier of this training image.
Based on the method for above-mentioned description information of image, the embodiment of the present invention also provides a kind of image-recognizing method, the training image that first image to be identified and training image are concentrated by the method is jointly as the multiple image in above-described embodiment, then the method for above-described embodiment is utilized to obtain the description information of image of image to be identified, based on the description information of image of this image to be identified, the sorter utilizing above-described embodiment to provide carries out discriminator to this image to be identified.
The embodiment of the present invention additionally provides a kind of device of the acquisition description information of image based on multiple image, and as shown in figure 11, this device can comprise:
Image segmentation unit 1101, is configured to split multiple image, obtains the subimage of this multiple image;
Factor acquirement unit 1102, is configured to based on subimage characteristic similarity each other, obtains the similarity factor of subimage;
Information acquisition unit 1103, is configured to the descriptor calculating the image corresponding to subimage based on the similarity factor of subimage.Information acquisition unit specifically can be configured to the feature of each position of image according to corresponding similarity Factors Weighting.
Image segmentation unit 1101 carries out this multiple image splitting the subimage obtaining multiple image, this cutting procedure can adopt a kind of partitioning scheme once to split every width image, also multiple different partitioning scheme can be adopted to carry out repeated segmentation to every width image, after the subimage obtaining each image, factor acquirement unit 1102 extracts feature to each subimage, then the similarity between the feature calculating each subimage in this multiple image, obtain the similarity factor of each subimage, information acquisition unit 1103, can to the feature of each position of image, the similarity factor corresponding according to this position is weighted, calculate the descriptor of the image corresponding to subimage.
The embodiment of the present invention calculates the similarity factor of the subimage of zones of different in every piece image by said units, designate the importance degree in each region in this image, thus when obtaining the descriptor of image, can calculate according to the similarity factor of the feature of this image and correspondence thereof, thus make the feature in the large region of importance degree in image occupy larger proportion, and the feature in the little region of importance degree occupies less proportion, foreground object in image is such as made to occupy larger proportion in final feature, and background content occupies less proportion, effectively reduce the impact of ground unrest, obtain description information of image more accurately.
In another embodiment of the invention, image segmentation unit, is further configured to and splits every piece image according to multiple different partitioning scheme; Factor acquirement unit, be further configured to each partitioning scheme, based on all subimages characteristic similarity each other that this partitioning scheme obtains, obtain the similarity factor of each subimage, the concrete method that can adopt ranking, subimage is sorted, obtains the rank value of each subimage; Information acquisition unit, is further configured to the weights weighting obtained according to the similarity factor based on the multiple subimages corresponding to described position the feature of each position of image.
In another embodiment of the invention, factor acquirement unit, can also be further configured to all subimages characteristic similarity each other obtained based on all partitioning schemes, obtain the similarity factor of each subimage.
In another embodiment of the invention, information acquisition unit, can also be further configured to the feature of each position of the image similarity Factors Weighting according to the subimage corresponding to described position.
In another embodiment of the invention, this device can also comprise normalization unit, be configured to before the described weights weighting that the feature of each position of image is obtained according to the similarity factor of the multiple subimages corresponding to described position, the similarity factor of the subimage obtained under different partitioning scheme is normalized, and then the weights weighting feature of each position of image obtained according to the similarity factor after the normalization of the multiple subimages corresponding to described position by described information acquisition unit.
In another embodiment of the invention, as shown in figure 12, be a kind of pattern recognition device, this device can comprise:
Description information of image acquisition device 1201, for obtaining the description information of image of image to be identified.
Sorter 1202, is configured to the description information of image based on image to be identified, treats recognition image and carries out discriminator.
Wherein, description information of image acquisition device 1201 is similar with the device of the acquisition description information of image based on multiple image in previous embodiment, repeat no more herein, using image to be identified as the wherein piece image in above-mentioned multiple image, after this description information of image acquisition device 1201 obtains the description information of image of image to be identified, by the description information of image of sorter 1202 based on this image to be identified, treat recognition image and carry out discriminator, wherein, this sorter 1202 adopts the sorter training method in previous embodiment to train.
In addition, should also be noted that above-mentioned series of processes and device also can be realized by software and/or firmware.When being realized by software and/or firmware, from storage medium or network to the computing machine with specialized hardware structure, general purpose personal computer 1300 such as shown in Figure 13 installs the program forming this software, and this computing machine, when being provided with various program, can perform various function etc.
In fig. 13, CPU (central processing unit) (CPU) 1301 performs various process according to the program stored in ROM (read-only memory) (ROM) 1302 or from the program that storage area 1308 is loaded into random access memory (RAM) 1303.In RAM 1303, also store the data required when CPU 1301 performs various process etc. as required.
CPU 1301, ROM 1302 and RAM 1303 are connected to each other via bus 1304.Input/output interface 1305 is also connected to bus 1304.
Following parts are connected to input/output interface 1305: importation 1306, comprise keyboard, mouse etc.; Output 1307, comprises display, such as cathode-ray tube (CRT) (CRT), liquid crystal display (LCD) etc., and loudspeaker etc.; Storage area 1308, comprises hard disk etc.; With communications portion 1309, comprise network interface unit such as LAN card, modulator-demodular unit etc.Communications portion 1309 is via network such as the Internet executive communication process.
As required, driver 1310 is also connected to input/output interface 1305.Detachable media 1311 such as disk, CD, magneto-optic disk, semiconductor memory etc. are installed on driver 1310 as required, and the computer program therefrom read is installed in storage area 1308 as required.
When series of processes above-mentioned by software simulating, from network such as the Internet or storage medium, such as detachable media 1311 installs the program forming software.
It will be understood by those of skill in the art that this storage medium is not limited to wherein having program stored therein shown in Figure 13, distributes the detachable media 1311 to provide program to user separately with equipment.The example of detachable media 1311 comprises disk (comprising floppy disk (registered trademark)), CD (comprising compact disc read-only memory (CD-ROM) and digital universal disc (DVD)), magneto-optic disk (comprising mini-disk (MD) (registered trademark)) and semiconductor memory.Or hard disk that storage medium can be ROM 1302, comprise in storage area 1308 etc., wherein computer program stored, and user is distributed to together with comprising their equipment.
Also it is pointed out that the step performing above-mentioned series of processes can order naturally following the instructions perform in chronological order, but do not need necessarily to perform according to time sequencing.Some step can walk abreast or perform independently of one another.
Although described the present invention and advantage thereof in detail, be to be understood that and can have carried out various change when not departing from the spirit and scope of the present invention limited by appended claim, substituting and conversion.And, the term of the embodiment of the present invention " comprises ", " comprising " or its any other variant are intended to contain comprising of nonexcludability, thus make to comprise the process of a series of key element, method, article or equipment and not only comprise those key elements, but also comprise other key elements clearly do not listed, or also comprise by the intrinsic key element of this process, method, article or equipment.When not more restrictions, the key element limited by statement " comprising ... ", and be not precluded within process, method, article or the equipment comprising described key element and also there is other identical element.
About the embodiment comprising above embodiment, following remarks is also disclosed:
Remarks 1. 1 kinds obtains the method for description information of image based on multiple image, comprising:
Split described multiple image, obtain the subimage of described multiple image;
Based on described subimage characteristic similarity each other, obtain the similarity factor of described subimage;
The similarity factor based on described subimage calculates the descriptor of the image corresponding to described subimage.
The method of remarks 2. according to remarks 1, wherein, the process that the similarity factor based on described subimage calculates the descriptor of the image corresponding to described subimage comprises:
To the feature of each position of image according to corresponding similarity Factors Weighting.
The method of remarks 3. according to remarks 2, comprises according to the process of corresponding similarity Factors Weighting the feature of each position of image:
To the feature of each position of the image similarity Factors Weighting according to the subimage corresponding to described position.
The method of remarks 4. according to remarks 2, wherein, the described multiple image of described segmentation comprises:
Every piece image is split according to multiple different partitioning scheme;
Wherein, the feature of each position of image is comprised according to the process of corresponding similarity Factors Weighting:
To the weights weighting that the feature of each position of image obtains according to the similarity factor based on the multiple subimages corresponding to described position.
The method of remarks 5. according to remarks 4, wherein, the process obtaining the similarity factor of described subimage comprises:
To each partitioning scheme, based on all subimages characteristic similarity each other that this partitioning scheme obtains, obtain the similarity factor of each subimage.
The method of remarks 6. according to remarks 4, wherein, the process obtaining the similarity factor of described subimage comprises:
Based on all subimages characteristic similarity each other that all partitioning schemes obtain, obtain the similarity factor of each subimage.
The method of remarks 7. according to remarks 4, wherein, before the weights weighting obtained according to the similarity factor based on the multiple subimages corresponding to described position the feature of each position of image, also comprises:
The similarity factor of the subimage obtained under different partitioning scheme is normalized.
The method of remarks 8. according to remarks 1, wherein, the process obtaining the similarity factor of described subimage comprises:
Adopt the method for sequence, based on described subimage characteristic similarity each other, described subimage is sorted, obtain the rank value of each subimage, using the similarity factor of the rank value of each subimage of acquisition as each subimage.
Remarks 9. 1 kinds of sorter training methods, comprising:
Utilize the method as described in remarks 1-8, using training image as described multiple image, calculate the description information of image of described training image;
Based on the description information of image training classifier of described training image.
Remarks 10. 1 kinds of image-recognizing methods, comprising:
The method as described in remarks 1-8 of utilization, using image to be identified and training image jointly as described multiple image, calculates the description information of image of described image to be identified;
Based on the description information of image of described image to be identified, the sorter utilizing the method based on remarks 9 to train, carries out discriminator to described image to be identified.
Remarks 11. 1 kinds obtains the device of description information of image based on multiple image, comprising:
Image segmentation unit, is configured to split described multiple image, obtains the subimage of described multiple image;
Factor acquirement unit, is configured to, based on described subimage characteristic similarity each other, obtain the similarity factor of described subimage;
Information acquisition unit, is configured to the descriptor calculating the image corresponding to described subimage based on the similarity factor of described subimage.
The device of remarks 12. according to remarks 11, wherein,
Described information acquisition unit is further configured to the feature of each position of image according to corresponding similarity Factors Weighting.
The device of remarks 13. according to remarks 12, wherein,
Described information acquisition unit is further configured to the feature of each position of the image similarity Factors Weighting according to the subimage corresponding to described position.
The device of remarks 14. according to remarks 12, wherein,
Described image segmentation unit is further configured to splits every piece image according to multiple different partitioning scheme;
Described information acquisition unit is further configured to the weights weighting obtained according to the similarity factor based on the multiple subimages corresponding to described position the feature of each position of image.
The device of remarks 15. according to remarks 14, wherein,
Described factor acquirement unit is further configured to each partitioning scheme, based on all subimages characteristic similarity each other that this partitioning scheme obtains, obtains the similarity factor of each subimage.
The device of remarks 16. according to remarks 14, wherein,
Described factor acquirement unit is further configured to all subimages characteristic similarity each other obtained based on all partitioning schemes, obtains the similarity factor of each subimage.
The device of remarks 17. according to remarks 14, also comprises:
Normalization unit, be configured to, before the described weights weighting obtained according to the similarity factor of the multiple subimages corresponding to described position the feature of each position of image, be normalized the similarity factor of the subimage obtained under different partitioning scheme.
The device of remarks 18. according to remarks 11, wherein,
Described factor acquirement unit is further configured to the method adopting sequence, based on described subimage characteristic similarity each other, described subimage is sorted, obtains the rank value of each subimage, using the similarity factor of the rank value of each subimage of acquisition as each subimage.
Remarks 19. 1 kinds of pattern recognition devices, comprising:
Description information of image acquisition device as described in remarks 11-18, for obtaining the description information of image of image to be identified;
Sorter, is configured to the description information of image based on described image to be identified, carries out discriminator to described image to be identified.
Claims (9)
1. obtain a method for description information of image based on multiple image, comprising:
Split described multiple image, obtain the subimage of described multiple image;
Based on described subimage characteristic similarity each other, obtain the similarity factor of described subimage;
The similarity factor based on described subimage calculates the descriptor of the image corresponding to described subimage,
Wherein, the process calculating the descriptor of the image corresponding to described subimage based on the similarity factor of described subimage comprises: to the feature of each position of image according to corresponding similarity Factors Weighting.
2. method according to claim 1, comprises according to the process of corresponding similarity Factors Weighting the feature of each position of image:
To the feature of each position of the image similarity Factors Weighting according to the subimage corresponding to described position.
3. method according to claim 1, wherein, the described multiple image of described segmentation comprises:
Every piece image is split according to multiple different partitioning scheme;
Wherein, the feature of each position of image is comprised according to the process of corresponding similarity Factors Weighting:
To the weights weighting that the feature of each position of image obtains according to the similarity factor based on the multiple subimages corresponding to described position.
4. method according to claim 3, wherein, the process obtaining the similarity factor of described subimage comprises:
To each partitioning scheme, based on all subimages characteristic similarity each other that this partitioning scheme obtains, obtain the similarity factor of each subimage.
5. method according to claim 3, wherein, the process obtaining the similarity factor of described subimage comprises:
Based on all subimages characteristic similarity each other that all partitioning schemes obtain, obtain the similarity factor of each subimage.
6. method according to claim 3, wherein, before the weights weighting obtained according to the similarity factor based on the multiple subimages corresponding to described position the feature of each position of image, also comprises:
The similarity factor of the subimage obtained under different partitioning scheme is normalized.
7. method according to claim 1, wherein, the process obtaining the similarity factor of described subimage comprises:
Adopt the method for sequence, based on described subimage characteristic similarity each other, described subimage is sorted, obtain the rank value of each subimage, using the similarity factor of the rank value of each subimage of acquisition as each subimage.
8. a sorter training method, comprising:
Utilize the method as described in claim 1-7, using training image as described multiple image, calculate the description information of image of described training image;
Based on the description information of image training classifier of described training image.
9. obtain a device for description information of image based on multiple image, comprising:
Image segmentation unit, is configured to split described multiple image, obtains the subimage of described multiple image;
Factor acquirement unit, is configured to, based on described subimage characteristic similarity each other, obtain the similarity factor of described subimage;
Information acquisition unit, is configured to the descriptor calculating the image corresponding to described subimage based on the similarity factor of described subimage,
Wherein, the descriptor of the image calculated corresponding to described subimage based on the similarity factor of described subimage comprises: to the feature of each position of image according to corresponding similarity Factors Weighting.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201110188933.4A CN102855497B (en) | 2011-06-27 | 2011-06-27 | Obtain the method for description information of image and device and sorter training method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201110188933.4A CN102855497B (en) | 2011-06-27 | 2011-06-27 | Obtain the method for description information of image and device and sorter training method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN102855497A CN102855497A (en) | 2013-01-02 |
| CN102855497B true CN102855497B (en) | 2015-07-29 |
Family
ID=47402073
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201110188933.4A Active CN102855497B (en) | 2011-06-27 | 2011-06-27 | Obtain the method for description information of image and device and sorter training method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102855497B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104123713B (en) * | 2013-04-26 | 2017-03-01 | 富士通株式会社 | Many image joint dividing methods and device |
| CN108200043B (en) * | 2017-12-28 | 2020-07-31 | 中国联合网络通信集团有限公司 | Picture verification code verification method and picture verification code verification device |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2006338313A (en) * | 2005-06-01 | 2006-12-14 | Nippon Telegr & Teleph Corp <Ntt> | Similar image retrieving method, similar image retrieving system, similar image retrieving program, and recording medium |
| CN101137046A (en) * | 2006-08-31 | 2008-03-05 | 索尼株式会社 | Image processing device, method and program |
-
2011
- 2011-06-27 CN CN201110188933.4A patent/CN102855497B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2006338313A (en) * | 2005-06-01 | 2006-12-14 | Nippon Telegr & Teleph Corp <Ntt> | Similar image retrieving method, similar image retrieving system, similar image retrieving program, and recording medium |
| CN101137046A (en) * | 2006-08-31 | 2008-03-05 | 索尼株式会社 | Image processing device, method and program |
Also Published As
| Publication number | Publication date |
|---|---|
| CN102855497A (en) | 2013-01-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Shen et al. | Multi-crop convolutional neural networks for lung nodule malignancy suspiciousness classification | |
| Bansal et al. | An efficient technique for object recognition using Shi-Tomasi corner detection algorithm | |
| Xie et al. | Comparison among dimensionality reduction techniques based on Random Projection for cancer classification | |
| CN109118564A (en) | A kind of three-dimensional point cloud labeling method and device based on fusion voxel | |
| CN105488515A (en) | Method for training convolutional neural network classifier and image processing device | |
| JP2015506026A (en) | Image classification | |
| CN101196905A (en) | Intelligent pattern searching method | |
| Wang et al. | Bag-of-features based classification of breast parenchymal tissue in the mammogram via jointly selecting and weighting visual words | |
| CN110347851A (en) | Image search method and system based on convolutional neural networks | |
| Khrissi et al. | Image Segmentation based on k-means and genetic algorithms | |
| CN107679553A (en) | Clustering method and device based on density peaks | |
| Qi et al. | Hep-2 cell classification: The role of gaussian scale space theory as a pre-processing approach | |
| CN104751175A (en) | Multi-label scene classification method of SAR (Synthetic Aperture Radar) image based on incremental support vector machine | |
| CN102902976A (en) | Image scene classification method based on target and space relationship characteristics | |
| Grycuk et al. | Content-based image retrieval optimization by differential evolution | |
| Wu et al. | Joint entropy based learning model for image retrieval | |
| CN102855497B (en) | Obtain the method for description information of image and device and sorter training method | |
| Chung et al. | Fast randomized algorithm for center-detection | |
| Teng et al. | An active contour model based on hybrid energy and fisher criterion for image segmentation | |
| Wang et al. | An efficient k-medoids clustering algorithm for large scale data | |
| Bruch et al. | Evaluation of semi-supervised learning using sparse labeling to segment cell nuclei | |
| Sengottuvelan et al. | Object classification using substance based neural network | |
| Patidar | An efficient class-based data clustering through K-means and KNN approach | |
| Diecente Cid et al. | Lung graph-model classiffication with SVM and CNN for tuberculosis severity assessment and automatic CT report generation: participation in the ImageCLEF 2019 Tuberculosis Task | |
| Xiao et al. | Dynamic semantic feature-based long-term cross-session learning approach to content-based image retrieval |
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 |