CN106570125A - Remote sensing image retrieval method and remote sensing image retrieval device based on rotation/zooming/translation invariance - Google Patents

Abstract

The invention relates to the field of remote sensing image retrieval technology, and particularly to a remote sensing image retrieval method and a remote sensing image retrieval device based on rotation/zooming/translation invariance. The retrieval method comprises the steps of a, after performing image transformation processing on each image in an image database, extracting edge point pixels of each image, and constructing a transaction set of each image; b, extracting an association rule in the image through the transaction set of each image; and c, calculating an association rule similarity between the to-be-retrieved image and the retrieval images in the image database according to a support degree and confidence index, and performing image retrieval according to the association rule similarity between the to-be-retrieved image and the retrieval images in the image database. According to the remote sensing image retrieval method and the remote sensing image retrieval device, the association rule of the remote sensing image is extracted from the constructed transition set; and then the association rule similarity between the to-be-retrieved image and the retrieval images in the image database is determined, thereby realizing remote sensing image retrieval and improving remote image retrieval accuracy.

Description

A kind of remote sensing image retrieval method and device of rotation scaling translation invariance

Technical field

The present invention relates to Remote Sensing Image Retrieval technical field, more particularly to a kind of remote sensing figure of rotation scaling translation invariance As search method and device.

Background technology

Remote sensing image has image breadth big, the characteristics of presentation content is more and complicated, " the different spectrum of jljl " and " foreign matter is with spectrum " Phenomenon it is very universal, bring larger difficulty to the retrieval of remote sensing image.Video search is searched for special containing specifying in database Levy or the image with Similar content, video search (the Content-Based Image based on content of current main-stream Retrieval, CBIR) the combined imaging process of method energy, information retrieval, machine learning, computer vision, artificial intelligence etc. is many The knowledge in field, by the visual signature automatically extracted from image as presentation content description；At present, the shadow based on content As retrieval achieves substantial amounts of achievement in research.

Visual Feature Retrieval Process has important function in video search, can be divided into two research directions, and one is research shadow The extraction of the low-level visual features such as spectrum, texture, the shape of picture and measuring similarity, including being carried based on curve of spectrum Absorption Characteristics The Hyperspectral imaging for taking is retrieved, color characteristic is extracted using color space, color moment, become using wavelet transformation, Contourlet Change, the method such as Gabor wavelet, generalized gaussian model, Table describe image textural characteristics, based on pixel shape index, PHOG (Pyramid Histogram of Oriented Gradients, laminated gradient direction histogram) shapes and small echo gold The remote sensing image shape facility of word tower describes method etc..The application comparative maturity of this kind of low-level visual feature, but cannot describe The semantic information of description image, cognition of the retrieval result often with human brain to remote sensing image that it is provided have quite poor away from, and It is not entirely satisfactory.

For this problem, another research direction is to set up the mapping model of low-level visual feature and semanteme, in language Adopted level improves the accuracy rate of video search.Main results are included based on the semantic retrieving method of statistical learning, such as pattra leaves The Bayesian network of this sorter model context of co-text, Bayesian network and EM (greatest hope) parameter Estimation etc.；Based on language The search method of justice mark, such as linguistic index model, Concept Semantic distributed model；Based on GIS (GIS-Geographic Information System, Geographic Information System) auxiliary semantic retrieving method, such as using the sky of vector element in GIS data Between and attribute information guide the semantic method for giving；Based on ontological semantic retrieving method, such as view-based access control model object domain sheet Method, GeoIRIS of body etc..This kind of method can to a certain extent reflect semantic understanding mistake of the human brain for video search Journey, is the development trend of following video search with higher accuracy rate.But current semantic retrieving method is often excessively closed Note low-level visual feature and the building process of Semantic mapping model, have ignored species, the semanteme of adopted low-level visual feature The factors such as learning method, eventually affect the precision ratio of semantic retrieval.

In recent years, human visual perception characteristic is introduced in video search field, is widely paid close attention to, but this kind of Method is still in the starting stage, and also many problems have to be solved：Physiology course such as human visual system, more meet human eye and regard The character description method of feel, bottom-up sensor model, notable feature are extracted and tolerance, top-down vision noticing mechanism Etc..In addition, mainly including Switzerland's RSIAII+III projects for the typical achievement of remote sensing image data retrieval, research is based on light The description and retrieval of the multi-resolution remote sensing image database of spectrum and textural characteristics；The original of Berkeley digital libraries exploitation Type system Blobworld, it, as data source, is allowed with aviation image, USGS orthographies and topographic map, SPOT satellite images etc. User can intuitively improve retrieval result；(RS) 2I projects of Nanyang Technological University, its research contents covers distant Sense image feature extracts the numerous aspects with the design of description, multi-dimensional indexing technology and distributed architecture；Stanford University SIMPLIcity, is determined using a kind of sane general area matching process (Integrated Region Matching, IRM) Similarity between adopted image, in Remote Sensing Image Retrieval of the satellite based on correlation rule good result is obtained；Microsoft grinds in Asia Study carefully the iFind of institute, the markup information constructing semantic network that system passes through image, and in relevant feedback with the visual signature of image Combine, have effectively achieved the relevant feedback on two levels.These systems achieve important achievement, but whether exist Feature extraction still remains a need for further further investigation in terms of characteristic features selection.

In sum, whether based on pixel or OO method for retrieving image, image is all focused on mostly whole The statistical information of the low-level features such as color, texture, the shape of body or local or subject area.It is directly based upon the retrieval of low-level feature Method cannot extract target interested, lack the ability being described to image space information, and existing characteristics dimension is too high, retouch State that imperfect, accuracy is poor, shortage is regular, the shortcomings of there is semantic gap in feature interpretation and human cognitive.At the same time, base Lack the theoretical and method of maturation again in the Remote Sensing Image Retrieval of high-layer semantic information.Between low-level feature and high-layer semantic information " semantic gap ", hinder the development and application of Remote Sensing Image Retrieval.

The content of the invention

The invention provides the remote sensing image retrieval method and device of a kind of rotation scaling translation invariance, it is intended at least exist One of above-mentioned technical problem of the prior art is solved to a certain extent.

In order to solve the above problems, the invention provides following technical scheme：

A kind of remote sensing image retrieval method of rotation scaling translation invariance, comprises the following steps：

Step a：Every width image in Image Database is carried out after image conversion process, the marginal point pixel of every width image is extracted, Build every width image transaction set；

Step b：The correlation rule in every width image is extracted by every width image transaction set；

Step c：The pass of the image to be retrieved and every width image in Image Database is calculated according to support and confidence indicator The regular similarity of connection, according to the correlation rule similarity of every width image in the image to be retrieved and Image Database image inspection is carried out Rope.

The technical scheme that the embodiment of the present invention is taken also includes：It is described by every width image in Image Database in step a Carry out image conversion process to be specially：Every width image is carried out into Radon conversion and Fourier-Mellin conversion；The Radon Transformation for mula is：

P (r, θ)=R (r, θ) f (x, y)=∫ ∫ f (x, y) δ (r-x cos θ-y sin θs) dxdy

In above-mentioned formula, the distance of | r | expression round dots to straight line, θ ∈ [0, π] represent the angle between straight line and y-axis, δ (r) is Dirac functions；

The Fourier-Mellin transformation for mula is：

In above-mentioned formula, u is real variable, and σ is the real constant more than 0.

The technical scheme that the embodiment of the present invention is taken also includes：In step a, every width image is through Radon After conversion and Fourier-Mellin conversion, the translating rotation of the image is phase place, and scale conversion is amplitude, Radon conversion and Fourier-Mellin conversion after function be：

Function Z (u, k) has the width and height as raw video, and the function is constant with rotating, scaling Property.

The technical scheme that the embodiment of the present invention is taken also includes：Step a also includes：To Radon conversion and Every width image after Fourier-Mellin conversion carries out pixel grayscale compression.

The technical scheme that the embodiment of the present invention is taken also includes：It is described to calculate the similar of correlation rule in step c The calculation of degree is：For all correlation rules of every width Extraction of Image, every correlation rule support is taken advantage of with confidence level Product is joined together, and constitutes a rule vector；By image relatively more to be retrieved and the rule vector of every width image in Image Database Similarity realizes video search；The measure formulas of the regular vector similarity are：

In above-mentioned formula, N is the correlation rule quantity of image, and r1 and r2 is respectively two rule vectors, μ₁And μ₂For two The average of width image, if two rule vectors are closer to while the average of two width images is closer to then the value of D is less, two width The similarity of image is higher.

Another technical scheme that the embodiment of the present invention is taken is：A kind of remote Sensing Image Retrieval of rotation scaling translation invariance Device, including：

Image conversion module：For the every width image in Image Database to be carried out into image conversion process；

Transaction set builds module：For extracting the marginal point pixel of every width image, every width image transaction set is built；

Correlation rule extraction module：For extracting the correlation rule in every width image by every width image transaction set；

Similarity calculation module：For being calculated in the image to be retrieved and Image Database according to support and confidence indicator The correlation rule similarity of every width image；

Video search module：For according to the correlation rule similarity of every width image in the image to be retrieved and Image Database Carry out video search.

The technical scheme that the embodiment of the present invention is taken also includes：The image conversion module enters every width image in Image Database Row image conversion process is specially：Every width image is carried out into Radon conversion and Fourier-Mellin conversion；The Radon becomes Changing formula is：

P (r, θ)=R (r, θ) f (x, y)=∫ ∫ f (x, y) δ (r-x cos θ-y sin θs) dxdy

In above-mentioned formula, the distance of | r | expression round dots to straight line, θ ∈ [0, π] represent the angle between straight line and y-axis, δ (r) is Dirac functions；

The Fourier-Mellin transformation for mula is：

In above-mentioned formula, u is real variable, and σ is the real constant more than 0.

The technical scheme that the embodiment of the present invention is taken also includes：Every width image is converted and Fourier- through Radon After Mellin conversion, the translating rotation of the image is phase place, and scale conversion is amplitude, Radon conversion and Fourier-Mellin Function after conversion is：

Function Z (u, k) has the width and height as raw video, and the function is constant with rotating, scaling Property.

The technical scheme that the embodiment of the present invention is taken also includes image compression module, and it is right that the image compression module is used for Every width image after Radon is converted and Fourier-Mellin is converted carries out pixel grayscale compression.

The technical scheme that the embodiment of the present invention is taken also includes：The similarity calculation module calculates the similar of correlation rule The calculation of degree is：All correlation rules of the every width Extraction of Image in for Image Database, every correlation rule support with The product of confidence level is joined together, and constitutes a rule vector；By image relatively more to be retrieved and every width image in Image Database The similarity of regular vector realizes video search；The measure formulas of the regular vector similarity are：

In above-mentioned formula, N is the correlation rule quantity of image, and r1 and r2 is respectively two rule vectors, μ₁And μ₂For two The average of width image, if two rule vectors are closer to while the average of two width images is closer to then the value of D is less, two width The similarity of image is higher.

Relative to prior art, the beneficial effect that the embodiment of the present invention is produced is：The rotation scaling of the embodiment of the present invention The remote sensing image retrieval method and device of translation invariance carries out Radon conversion and Fourier-Mellin to remote sensing image first Conversion, eliminates image rotation, scaling, the impact of translation；Pixel is carried out to the image after Radon and Fourier-Mellin conversion Gray-scale compression, reduces the amount of calculation of association rule mining；By building transaction set, remote sensing image is extracted by transaction set Correlation rule, then compares the similarity between correlation rule, so as to realize the retrieval of remote sensing image；The invention provides a kind of It is feasible from low-level visual feature to high-layer semantic information realizing the new way based on the Remote Sensing Image Retrieval of correlation rule, carry The high accuracy rate of Remote Sensing Image Retrieval.

Description of the drawings

Fig. 1 is the flow chart of the remote sensing image retrieval method of the rotation scaling translation invariance of the embodiment of the present invention；

Fig. 2 is the grey scale pixel value schematic diagram on 4 directions of marginal point pixel；

Fig. 3 (a) is original remote sensing image, and Fig. 3 (b) is the remote sensing image after compression, and Fig. 3 (c) is rim detection image；

Fig. 4 scales the structural representation of the remote Sensing Image Retrieval device of translation invariance for the rotation of the embodiment of the present invention；

Fig. 5 is 23 width remote sensing image schematic diagrames；

Fig. 6 is before all kinds of search method rankings 5 average precision；

Fig. 7 is before all kinds of search method rankings 10 average precision；

Fig. 8 is before all kinds of search method rankings 15 average precision；

Fig. 9 is before all kinds of search method rankings 20 average precision；

Figure 10 is the average precision of all retrieval results of all kinds of search methods；

Figure 11 is the ensemble average precision ratio of all search methods.

Specific embodiment

In order that the objects, technical solutions and advantages of the present invention become more apparent, it is right below in conjunction with drawings and Examples The present invention is further elaborated.It should be appreciated that specific embodiment described herein is only to explain the present invention, not For limiting the present invention.

Fig. 1 is referred to, is the flow process of the remote sensing image retrieval method of the rotation scaling translation invariance of the embodiment of the present invention Figure.The remote sensing image retrieval method of the rotation scaling translation invariance of the embodiment of the present invention is comprised the following steps：

Step 100：Every width image in Image Database is carried out into Radon and Fourier-Mellin conversion；

In step 100, the embodiment of the present invention transform to another domain by the image will with rotation, scaling, translation, Its rotation, scaling, the impact of translation are eliminated, the correlation rule of image is then extracted again.First Radon conversion is carried out to image, Transformation for mula is as follows：

P (r, θ)=R (r, θ) f (x, y)=∫ ∫ f (x, y) δ (r-x cos θ-y sin θs) dxdy (1)

In formula (1), the distance of | r | expression round dots to straight line, θ ∈ [0, π] represent the angle between straight line and y-axis, δ R () is Dirac functions.Radon conversion is to be integrated f (x, y) along straight line r-x cos θ-y sin θ=0, to obtain Summing value of the optional position (r, θ) place f (x, y) along the straight line.

Then Fourier-Mellin conversion is carried out to the image after Radon is converted again, and Fourier-Mellin becomes Change formula as follows：

In formula (2), u is real variable, and σ is the real constant more than 0, and its value typically takes 0.5.

Through Radon conversion and Fourier-Mellin conversion, the rotation of image translates into phase place to image, and scales just Be converted to amplitude.Finally define with minor function：

The function has the width and height as raw video, therefore its value can be considered as into an image.The letter Number has rotation, scaling consistency.With reference to correlation rule, then with rotation, scaling, translation invariance, carry out on this basis Video search.

Step 200：Pixel grayscale compression is carried out to the every width image after Radon and Fourier-Mellin conversion；

In step 200, because gray level is excessive, the support of frequent item set can be very little, is unfavorable for extracting support The all sufficiently large correlation rule of degree and confidence level, therefore before rule digging is associated, needs first enter to image to be retrieved Row pixel grayscale is compressed, by image compression to be retrieved to a few gray level, to reduce the amount of calculation of association rule mining.

In embodiments of the present invention, the method for every width image pixel gray-scale compression is specially：By the function Z after conversion (u, k) is considered as an image, and its value is considered as gray value, image compression is carried out with neighboring mean value and variance, on remote sensing image Pixel in each 3*3 neighborhood, calculates the mean μ and standard deviation sigma of the neighborhood, then calculates the centre of neighbourhood pixel using following formula Gray level upon compression：

In formula (4), pixel original gray level centered on g, g' is the center pixel gray level after compression, and c is ratio Coefficient, span is between [0.1,0.5].With it, can be by image compression to be retrieved to 0,1,2 this 3 gray scales Level.

In an alternative embodiment of the invention, according to the gray level by the way of homogenous segmentations, each grey level quantization is arrived The scope of [1, G], specially：Using the method for average compression, all gray levels are evenly distributed in several gray levels,

Wherein maxG is original maximum gradation value, and G is maximum gray scale after compression, and G=8, ceil () are the letters that rounds up Number, g+1 is in order that the gray level of image is compressed to 1～8.

According to the gray level by the way of homogenous segmentations, by the scope of each gray-level quantization to [1, G], specifically, It is compressed using the method for linear segmented, the maximum gray scale gMax and minimal gray level gMin of image is calculated first, then The gray level after compression is calculated using following formula：

Wherein G be maximum gray scale, G=8.

Gray level after compression is more, then the amount of calculation for being associated rule digging is bigger, but between the pixel for reflecting Relation be closer to truly；Otherwise the gray level after compression is fewer, and the difference after compression between pixel can be less, is more unfavorable for Significant correlation rule is excavated, therefore selects a suitable number of greyscale levels extremely important.In the embodiment of the present invention most High-gray level series is chosen to be 8, and the compress mode for adopting is average compression, and compression formula is as follows：

In formula (7), maxG is original maximum gradation value, and G is maximum gray scale number after compression, and G=8, ceil () are Round up function, and g+1 is in order that the pixel grayscale of image to be retrieved is compressed to 1～8.

Step 300：The marginal point pixel that rear image is changed per amplitude variation is extracted, every width image transaction set is built；

In step 300, the building process of transaction set is described by taking original remote sensing image as an example, the structure of the image after conversion Process is similar.Remote sensing image is needed on the basis of pixel based on pixel in the data mining of remote sensing image Upper structure transaction set.Build transaction set method be：In units of neighborhood, made with the arrangement of all pixels gray value in the neighborhood For some affairs in transaction set, such as the gray value of 9 pixels may make up an affairs in 3*3 neighborhoods.So for The image of 100*100 pixel sizes, may be constructed the transaction set being made up of 98*98=9604 affairs, and each transaction packet contains 9 .Remote sensing image is bigger, then the affairs for constituting are more, and the transaction set of composition is bigger；The Xiang Yue that one transaction packet contains is more, then need The frequent item set of calculating is bigger, and amount of calculation is also bigger, it is therefore desirable to do certain restriction to the item number that each transaction packet contains.Examine The edge for considering remote sensing image contains substantial amounts of useful information, while edge has directionality, therefore the present invention is sharp first With the edge of canny operator extraction remote sensing images, 4 directions of marginal point pixel are then extracted, with 3 pictures on each direction Element of the plain gray value as an item.For a marginal point pixel, the grey scale pixel value on its 4 directions is concrete such as Fig. 2 It is shown, it is the grey scale pixel value schematic diagram on 4 directions of a marginal point pixel.

Because the marginal point pixel on remote sensing image accounts for the ratio very little of whole image, while each affairs only includes 3 , therefore amount of calculation can reduce significantly.Fig. 3 (a), Fig. 3 (b) and Fig. 3 (c) are specifically referred to, Fig. 3 (a) is original remote sensing Image, size is 128*128 pixels；Fig. 3 (b) is the remote sensing image after compression, is to show convenient, and each gray level is multiplied by 256/G, although from compressing image as can be seen that number of greyscale levels is contracted to 8 after compression, the content of remote sensing image does not have not too Big change；Fig. 3 (c) is rim detection image, and the quantity of marginal point for detecting is 1964, therefore the transaction set size for constituting is 1964*4=7856, each transaction packet contains 3 items.If the atural object in image relatively enriches, then the marginal point for detecting will More, then the transaction set for finally constituting will be bigger.Table 1 shows the partial transaction in the transaction set.

Subitem in the transaction set of table 1

Step 400：Correlation rule in every width image is extracted by every width image transaction set；

In step 400, after constructing every width image transaction set, because each affairs only includes 3 items in transaction set, Can concentrate from affairs and build three-dimensional data cube, on this basis, extract the correlation rule in image to be retrieved.Extract and close The method of connection rule is specifically included：

Firstly the need of calculating frequent item set；It is a very crucial step in association rule mining to calculate frequent item set, Its amount of calculation directly affects the amount of calculation of whole association rule mining.For the item being made up of 3 elements, [a b c] is set to, Discounting for support, its correlation rule for extracting has 12 shown in table 2 below：

The correlation rule that 23 items of table are generated

A=>b	A=>c	B=>c
			B=>a	C=>a	C=>b
Ab=>c	Ac=>b	Bc=>a
			A=>bc	B=>ac	C=>ab

Because front 6 correlation rules relate only to the relation between two elements, still it is not enough to express the interior of remote sensing image Hold, and the operand of association rule mining and follow-up Similarity Measure can be increased, thus the embodiment of the present invention using Apriori or The frequent item set algorithm of the Mining Association Rules such as person FP-Growth from build transaction set in extract satisfaction specify confidence level with 6 correlation rules of support.

For simplicity, the embodiment of the present invention only have selected three pixels near marginal point, specifically can be according to retrieval Demand is selected.In general, the pixel for selecting is more, then the Xiang Yue that each office includes is more, and amount of calculation is bigger, but It is that the physical significance that every correlation rule is shown is abundanter, retrieval precision can be higher.

Step 500：Image to be retrieved is calculated according to support and confidence indicator to associate with every width image in Image Database Regular similarity, according to the correlation rule similarity of every width image in image to be retrieved and Image Database video search is carried out；

In step 500, support reflects distribution frequency of the correlation rule in remote sensing image, and confidence level then reflects The credibility of the correlation rule, therefore the correlation rule can be represented in remote sensing shadow with the product of support and confidence level True specific gravity as in.

The similarity calculating method of correlation rule is：For some correlation rules that a width remote sensing image is obtained, per bar Correlation rule support is joined together with the product of confidence level, that is, constitute a rule vector, and with the regular vector this is described The content of remote sensing image.The remote sensing image same or like for content, its regular vector should be similar.

Simultaneously the visual characteristic of human visual system HVS can be described with Weber's law.According to Weber's law, HVS pair The sensitiveness that relative luminance changes is higher than absolute brightness change, and the average of remote sensing image reflects entirety of the human eye for image Impression, if the average of remote sensing image is μ before brightness change₁, R is the brightness change relative to background luminance, distant after brightness change The average of sense image can be expressed as μ₂=(1+R) μ₁, then expression formulaCan be used to weigh the remote sensing image change Overall feeling：

Therefore,Only the function of R, shows it is consistent with Weber's law, can express human eye for image is bright The reaction of degree.Work as μ₁With μ₂Closer to,1 is closer to, illustrates that similarity of the two width remote sensing images in brightness is higher.

If the correlation rule of image to be retrieved has N bars, in Image Database it is arbitrary retrieval image correlation rule have M bars, then with It is on the basis of the N bar correlation rules of image to be retrieved, the correlation rule that image is retrieved in Image Database is matching, two association rule The condition that then the match is successful is that the former piece of two correlation rules is identical with consequent difference.If the match is successful for two correlation rules, Then retain the support of the correlation rule and the product of confidence level；If matching is unsuccessful, the support of the correlation rule with The product of confidence level is set to 0.Therefore the retrieval image can equally generate the rule vector of a N-dimensional, by relatively more to be checked The rope image similarity vectorial with the N-dimensional rule of all images in Image Database, you can realize video search.

In the embodiment of the present invention, using Kullback-Leibler divergences first approximation distance as weigh two rules to The tolerance of similarity is measured, its expression formula is：

In formula (6), r1 and r2 is respectively two rule vectors, while considering human-eye visual characteristic, final phase It is like degree Measure Indexes：

If two rule vectors are closer to while the average of two width remote sensing images is closer to then the value of D is less, similar Degree is higher.

Fig. 4 is referred to, is the structure of the remote Sensing Image Retrieval device of the rotation scaling translation invariance of the embodiment of the present invention Schematic diagram.The remote Sensing Image Retrieval device of the rotation scaling translation invariance of the embodiment of the present invention includes image conversion module, shadow As compression module, transaction set build module, correlation rule extraction module, similarity calculation module and video search module.

Image conversion module is used to for the every width image in Image Database to carry out Radon and Fourier-Mellin conversion；Its In, the embodiment of the present invention by will have rotation, scaling, translation image transform to another domain, eliminate its rotation, scaling, The impact of translation, then extracts again the correlation rule of image.First Radon conversion is carried out to image, transformation for mula is as follows：

P (r, θ)=R (r, θ) f (x, y)=∫ ∫ f (x, y) δ (r-x cos θ-y sin θs) dxdy (1)

In formula (1), the distance of | r | expression round dots to straight line, θ ∈ [0, π] represent the angle between straight line and y-axis, δ R () is Dirac functions.Radon conversion is to be integrated f (x, y) along straight line r-x cos θ-y sin θ=0, to obtain Summing value of the optional position (r, θ) place f (x, y) along the straight line.

Then Fourier-Mellin conversion is carried out to the image after Radon is converted again, and Fourier-Mellin becomes Change formula as follows：

In formula (2), u is real variable, and σ is the real constant more than 0, and its value typically takes 0.5.

Through Radon conversion and Fourier-Mellin conversion, the rotation of image translates into phase place to image, and scales just Be converted to amplitude.Finally define with minor function：

The function has the width and height as raw video, therefore its value can be considered as into an image.The letter Number has rotation, scaling consistency.With reference to correlation rule, then with rotation, scaling, translation invariance, carry out on this basis Video search.

Image compression module is used to carry out pixel grayscale to the every width image after Radon and Fourier-Mellin conversion Compression；Wherein, because gray level is excessive, the support of frequent item set can be very little, is unfavorable for extracting support and confidence level All sufficiently large correlation rule, therefore before rule digging is associated, needs first carry out pixel grayscale to every width image Compression, will change rear image compression to a few gray level, to reduce the amount of calculation of association rule mining per amplitude variation.

In embodiments of the present invention, the method that rear image pixel gray-scale compression is changed per amplitude variation is specially：After conversion Function Z (u, k) is considered as an image, and its value is considered as gray value, image compression is carried out with neighboring mean value and variance, on image Pixel in each 3*3 neighborhood, calculates the mean μ and standard deviation sigma of the neighborhood, then calculates the centre of neighbourhood pixel using following formula Gray level upon compression：

In formula (4), pixel original gray level centered on g, g' is the center pixel gray level after compression, and c is ratio Coefficient, span is between [0.1,0.5].With it, original remote sensing image can be compressed to into 0,1,2 this 3 ashes Degree level.

In an alternative embodiment of the invention, according to the gray level gray-level quantization is arrived by the way of homogenous segmentations The scope of [1, G], specially：Using the method for average compression, all gray levels are evenly distributed in several gray levels,

Wherein maxG is original maximum gradation value, and G is maximum gray scale after compression, and G=8, ceil () are the letters that rounds up Number, g+1 is in order that the gray level of image is compressed to 1～8.

According to the gray level by the way of homogenous segmentations, by the scope of each gray-level quantization to [1, G], specifically, It is compressed using the method for linear segmented, the maximum gray scale gMax and minimal gray level gMin of image is calculated first, then The gray level after compression is calculated using following formula：

Wherein G be maximum gray scale, G=8.

Gray level after compression is more, then the amount of calculation for being associated rule digging is bigger, but between the pixel for reflecting Relation be closer to truly；Otherwise the gray level after compression is fewer, and the difference after compression between pixel can be less, is more unfavorable for Significant correlation rule is excavated, therefore selects a suitable number of greyscale levels extremely important.In the embodiment of the present invention most High-gray level series is chosen to be 8, and the compress mode for adopting is average compression, and compression formula is as follows：

In formula (7), maxG be original maximum gradation value, G for compression after maximum gray scale, G=8, ceil () be to Flow in upper plenum, g+1 is in order that the pixel grayscale of remote sensing image is compressed to 1～8.

Transaction set builds module to be used to extract the marginal point pixel that every amplitude variation changes rear image, builds image transaction set；Wherein, Build transaction set method be：In units of neighborhood, the arrangement of all pixels gray value is as in transaction set using in the neighborhood The gray value of 9 pixels may make up an affairs in some affairs, such as 3*3 neighborhoods.It is so big for 100*100 pixels Little image, may be constructed the transaction set being made up of 98*98=9604 affairs, and each transaction packet contains 9 items.Remote sensing image is got over Greatly, then the affairs for constituting are more, and the transaction set of composition is bigger；The Xiang Yue that one transaction packet contains is more, then calculative frequent episode Collection is bigger, and amount of calculation is also bigger, it is therefore desirable to do certain restriction to the item number that each transaction packet contains.In view of remote sensing image Edge contain substantial amounts of useful information, while edge has directionality, therefore the present invention is carried first with canny operators Take the edge of remote sensing image, then extract marginal point pixel 4 directions, using 3 grey scale pixel values on each direction as The element of one item.For a marginal point pixel, the grey scale pixel value on its 4 directions is concrete as shown in Fig. 2 being a side Grey scale pixel value schematic diagram on 4 directions of edge point pixel.

Because the marginal point pixel on remote sensing image accounts for the ratio very little of whole image, while each affairs only includes 3 , therefore amount of calculation can reduce significantly.Fig. 3 (a), Fig. 3 (b) and Fig. 3 (c) are specifically referred to, Fig. 3 (a) is original remote sensing Image, size is 128*128 pixels；Fig. 3 (b) is the remote sensing image after compression, is to show convenient, and each gray level is multiplied by 256/G, although from compressing image as can be seen that number of greyscale levels is contracted to 8 after compression, the content of remote sensing image does not have not too Big change；Fig. 3 (c) is rim detection image, and the quantity of marginal point for detecting is 1964, therefore the transaction set size for constituting is 1964*4=7856, each transaction packet contains 3 items.If the atural object in image relatively enriches, then the marginal point for detecting will More, then the transaction set for finally constituting will be bigger.

Correlation rule extraction module is used for the correlation rule changed in rear image transaction set extraction image by every amplitude variation；Its In, after constructing image transaction set, because each affairs only includes 3 items in transaction set, can concentrate from affairs and build three Dimension data cube, on this basis, extracts the correlation rule in image.The method for extracting correlation rule is specifically included：

Firstly the need of calculating frequent item set；It is a very crucial step in association rule mining to calculate frequent item set, Its amount of calculation directly affects the amount of calculation of whole association rule mining.For the item being made up of 3 elements, [a b c] is set to, Discounting for support, its correlation rule for extracting has 12 shown in table 2 below：

The correlation rule that 23 items of table are generated

A=>b	A=>c	B=>c
			B=>a	C=>a	C=>b
Ab=>c	Ac=>b	Bc=>a
			A=>bc	B=>ac	C=>ab

Because front 6 correlation rules relate only to the relation between two elements, still it is not enough to express the interior of remote sensing image Hold, and the operand of association rule mining and follow-up Similarity Measure can be increased, thus the embodiment of the present invention using Apriori or The frequent item set algorithm of the Mining Association Rules such as person FP-Growth from build transaction set in extract satisfaction specify confidence level with 6 correlation rules of support.

For simplicity, the embodiment of the present invention only have selected three pixels near marginal point, specifically can be according to retrieval Demand is selected.In general, the pixel for selecting is more, then the Xiang Yue that each office includes is more, and amount of calculation is bigger, but It is that the physical significance that every correlation rule is shown is abundanter, retrieval precision can be higher.

Similarity calculation module is used to calculate image to be retrieved with retrieval in Image Database according to support and confidence indicator The correlation rule similarity of image；Wherein, support reflects distribution frequency of the correlation rule in remote sensing image, and confidence level The credibility of the correlation rule is then reflected, therefore the correlation rule can be represented with the product of support and confidence level and existed True specific gravity in remote sensing image.

The similarity calculating method of correlation rule is：For some correlation rules that a width remote sensing image is obtained, per bar Correlation rule support is joined together with the product of confidence level, that is, constitute a rule vector, and with the regular vector this is described The content of remote sensing image.The remote sensing image same or like for content, its regular vector should be similar.

Simultaneously the visual characteristic of human visual system HVS can be described with Weber's law.According to Weber's law, HVS pair The sensitiveness that relative luminance changes is higher than absolute brightness change, and the average of remote sensing image reflects entirety of the human eye for image Impression, if the average of remote sensing image is μ before brightness change₁, R is the brightness change relative to background luminance, distant after brightness change The average of sense image can be expressed as μ₂=(1+R) μ₁, then expression formulaCan be used to weigh the remote sensing image change Overall feeling：

Therefore,Only the function of R, shows it is consistent with Weber's law, can express human eye for image is bright The reaction of degree.Work as μ₁With μ₂Closer to,1 is closer to, illustrates that similarity of the two width remote sensing images in brightness is got over It is high.

Video search module is used to be carried out with the correlation rule similarity of retrieval image in Image Database according to image to be retrieved Video search；If the correlation rule of image to be retrieved has N bars, in Image Database it is arbitrary retrieval image correlation rule have M bars, then with It is on the basis of the N bar correlation rules of image to be retrieved, the correlation rule that image is retrieved in Image Database is matching, two association rule The condition that then the match is successful is that the former piece of two correlation rules is identical with consequent difference.If the match is successful for two correlation rules, Then retain the support of the correlation rule and the product of confidence level；If matching is unsuccessful, the support of the correlation rule with The product of confidence level is set to 0.Therefore the retrieval image can equally generate the rule vector of a N-dimensional, by relatively more to be checked The rope image similarity vectorial with the N-dimensional rule of all images in Image Database, you can realize video search.

In the embodiment of the present invention, using Kullback-Leibler divergences first approximation distance as weigh two rules to The tolerance of similarity is measured, its expression formula is：

In formula (9), r1 and r2 is respectively two rule vectors, while considering human-eye visual characteristic, final phase It is like degree Measure Indexes：

If two rule vectors are closer to while the average of two width remote sensing images is closer to then the value of D is less, similar Degree is higher.

In order to verify effectiveness of the invention, following Remote Sensing Image Retrieval experiment has been carried out：

By selecting totally 23 width remote sensing images (size is 300*300) to make from QuickBird and WorldView-2 Image Databases On the basis of image, it is concrete as shown in figure 5, being 23 width remote sensing image schematic diagrames.Wherein Q1～Q8 represents QuickBird images, W1 ～W15 represents WorldView-2 images, presentation content comprising road, sparse wood, settlement place, thick forest ground, open ground, waters, Meadow, farmland etc..

Rotation Remote Sensing Image Retrieval experiment

First 23 width benchmark images are turned clockwise, every 15 degree of rotations once, 552 width images, Ran Houcong is obtained This 552 width picture centre intercepts subgraph of the size for 128*128 pixels, used as final image library.The rotation of the present invention is contracted The remote sensing image retrieval method and existing auto-relativity function method, Gabor wavelet for being laid flat motion immovability is converted, DT-CWT (Dual- Tree Complex Wavelet Transform) and the retrieval result of NSCT search methods contrasted.Arranging support is 0.015, confidence level is 0.3, extracts the correlation rule of all images, is then input into an image to be retrieved, calculates figure to be retrieved Similarity in picture and image library between the correlation rule of all images, and according to similar size front 24 width image is taken as retrieval As a result.Count respectively in all kinds of search method image returning results 5 before ranking, front 10, front 15, front 20 width image and all retrievals As a result average precision, and the ensemble average precision ratio of all images, concrete as shown in Fig. 6 to Figure 11, Fig. 6 is all kinds of inspections 5 average precision before Suo Fangfa rankings, Fig. 7 is before all kinds of search method rankings 10 average precision, and Fig. 8 is all kinds of retrievals 15 average precision before method ranking, Fig. 9 is before all kinds of search method rankings 20 average precision, and Figure 10 is all kinds of retrievals The average precision of all retrieval results of method, Figure 11 is the ensemble average precision ratio of all search methods.Can from Figure 11 Go out, the average precision of the remote sensing image retrieval method of the rotation scaling translation invariance of the present invention is 99.77%, and auto-correlation Function method, Gabor wavelet conversion, the average precision of DT-CWT and NSCT are only 32.71%, 83.38%, 66.94% and 71.74%.The remote sensing image retrieval method for illustrating the rotation scaling translation invariance of the embodiment of the present invention has very high inspection Suo Jingdu.

The remote sensing image retrieval method and device of the rotation scaling translation invariance of the embodiment of the present invention is first to remote sensing shadow As carrying out Radon conversion and Fourier-Mellin conversion, image rotation, scaling, the impact of translation are eliminated；To Radon and Image after Fourier-Mellin conversion carries out pixel grayscale compression, reduces the amount of calculation of association rule mining；Pass through Transaction set is built, the correlation rule of remote sensing image is extracted by transaction set, then compare the similarity between correlation rule, so as to Realize the retrieval of remote sensing image；The invention provides it is a kind of it is feasible from low-level visual feature to high-layer semantic information realizing base In the new way of the Remote Sensing Image Retrieval of correlation rule, the accuracy rate of Remote Sensing Image Retrieval is improve.

The foregoing description of the disclosed embodiments, enables professional and technical personnel in the field to realize or using the present invention. Various modifications to these embodiments will be apparent for those skilled in the art, as defined herein General Principle can be realized in other embodiments without departing from the spirit or scope of the present invention.Therefore, the present invention The embodiments shown herein is not intended to be limited to, and is to fit to and principles disclosed herein and features of novelty phase one The most wide scope for causing.

Claims (10)

1. a kind of rotation scales the remote sensing image retrieval method of translation invariance, it is characterised in that comprise the following steps：

Step a：Every width image in Image Database is carried out after image conversion process, the marginal point pixel of every width image is extracted, is built Every width image transaction set；

Step b：The correlation rule in every width image is extracted by every width image transaction set；

Step c：The image to be retrieved is calculated according to support and confidence indicator and associates rule with retrieval image in Image Database Then similarity, according to the correlation rule similarity of every width image in the image to be retrieved and Image Database video search is carried out.

2. rotation according to claim 1 scales the remote sensing image retrieval method of translation invariance, it is characterised in that in institute It is described every width image in Image Database is carried out into image conversion process to be specially in stating step a：Every width image is carried out into Radon changes Change and Fourier-Mellin conversion, the Radon transformation for mula is：

P (r, θ)=R (r, θ) f (x, y)=∫ ∫ f (x, y) δ (r-xcos θ-ysin θ) dxdy

In above-mentioned formula, the distance of | r | expression round dots to straight line, θ ∈ [0, π] represent the angle between straight line and y-axis, δ (r) It is Dirac functions；

The Fourier-Mellin transformation for mula is：

$M(u,k)={\∫}_0^{\mathrm{\∞}}{\∫}_0^{2\mathrm{\π}}P(r,\mathrm{\θ})\·r^{\mathrm{\σ}-iu-1}\·\exp (-ik\mathrm{\θ})drd\mathrm{\θ}$

In above-mentioned formula, u is real variable, and σ is the real constant more than 0.

3. rotation according to claim 2 scales the remote sensing image retrieval method of translation invariance, it is characterised in that in institute In stating step a, after Radon conversion and Fourier-Mellin conversion, the translating rotation of the image is every width image Phase place, scale conversion is amplitude, and the function after Radon conversion and Fourier-Mellin conversion is：

$Z(u,k)=M{(0,0)}^{-\frac{\mathrm{\σ}-iu+1}{\mathrm{\σ}+1}}{e}^{ik\arg \left(M\right(0,1))}M(u,k)$

Function Z (u, k) has the width and height as raw video, and the function has rotation, scaling consistency.

4. rotation according to claim 3 scales the remote sensing image retrieval method of translation invariance, it is characterised in that described Step a also includes：Pixel grayscale compression is carried out to the every width image after Radon conversion and Fourier-Mellin conversion.

5. rotation according to claim 1 scales the remote sensing image retrieval method of translation invariance, it is characterised in that in institute In stating step c, the calculation of the similarity for calculating correlation rule is：For the relevant rule of institute of every width Extraction of Image Then, every correlation rule support is joined together with the product of confidence level, constitutes a rule vector；By shadow relatively more to be retrieved As realizing video search with the similarity of the rule vector of every width image in Image Database；The tolerance of the regular vector similarity is public Formula is：

$D=\underset{i=1}{\overset{N}{\Σ}}\frac{{\left(r1\right(i)-r2(i\left)\right)}^2}{r1\left(i\right)+r2\left(i\right)}+(1-\frac{2{\mathrm{\μ}}_1{\mathrm{\μ}}_2}{{\mathrm{\μ}}_1^2+{\mathrm{\μ}}_2^2})$

In above-mentioned formula, N is the correlation rule quantity of image, and r1 and r2 is respectively two rule vectors, μ₁And μ₂For two width shadows The average of picture, if two rule vectors are closer to while the average of two width images is closer to then the value of D is less, two width images Similarity it is higher.

6. a kind of rotation scales the remote Sensing Image Retrieval device of translation invariance, it is characterised in that include：

Image conversion module：For every width image in Image Database to be carried out into image conversion process；

Transaction set builds module：For extracting the marginal point pixel of every width image, every width image transaction set is built；

Correlation rule extraction module：For extracting the correlation rule in every width image by every width image transaction set；

Similarity calculation module：For being calculated in the image to be retrieved and Image Database per width according to support and confidence indicator The correlation rule similarity of image；

Video search module：For being carried out with the correlation rule similarity of every width image in Image Database according to the image to be retrieved Video search.

7. rotation according to claim 6 scales the remote Sensing Image Retrieval device of translation invariance, it is characterised in that described Every width image in Image Database is carried out image conversion process and is specially by image conversion module：Every width image is carried out into Radon conversion With Fourier-Mellin conversion；The Radon transformation for mula is：

P (r, θ)=R (r, θ) f (x, y)=∫ ∫ f (x, y) δ (r-xcos θ-ysin θ) dxdy

In above-mentioned formula, the distance of | r | expression round dots to straight line, θ ∈ [0, π] represent the angle between straight line and y-axis, δ (r) It is Dirac functions；

The Fourier-Mellin transformation for mula is：

$M(u,k)={\∫}_0^{\mathrm{\∞}}{\∫}_0^{2\mathrm{\π}}P(r,\mathrm{\θ})\·r^{\mathrm{\σ}-iu-1}\·\exp (-ik\mathrm{\θ})drd\mathrm{\θ}$

In above-mentioned formula, u is real variable, and σ is the real constant more than 0.

8. rotation according to claim 7 scales the remote Sensing Image Retrieval device of translation invariance, it is characterised in that described After Radon conversion and Fourier-Mellin conversion, the translating rotation of the image is phase place to every width image, and scale conversion is Amplitude, Radon is converted and is with the function after Fourier-Mellin conversion：

$Z(u,k)=M{(0,0)}^{-\frac{\mathrm{\σ}-iu+1}{\mathrm{\σ}+1}}{e}^{ik\arg \left(M\right(0,1))}M(u,k)$

Function Z (u, k) has the width and height as raw video, and the function has rotation, scaling consistency.

9. rotation according to claim 8 scales the remote Sensing Image Retrieval device of translation invariance, it is characterised in that also wrap Image compression module is included, the image compression module is used for the every width shadow after Radon conversion and Fourier-Mellin conversion As carrying out pixel grayscale compression.

10. rotation according to claim 6 scales the remote Sensing Image Retrieval device of translation invariance, it is characterised in that institute State similarity calculation module and calculate the calculation of similarity of correlation rule and be：Every width Extraction of Image in for Image Database Correlation rule, every correlation rule support is joined together with the product of confidence level, constitutes a rule vector；Treated by comparing Retrieval image realizes video search with the similarity of the rule vector of every width image in Image Database；The regular vector similarity Measure formulas are：

$D=\underset{i=1}{\overset{N}{\Σ}}\frac{{\left(r1\right(i)-r2(i\left)\right)}^2}{r1\left(i\right)+r2\left(i\right)}+(1-\frac{2{\mathrm{\μ}}_1{\mathrm{\μ}}_2}{{\mathrm{\μ}}_1^2+{\mathrm{\μ}}_2^2})$

In above-mentioned formula, N is the correlation rule quantity of image, and r1 and r2 is respectively two rule vectors, μ₁And μ₂For two width shadows The average of picture, if two rule vectors are closer to while the average of two width images is closer to then the value of D is less, two width images Similarity it is higher.