CN104077770A - Plant leaf image local self-adaption tree structure feature matching method - Google Patents

Abstract

The invention relates to a plant leaf image local self-adaption tree structure feature matching method. A DOG feature point matching substrate is constructed, feature line structures are extracted between matched feature points and unmatched feature points of the substrate, and spatial distribution limitation between local feature descriptor information and the feature points can be combined in a self-adaption mode through the feature line structures to form a dynamic reference relational tree. The problem that the good matching result can be difficult to obtain under the conditions of zooming, rotation combination, local high-repetition modes and the like under high noise points when plant leaves are irregularly deformed is effectively solved, the high-robustness plant leaf image feature matching method is achieved, and the adaptability and the accuracy meet the actual application requirement. Meanwhile, the plant leaf image local self-adaption tree structure feature matching method can be easily combined with a large number of existing feature descriptors based on pictures, various feature matching methods with different features and based on images are expanded, and the plant leaf image local self-adaption tree structure feature matching method is of great significance in the typical problem in computer visions.

Description

A kind of leaf image local auto-adaptive tree structure feature matching method

Technical field

The invention belongs to area of pattern recognition, relate to computer graphics techniques field, relate in particular to a kind of leaf image local auto-adaptive tree structure feature matching method.

Background technology

Image Feature Matching is the corresponding relation of setting up two unique point set in picture, is the requisite base support section of application such as Image Mosaics, scene Recognition, image retrieval, three-dimensional modeling.Compare with the industrial part having compared with regular geometric shapes, under state of nature, the surface profile of plant and resemblance are all more complicated.When processing feature is mated, the performance of above-mentioned difficult point is also more outstanding.Blade as one of important plant organ, it is carried out based on Image Feature Matching, specifically may be summarized to be following several difficult point: (1) plant leaf blade surface has obvious local similar feature, such as symmetry of texture, leaf color and space distribution etc.High Local Gravity And complex pattern is like this difficult point in existing Feature Correspondence Algorithm always; (2), in natural shooting situation, plant leaf blade itself has significantly irregular around deformation such as volume, distortions.Due to the scrambling of deformation, brought difficulty to the unified constraint of Matching Model and concrete feature description; (3) under state of nature, the actual photographed environment of leaves of plants portion image is easily subject to the impact of the factors such as focal length, shooting angle, so actual while carrying out characteristic matching, often to be processed is the combinatorial problem of the noise spot that mixes in a large number and rotation, scale transformation, and this brings difficulty to equally coupling work.As above the feature point selection method that a few class difficult points cause leaves of plants portion Image Feature Matching not only to need, also needs to have more the feature matching method of robustness and accuracy.

In sum, existing leaves of plants portion Image Feature Matching method is due to for being all the object under specific shooting condition mostly, there is the features such as illumination is even, noise interference is little, locus conversion is simple, blade surface deformation is relatively regular, therefore also come with some shortcomings, be mainly manifested in: (1) faces the local irregularities of plant leaf blade around volume, twist distortion, existing method is difficult to build unified model and describes it, cannot introduce effective space geometry constraint rule and improve characteristic matching precision; (2) in the situation that cannot introducing regular geometric constraint, the similarity of simple use local description is processed matching problem, is difficult to obtain satisfied matching result; (3) when processing much noise point mixes and rotates the combination of scale transformation, the precision of coupling is difficult to improve.

Summary of the invention

For solving the problems such as the strong noise point interference introduced in the actual photographed process existing in prior art and image rotation, convergent-divergent combined transformation, the object of the present invention is to provide a kind of leaf image local auto-adaptive tree structure feature matching method, can carry out flexibly characteristic matching, thereby solve in actual characteristic matching application, existing method is difficult to obtain the problem of matched well in these problems.

To achieve these goals, leaf image local auto-adaptive tree structure feature matching method provided by the invention, built the substrate of DOG Feature Points Matching, utilize substrate to extract characteristic curve structure between the unique point of having mated and do not mate, by characteristic curve structure can adaptive combination local description information and the space distribution of unique point retrain, form dynamic referring-to relation tree, specifically comprise the following steps:

Step S1: build the multi-resolution Gaussian difference pyramid substrate of leaf image, the unique point having extracted is mapped in Gaussian difference substrate;

Step S2: in the situation that step S1 completes unique point mapping, combining space information is set up out one with reference to triangular structure in unique point set;

Step S3: utilize the reference triangular structure of step S2, calculate local yardstick information and the rotation information of blade, according to this information architecture partial transformation model, complete the spatial division to unique point;

Step S4: certain point carries out the growth coupling of characteristics tree from the reference triangle of step S3, expands to unknown point from known point is constantly adaptive.By unique point top-stitching feature and space distribution in conjunction with building in local feature description, substrate, screen, until complete whole tree structure coupling;

Step S5: repeat the process of S2-S4, until filter out with reference to triangular structure in the unique point that cannot never mate, coupling completes.

Wherein, the concrete steps of step S1 are as follows:

Step S11: pass through formula:

$L=(x,y,\mathrm{\δ})=G(x,y,\mathrm{\δ})\⊗I(x,y)$

Calculate Gauss's metric space image, wherein expression is to the convolution operation in x and y direction, and G (x, y, δ) is a Gaussian function that becomes yardstick.

Step S12: the gaussian kernel convolved image of adjacent two different scales is subtracted each other, obtain Gauss's error image D (x, y, δ).

Step S13: introduce adjacent two metric space multiple constant k adjacent picture is similar to sampling, pass through formula:

$D(x,y,\mathrm{\δ})=(G(x,y,\mathrm{k\δ})-G(x,y,\mathrm{\δ}))\⊗I(x,y)$

Obtaining pyramid Gauss error image, is wherein 1.5 or 2 with image drop sampling multiplying power k.

Wherein, the concrete steps of step S2 are as follows:

Step S21: obtain the unique point set of former figure and the unique point set of target figure, random one section of continuous unique point subset of extraction from the unique point set of former figure, utilize local feature description's, mate with the continuous unique point subset of randomly drawing in the set of target signature point successively.

Step S22: utilize ratio screening method to form candidate matches set on a small scale, therefrom randomly draw three pairs of unique points, further verify the correctness of coupling by the similarity of three pairs of characteristic edge forming.If there are all authentication faileds of two characteristic edge on triangle, delete two unique points that limit is folded, from the set of candidate matches unique point, random choosing is put continuing coupling again, until three pairs of characteristic edge are all by checking, thereby formation is with reference to triangular structure.

Wherein, the concrete steps of step S3 are as follows:

Step S31: utilize with reference to leg-of-mutton constraint, unique point to be matched is divided according to space distribution situation.

Step S32: suppose that in figure, unique point distribution to be matched is relatively even, according to calculate best mesh segmentation parameter with reference to triangle, then complete unique point distributing position to the mapping of the mesh space of coarseness.

Wherein, the concrete steps of step S4 are as follows:

Step S41: establish A ^sfor in the set of former figure unique point with reference on triangle a bit, make the tree root into whole coupling tree.

Step S42: from A ^sset out, with radius r ^branchin scope, select arbitrarily a some D ^s, utilize the distance vector in former figure and characteristic curve in target figure, find out a qualified match point D ^t.

Step S43: with D in former figure ^scentered by, at it, do not comprise D ^sin eight neighborhoods of the grid scope at place, find out lower some E ^s, the relative Corresponding matching position E that obtains in target figure ^vir, with E ^vircentered by extract adjacent mesh, all unique points of taking out in grid are alternative features point.

Step S44: utilize the characteristic curve Policy Filtering of local feature and dynamic construction to go out to select optimal match point, as E ₁ ^t.Then, the unique point E newly to mate ₁ ^tfor new originating point continues coupling, expanded search and the new unique point of coupling, until complete the coupling work of whole coupling tree.

The present invention has effectively avoided existing method the convergent-divergent under irregular deformation, strong noise point occurs and rotate the problem that is difficult to obtain matched well result in the situations such as combination, local high repeat pattern at plant leaf blade, realized the leaves of plants portion Image Feature Matching method of high robust, its adaptability and higher accuracy has flexibly reached the requirement of practical application.Meanwhile, the present invention is easy to be combined with existing a large amount of Feature Descriptors based on picture, expands all kinds of feature matching methods based on image with different qualities, for target following, scene modeling, Image Mosaics, the classical problem in the computer visions such as image retrieval has great significance.

Accompanying drawing explanation

Fig. 1 is the method flow diagram of leaf image local auto-adaptive tree structure feature matching method.

Fig. 2 is the phoenix tree leaves with irregular curling and deformation.

Fig. 3 is the 1st, 3,5,7 layers of feature substrate stratal diagram of Chinese parasol leaf sheet.

Fig. 4 is feature point extraction and the mark figure of Chinese parasol leaf sheet.

Fig. 5 is the unique point subdivision of Chinese parasol leaf sheet and builds with reference to triangle.

Fig. 6 is the irregular deformation of Chinese parasol leaf sheet and the tree-like coupling under rotation.

Fig. 7 is the Feature Points Matching line graph of Chinese parasol leaf sheet.

Fig. 8 is the coupling under the high repeat pattern of maple blade and rotation.

Embodiment

Below in conjunction with accompanying drawing, describe each related detailed problem in the technology of the present invention method in detail.Be to be noted that described embodiment is only intended to be convenient to the understanding of the present invention, and it is not played to any restriction effect.

A leaf image local auto-adaptive tree structure feature matching method, its flow process as shown in Figure 1.This feature matching method has built the substrate of DOG Feature Points Matching, utilize substrate to extract characteristic curve structure between the unique point of having mated and do not mate, by characteristic curve structure can adaptive combination local description information and the space distribution of unique point retrain, form dynamic referring-to relation tree.

Specifically comprise the following steps:

Step S1: build the poor pyramid substrate of multi-resolution Gaussian of leaf portion image.

The phoenix tree leaves with irregular curling and deformation of take is as shown in Figure 2 example, and the unique point having extracted is mapped in basis space: first calculate Gauss's metric space image wherein expression is to the convolution operation in x and y direction, and G (x, y, δ) is a Gaussian function that becomes yardstick.The image subtraction of the gaussian kernel convolution by two adjacent different scales, draws Gauss's error image D (x, y, δ), then introduces adjacent two metric space multiple constant k adjacent picture is similar to sampling, by $D(x,y,\mathrm{\δ})=(G(x,y,\mathrm{k\δ})-G(x,y,\mathrm{\δ}))\⊗I(x,y)$ Satisfied convergent-divergent and deformation requirements, select suitable pyramid level according to specific needs.The pyramid response image finally obtaining.Wherein, image drop sampling multiplying power constant k value is 1.5 or 2.

As shown in Figure 3 select this down-sampled pyramid of 7 floor heights, during k=1.5, the substrate that phoenix tree leaves is the 1st, 3,5,7 layers.

Step S2: never in matching characteristic point, screening is set up with reference to triangular structure.

Complete in the situation of unique point mapping, obtain the unique point set of former figure and the unique point set of target figure, random one section of continuous unique point subset of extraction from the unique point set of former figure, utilize local feature description's, mate with the continuous unique point subset of randomly drawing in the set of target signature point successively.The feature point extraction of Chinese parasol leaf sheet as shown in Figure 4 and mark figure.

Unique point set to be matched is divided into the subset that granularity is less, and the rapid screening by simple ratioing technigue between subclass goes out candidate matches point, randomly draws the candidate point that 3 spaces approach relatively and Feature Descriptor successfully mates.Finally, use three limits of the route characteristic diabolo building in the substrate of Gaussian difference pyramid judge and screen, if there are all authentication faileds of two characteristic edge on triangle, delete two unique points that limit is folded, from the set of candidate matches unique point, random choosing is put continuing coupling again, until the three pairs of characteristic edge are all by checking, thereby form with reference to triangular structure.Finally obtain a stable part with reference to triangular structure.

The unique point subdivision of Chinese parasol leaf sheet as shown in Figure 5 and building with reference to triangle.

Step S3: utilize feature triangle to build local convergent-divergent and rotation matching model, complete the subdivision to space characteristics point.

Utilization is with reference to three couples of summit (A corresponding to triangle ^s, B ^s, C ^s) → (A ^t, B ^t, C ^t) calculate former figure to the local scaling of target figure, according to local zooming parameter, obtain coupling basalis pair best in characteristic matching, follow-up feature line extraction all from then on feature basalis extracts, to have overcome the matching problem in convergent-divergent situation.

Calculate with reference to triangle centre of form O separately ^sand O ^t, further obtain local feature growth reference direction with in the point search matching process in conjunction with local space, when running into direction of search judgement, all by growth reference direction more separately, judge and screen, effectively to overcome the Rotation in coupling.

Initialization pantograph ratio and rotary reference direction, suppose that unique point distributes relatively even, by regular grid, even subdivision carried out in unique point distributed areas.By a lattice structure is set, all unique point numberings are mapped in corresponding grid, facilitate follow-up space querying fast and search relatively.Division principle is carried out in the space of unique point is: the feature as far as possible reducing in each cell is counted, and improves matching speed; The enough cell size of suitable reservation, to increase the discrimination of characteristic edge, increases the differentiation degree of single match.By mesh mapping, unique point will more easily be carried out Local Search and inquiry in follow-up space search coupling, has avoided a large amount of erroneous matching simultaneously.Tree-like coupling under the irregular deformation of Chinese parasol leaf sheet as shown in Figure 6 and rotation.

Step S4, from reference to certain point triangle, utilizes tree-like growth matching strategy to carry out the growth coupling of characteristics tree.

If A ^sfor on the reference triangle extracting from former figure a bit, making it is the tree root of whole coupling tree.From A ^sset out, with radius r ^branchin scope, select arbitrarily a some D ^s, utilize the unique point vector in former figure and the characteristic curve forming in substrate between unique point filter out qualified match point D in target figure ^t.

With D in former figure ^scentered by, at it, do not comprise D ^sin eight neighborhoods of the grid scope at place, find out lower some E ^s, the relative Corresponding matching position E that obtains in target figure ^vir, with E ^vircentered by extract adjacent mesh, all unique points of taking out in grid are alternative features point.

Half r with the regular grid length of side ^leaffor radius, draw circle, extract the grid crossing with circle, all unique points of taking out in grid are alternative unique point E to be matched _n ^t, utilize local feature description's and dynamic Characteristic constraint line to filter out optimal match point E ₁ ^t, with the unique point E newly mating ₁ ^tfor new originating point, continue Downward match.With the whole region of News Search decision search of this increment type expansion, until whole coupling set all couplings that can expand.

The Feature Points Matching line graph of Chinese parasol leaf sheet as shown in Figure 7.

Step S5, Iterative matching.

Repeated execution of steps S2～S4, reject the unique point that has completed coupling, in remaining point, continue to find with reference to triangular structure, set up new Local Search Matching Model and search for, until cannot extract the reference triangular structure making new advances, now coupling finally completes.Coupling under the high repeat pattern of maple blade as shown in Figure 8 and rotation.

The above; be only the embodiment in the present invention, but protection scope of the present invention is not limited to this, any people who is familiar with this technology is in the disclosed technical scope of the present invention; can manage conceivable conversion and remodeling, within also should being considered as covereding in protection scope of the present invention.

Claims (5)

1. a leaf image local auto-adaptive tree structure feature matching method, it is characterized in that first having built the substrate of DOG Feature Points Matching, utilize substrate to extract characteristic curve structure between the unique point of having mated and do not mate, by characteristic curve structure can adaptive combination local description information and the space distribution of unique point retrain, form dynamic referring-to relation tree, specifically comprise the following steps:

Step S1: build the multi-resolution Gaussian difference pyramid substrate of leaf image, the unique point having extracted is mapped in Gaussian difference substrate;

Step S2: in the situation that step S1 completes unique point mapping, combining space information is set up out one with reference to triangular structure in unique point set;

Step S3: utilize the reference triangular structure of step S2, calculate local yardstick information and the rotation information of blade, according to this information architecture partial transformation model, complete the spatial division to unique point;

Step S4: certain point carries out the growth coupling of characteristics tree from the reference triangle of step S3, expands to unknown point from known point is constantly adaptive.By unique point top-stitching feature and space distribution in conjunction with building in local feature description, substrate, screen, until complete whole tree structure coupling;

Step S5: repeat the process of S2-S4, until filter out with reference to triangular structure in the unique point that cannot never mate, coupling completes.

2. leaf image local auto-adaptive tree structure feature matching method according to claim 1, is characterized in that: described step S1 concrete steps are as follows:

Step S11: pass through formula:

$L=(x,y,\mathrm{\δ})=G(x,y,\mathrm{\δ})\⊗I(x,y)$

Calculate Gauss's metric space image, wherein expression is to the convolution operation in x and y direction, and G (x, y, δ) is a Gaussian function that becomes yardstick.

Step S12: the gaussian kernel convolved image of adjacent two different scales is subtracted each other, obtain Gauss's error image D (x, y, δ).

Step S13: introduce adjacent two metric space multiple constant k adjacent picture is similar to sampling, pass through formula:

$D(x,y,\mathrm{\δ})=(G(x,y,\mathrm{k\δ})-G(x,y,\mathrm{\δ}))\⊗I(x,y)$

Obtaining pyramid Gauss error image, is wherein 1.5 or 2 with image drop sampling multiplying power k.

3. leaf image local auto-adaptive tree structure characteristic point matching method according to claim 1, is characterized in that: described step S2 concrete steps are as follows:

Step S21: obtain the unique point set of former figure and the unique point set of target figure, random one section of continuous unique point subset of extraction from the unique point set of former figure, utilize local feature description's, mate with the continuous unique point subset of randomly drawing in the set of target signature point successively.

Step S22: utilize ratio screening method to form candidate matches set on a small scale, therefrom randomly draw three pairs of unique points, further verify the correctness of coupling by the similarity of three pairs of characteristic edge forming.If there are all authentication faileds of two characteristic edge on triangle, delete two unique points that limit is folded, from the set of candidate matches unique point, random choosing is put continuing coupling again, until three pairs of characteristic edge are all by checking, thereby formation is with reference to triangular structure.

4. leaf image local auto-adaptive tree structure feature matching method according to claim 1, is characterized in that: the concrete steps of described step S3 are as follows:

Step S31: utilize with reference to leg-of-mutton constraint, unique point to be matched is divided according to space distribution situation.

Step S32: suppose that in figure, unique point distribution to be matched is relatively even, according to calculate best mesh segmentation parameter with reference to triangle, then complete unique point distributing position to the mapping of the mesh space of coarseness.

5. leaf image local auto-adaptive tree structure feature matching method according to claim 1, is characterized in that: the concrete steps of described step S4 are as follows:

Step S41: establish A ^sfor in the set of former figure unique point with reference on triangle a bit, make the tree root into whole coupling tree.

Step S42: from A ^sset out, with radius r ^branchin scope, select arbitrarily a some D ^s, utilize the distance vector in former figure and characteristic curve in target figure, find out a qualified match point D ^t.

Step S43: with D in former figure ^scentered by, at it, do not comprise D ^sin eight neighborhoods of the grid scope at place, find out lower some E ^s, the relative Corresponding matching position E that obtains in target figure ^vir, with E ^vircentered by extract adjacent mesh, all unique points of taking out in grid are alternative features point.

Step S44: utilize the characteristic curve Policy Filtering of local feature and dynamic construction to go out to select optimal match point, as E ₁ ^t.Then, the unique point E newly to mate ₁ ^tfor new originating point continues coupling, expanded search and the new unique point of coupling, until complete the coupling work of whole coupling tree.