CN106897724A - A kind of plant leaf identification method based on contour line shape facility - Google Patents

Abstract

Plant leaf blade identification is the hot issue in image procossing and pattern identification research field.But there is bottleneck problem in plant leaf blade identification, i.e., similar blade interior difference is big, and difference is small between inhomogeneity blade.The present invention proposes a kind of plant leaf identification method based on contour line shape facility-inside and outside chord length characteristic vector (Inner＆Outer Characteristic Vector) and attempts to solve this bottleneck problem to a certain extent, belongs to digital image processing techniques field.The length of each section of string of the target profile curve of extraction is divided into internal chord length and outside two parts of chord length by the present invention according to target region occupied in the picture.Internal chord length is intended to describe the convex characteristic of contour line, and outside chord length is then conceived to the recessed feature of Extract contour line, and the inside chord length of multiple yardsticks and outside chord length have respectively constituted the interior chord length matrix and outer chord length matrix of description shape.Interior chord length matrix and outer chord length matrix are pulled into one-dimensional vector respectively, combination couples together composition IOCV.The vision system of the mankind is very sensitive to shape convex-concave characteristic, cognitive using this, and IOCV proposed by the present invention will be greatly improved to the distinguishing ability of target shape.

Description

A kind of plant leaf identification method based on contour line shape facility

Technical field

The present invention relates to a kind of description method of leaf image contour line shape facility, belong to digital image processing techniques field.

Background technology

The present invention is the feature extraction based on binary digital image, and the leaf image collected in reality often RGB true color images, need for true color image to be converted to black and white binary image that (shape area pixel value is 1, and background area pixels value is for 0).Obtain binary digital image and generally include following steps：

1st, successively by the digital picture that the scanned instrument generation euchroic of original analog leaf image is color, then digital picture is sent in computer by connecting mouth, obtains and complete whole image, build RGB leaf image database；

2nd, coloured image f (x, y) is converted to gray level image, is, according to 3 color component relations of red, green, blue, to be converted into YUV black and white gray scale color model by the very color color color models of RGB and obtain h (x, y),

H (x, y)=0.299*R+0.587*G+0.144*B (1)

3rd, by Two-peak method threshold process, according to the difference of background, adjust automatically threshold value, the image after normalization is to be converted into bianry image g (x, y) to gray level image h (x, y), and T is threshold value.The gray value of each pixel in image is compared with threshold value, shape area is split from background, to carry out the boundary tracking of next step.

4th, the tracking of longest edge circle is carried out to bianry image g (x, y).The bianry image of closing of the frontier can extract border line coordinates with Contour tracing algorithm.The core of Edge tracking of binary image extraction algorithm is exactly to empty target internal point.If a stain in target image, when its 8 connected region consecutive points are all stain, then the point is deleted.The basic ideas of tracking are：First from top to bottom, data image is scanned line by line from left to right.The first stain for scanning as profile first boundary point, and savepoint coordinate (x, y).Then choose along upper left side as initial search direction.If top point is stain, then record saves as boundary point, otherwise rotates 45 degree according to the direction of search clockwise.Then traveled through according to above-mentioned steps, until finding first for stain.This stain is then set to new boundary point.It is rotated by 90 ° counterclockwise on current search direction.Search by that analogy preserves next stain coordinate, untill persistently traveling through until finding initial boundary point.If there is the boundary line of multiple target objects in picture, we are given tacit consent to it is contemplated that the border of the curve most long in image.

The method of Shape Classification is divided into two kinds, and a kind of is the method based on shape area, and another kind is the method based on shaped wheel profile.Method based on region needs to obtain all information of image pixel, and amount of calculation and amount of storage are all than larger；Method based on contour line only needs to obtain the Pixel Information of shaped wheel profile, and amount of calculation is small, easy extraction and storage, and description of construction more compacts.And profile wire shaped most discrimination, human vision can just make a distinction often through contour of object wire shaped to it.In sum, the description method of shaped wheel profile has more advantage.

In order to effectively describe profile wire shaped, there has been proposed a series of classical contour line shape descriptors.The sorting technique of contour line shape descriptor has a lot, and relatively conventional two methods are：(1) difference based on description son description scope, is divided into global description's and local description；(2) difference based on description subrepresentation space, is divided into spatial domain description and frequency domain describes son.Table 1 gives some classical contour line shape description methods according to the sorting technique of spatial domain and frequency domain.

Table 1. some classical contour line shape description methods

And wherein Fourier descriptor (Fourier Descriptor) is a kind of classical shape description method based on contour line.Two-dimensional shapes contour line is expressed as one-dimensional profile line function by the method first, also known as shapes ignature (shape signature), the characteristic vector of description shape is constituted with the coefficient of the Fourier transform of the function.FD calculates the characteristics of simple, definition is clear, energy is concentrated to low frequency, is with one of widest shape descriptor.

It is proposing in recent years as follows to the constitution step of description profile wire shaped similar herein MDM description：

Step 1：Following the trail of maximum boundary the shaped wheel profile for extracting carries out uniform sampling, and sampled point number is N, constitutes sampling point set C={ P₁, P₂..., P_N, P_i=(x_i, y_i)；

Step 2：Calculate point centrostigma P_iHere it is Euclidean distance to remaining the N-1 chord length distance of point, obtains L₁, L₂..., L_NSection chord length, as the i-th row of matrix D, allows i to change to n from 1 successively；

Each row of step 3, spin matrix D so that first element of each row is 0.Structural matrix D in this way_m, D_mThe first row element be all 0 because its N number of point for representing is to the distance of its own, the second row element represents the distance between two neighboring point；

Step 4, the every a line ascending order arrangement to Dm, produce matrix D_ms.Due to the matrix D after sequence_msIt is symmetrical on (n/2) row, to avoid redundancy, choose matrix D_msThe second row to (n/2), construct MDM matrixes.

It is a kind of conventional effective ways that profile wire shaped is portrayed using chord length.Human visual system is very sensitive for the convex-concave characteristic of profile wire shaped, is the very effective feature for target identification.Cognitive based on this, the distinguishing ability for extracting the inside and outside chord length characteristic vector of profile wire shaped convex-concave characteristic will be greatly improved.Every section of string of contour line is fallen and be divided into two parts in the inside and outside of shape area by the present invention according to it, and the chord length inside shape area that falls is referred to as interior chord length, and the chord length outside shape area that falls is referred to as outer chord length.Interior chord length is used for describing the convex feature of shaped wheel profile, and outer chord length is used for describing the recessed characteristic of contour line.And multiple yardsticks upper contour line convex-concave characteristic of level is portrayed, can be from details to entirety and thick to carefully excavating contour line shape facility.Inside and outside chord length is effectively combined just can be from part to the convex-concave characteristic for globally extracting profile wire shaped.In sum, it may be considered that the distinguishing ability for extracting description of profile wire shaped convexo-concave characteristic will be greatly enhanced, so that difference is big in class in the identification of solution plant leaf blade to a certain extent, the small bottleneck problem of class inherited.

The content of the invention

The technical problem to be solved in the present invention is to propose a kind of shape descriptor that can extract profile wire shaped available characteristic-convex-concave characteristic, so as to more efficiently be described to profile wire shaped, improves the discrimination of plant leaf blade.

The present invention uses following technical scheme：

True color RGB image is converted into bianry image, the digital picture of binaryzation can be represented with following form：

Here x, y represent the transverse and longitudinal coordinate of target image shape.D is target shape distributed areas in the picture.F ' represents the shaped wheel profile for extracting.Point set C={ P are constituted by uniform sampling counterclockwise to the profile line boundary most long that tracking is obtained_i(x_i, y_i), i=1,2 ..., N, N takes even number here, because target profile curve is closure, here P_N+i=P_i.To a point P1 (x on contour line₁, y₁), along mobile s segmental arc counterclockwise to point P2 (x₂, y₂), its chord length is represented with equation below：

Falling to cut string shape area is inside and outside according to the string.The length l of the part that the string falls in shape area^(l)Represent, be called interior chord length；The length l fallen outside shape area^(O)Represent, be called outer chord length.Make x₁=min { x₁, x₂, x₂=max { x₁, x₂, y₁=min { y₁, y₂, y₂=max { y₁, y₂, then l^(l)And l^(O)Can be represented with following form：

l^(O)=l-l^(l) (6)

Here A=(y₂-y₁)/l, B=(x₁-x₂)/l, C=(x₁y₂-y₁x₂)/l.δ () represents Dirac functions.It is pointed out that Ax+By+C=0 is point P here₁With point P₂Connected normal form of a straight line equation.Accompanying drawing 1 is the schematic diagram of inside and outside chord length, and the part of solid marks represents interior chord length, and the part of dashed lines labeled represents outer chord length.

Point set C midpoint P of the present invention to shaped wheel profile uniform sampling₁(x₁, y₁) and point P₂(x₂, y₂) calculating of the inside and outside chord length of single chord length that is connected, specifically include following steps：

Step A, use formulaCalculate point P₁With point P₂Between chord length, i.e., with Euclidean distance, this measures to represent the chord length between 2 points；

Step B, will respectively calculate (x₁-x₂) and (y₁-y₂) value that obtains takes absolute value and distinguish assignment and X and Y, takes the greater and assignment and N in X and Y；

Step C, the size for comparing X and Y, if X >=Y, perform step D, otherwise, perform step J；

Step D, compare x₁And x₂Size, if x₁＞ x₂, then step E is performed, otherwise, perform step F；

Step E, x is exchanged respectively₁、x₂And y₁、y₂Size, even if point P₁Represent P₁、P₂The less pixel of transverse and longitudinal scale value in 2 points, point P₂Represent the larger pixel of abscissa value；

Step F, i is entered as 1, and performs step G；

Step G, calculating λ=i/ (N-i),Calculate (y₁+λ*y₂)/(1+ λ) and round assignment withWhether step H, i adds 1 and assignment and i, and judges i more than N-1, if it is not, step G is then performed, if so, then performing step N；

Step I, compare y₁And y₂Size, if y₁＞ y₂, then step J is performed, otherwise, perform step K；

Step J, y is exchanged respectively₁、y₂And x₁、x₂Size, even if point P₁Represent P₁、P₂The less pixel of vertical scale value, point P in 2 points₂Represent the larger pixel of ordinate value；

Step K, i is entered as 1, and performs step L；

Step L, calculating λ=i/ (N-i),Calculate (x₁+λ*x₂)/(1+ λ) and round assignment with

Whether step M, i adds 1 and assignment and i, and judges i more than N-1, if it is not, then performing step L；If so, then performing step N；

Step N, by x₂Assignment withy₂Assignment withK is entered as 0, i and is entered as 1；

Step O, calculating k=k+f (x_i, y₁)；

Whether step P, i adds 1 and assignment and i, and judges i more than N, if it is not, step O is then performed, if so, then performing step Q；

Step Q, calculating r=k/N, return to interior chord length l^(I)=r*l, outer chord length l^(o)=(1-r) * l.

Order point P₁Subscript change to every bit on N, i.e. contour line all as a starting point from 1, calculate starting point P_i(i=1,2 ..., N) moves s isometric segmental arc point of arrival P counterclockwise_i+sCorresponding string.By calculating each section of inside and outside chord length of string, we obtain interior chord length sequenceWith outer chord length sequenceIt is pointed out that because contour line is closure, any one section of string all corresponds to arc length and be respectively s unit length and N-s the two of unit length sections of arcs, to avoid data redundancy, we allow arc length s values 1,2 ..., M=N/2.Thus two matrixes of M × N, interior chord length matrix ICM and outer chord length matrix OCM are obtained.Inside and outside chord length matrix is pulled into one-dimensional vector respectively, inside and outside chord length characteristic vector (IOCV) of just constitute that two combination of eigenvectors being got up.

Here segmental arc number is exactly s, represents yardstick level.Accompanying drawing 3 for contour line (128 sampled points) 5 yardsticks level inside and outside chord length an example (from left figure to right figure, yardstick level be followed successively by s=4,8,16,32).If observed with the short segmental arc that small yardstick observes blade wheel profile, i.e. contouring line, it can be found that they are mostly convex.The inside and outside chord length that leftmost figure (correspondence yardstick s=4) in the figure is marked has reacted this visual characteristic of contour line well, and in the figure, only least a portion of string has the part (marked with dotted line) outside shape area.And the figure (s=8 is observed using big yardstick, 16,32), when i.e. the big segmental arc of contouring line is observed, it can be found that the recessed characteristic of contour line, s=8 in the figure, 16,32 three figures, have considerable string section or fully fall in the outside of shape area, and this characteristic is consistent with the Visual Observations Observations of people.

Contour line shape description method proposed by the present invention, interior chord length and outer chord length two parts are divided into by single chord length.Interior chord length describes the convex characteristic of profile wire shaped, and outer chord length describes the recessed characteristic of profile wire shaped.Human visual system is very sensitive to the convex-concave characteristic of profile wire shaped, therefore contour line convex-concave characteristic is very effective contour line shape recognition feature.Cognitive based on this, the distinguishing ability for extracting the inside and outside chord length characteristic vector of profile wire shaped convex-concave characteristic will be greatly enhanced.Compared to the existing similar multiple dimensioned distance matrix (Multiscale Distance Matrix) merely with chord length Extracting contour shape, the profile wire shaped convex-concave characteristic for extracting more distinguishing ability proposed by the present invention, inside and outside chord length characteristic vector (Inner＆Outer Characteristic Vector) of construction, contour line style characteristic can be more efficiently portrayed, recognition capability is stronger.

Brief description of the drawings

Fig. 1 is the direct feature recognition FB(flow block) of typical image；

Fig. 2 is dimensionality reduction property data base construction FB(flow block) in leaf image identification, and the numeral of circles mark represents the execution step of construction dimensionality reduction property data base；

Fig. 3 is inside and outside chord length schematic diagram, and wherein solid line represents interior chord length, and dotted line represents outer chord length；

Fig. 4 for contour line (128 sampled points) 5 yardsticks level inside and outside chord length an example (from left figure to right figure, yardstick level be followed successively by s=4,8,16,32).

Fig. 5 is 15 class plant sample black white images in Swedish blade databases.

Fig. 6 is 100 class plant sample black white images in Leaf100 blade databases.

Fig. 7 is 12 blade black white images of similar plant species in Leaf100

Specific embodiment

Technical scheme is described in detail below in conjunction with the accompanying drawings：

Accompanying drawing 1 is typical image dimensionality reduction identification process block diagram, and wherein dotted line frame represents existing similar contour line shape description method, and solid box is institute's contour line shape description method of the present invention.Apply the present invention to two leaf image databases, respectively classical Swedish blades database and we oneself from the Leaf100 databases of field acquisition, and be compared to the discrimination of similar multiple dimensioned distance matrix.Specifically according to following steps：

The discrimination of step 1, present invention experiment measuring and calculating plant leaf blade, therefore blade database is divided into two subsets of training set and test set.

A leaf image is concentrated in step 2, input test：

Step 201, judge input leaf image whether be RGB true color images, if so, then perform step 202, if it is not, then perform step 203；

Step 202, the very color color color models of RGB are converted into YUV black and white gray scale color model using formula (1) obtain h (x, y)；

Step 203, judge input leaf image whether be bianry image, if then perform step 3, if it is not, then perform step 204；

Step 204, gray level image h (x, y) are converted into bianry image by Two-peak method threshold process, such as shown in formula (2).

Step 3, two-value leaf image can be expressed as formula (3) form, and carry out maximum boundary tracking to bianry image, extract shaped wheel profile.Uniform sampling is carried out to shaped wheel profile, point set C={ P are constituted_i, i=1,2 ..., N.Here we take N=128；

Step 4, make l_siRepresent the point P from profile_i(x_i, y_i) set out, another point P is reached in the counterclockwise direction_i+s(x_i+s, y_i+s) length of the string of segmental arc passed through.Calculate string l_siThe length that the string for being connected falls inside shape area and the length fallen outside shape area, use interior chord length respectivelyWith outer chord lengthRepresent.

For two point P in point set C₁(x₁, y₁) and P₂(x₂, y₂), calculating interior chord lengths and outer chord length of the single chord length for being connected at 2 points of, can be calculated by following specific steps：

Step 401, use formulaCalculate point P₁With point P₂Between chord length, i.e., with Euclidean distance, this measures to represent the chord length between 2 points；

Step 402, will respectively calculate (x₁-x₂) and (y₁-y₂) value that obtains takes absolute value and distinguish assignment and X and Y, takes the greater and assignment and N in X and Y；

Step 403, the size for comparing X and Y, if X >=Y, perform step 404, otherwise, perform step 410；

Step 404, compare x₁And x₂Size, if x₁＞ x₂, then step 405 is performed, otherwise, perform step 406；

Step 405, x is exchanged respectively₁、x₂And y₁、y₂Size, even if point P₁Represent P₁、P₂The less pixel of transverse and longitudinal scale value in 2 points, point P₂Represent the larger pixel of abscissa value；

Step 406, i is entered as 1, and performs step 407；

Step 407, calculating λ=i/ (N-i),Calculate (y₁+λ*y₂)/(1+ λ) and round assignment with

Whether step 408, i add 1 and assignment and i, and judge i more than N-1, if it is not, step 407 is then performed, if so, then performing step 414；

Step 409, compare y₁And y₂Size, if y₁＞ y₂, then step 410 is performed, otherwise, perform step 411；

Step 410, y is exchanged respectively₁、y₂And x₁、x₂Size, even if point P₁Represent P₁、P₂The less pixel of vertical scale value, point P in 2 points₂Represent the larger pixel of ordinate value；

Step 411, i is entered as 1, and performs step 412；

Step 412, calculating λ=i/ (N-i),Calculate (x₁+λ*x₂)/(1+ λ) and round assignment with

Whether step 413, i add 1 and assignment and i, and judge i more than N-1, if it is not, then performing step 412；If so, then performing step 414；

Step 414, by x₂Assignment withy₂Assignment withK is entered as 0, i and is entered as 1；

Step 415, calculating k=k+f (x_i, y_i)；

Whether step 416, i add 1 and assignment and i, and judge i more than 414, if it is not, step O is then performed, if so, then performing step 417；

Step 417, calculating r=k/N, return to interior chord length l^(l)=r*l, outer chord length l^(o)=(1-r) * l.

Step 5, fixed s, make i change to N from 1, and we obtain the N sections of string being all corresponding to s unit segmental arc, i.e. every bit on contour line as a starting point.By calculating each section of inside and outside chord length of string, we obtain interior chord length sequenceWith outer chord length sequenceIt is pointed out that because contour line is closure, any one section of string all corresponds to arc length and be respectively s unit length and N-s the two of unit length sections of arcs, to avoid data redundancy, we allow arc length s values 1,2 ..., M=N/2.Here N=128, M=64, thus obtain the matrix of two 64 × 128, are referred to as interior chord length matrix ICM and outer chord length matrix OCM；

Step 6, inside and outside chord length characteristic vector ICM and OCM is set to meet translation, selection and flexible geometric invariance.Specifically include following steps：

Step 601, here chord length are calculated with the Euclidean distance between 2 points, and Euclidean distance is relative distance, it is easy to prove that the inside and outside chord length of string keeps constant.Therefore, the interior chord length matrix and outer chord length matrix of target shape have the inherent translation invariance to target；

Step 602, when target shape is rotated, the length value of inside and outside chord length will not change.But rotation can make the initial point that contour line is used change, when inside and outside chord length characteristic vector is calculated, can make matrix that cyclic shift occurs per a line.As the row of matrix i-th, the element of j row move to the i-th row, j+t is arranged, wherein 1≤t≤n is displacement.

Fourier descriptor is a kind of classical description method based on contour shape, and rotation normalizing is carried out using the method, can retain the relative position relation between matrix row element.Specific practice is as follows：Regard matrix as one-dimensional discrete signal per a line, and carry out one dimensional fourier transform, original row is replaced with the mould of Fourier transform coefficient.One-dimensional discrete signal Fourier transform is as follows.

Here occluding contour the point set { (x for sampling_i, y_i) an one-dimensional sequence of complex numbers { C (i) } can be regarded as, wherein we make C_i=x_i+jy_i, it is expressed as { C (i) }, i=1,2 ..., 128；

Step 603, when target shape occur stretching be that object function f (x, y) is changed into f (α x, α y), if α ＞ 0 be zoom factor.So inside and outside chord lengthWith outer chord lengthIt is changed intoWithWe zoom in and out normalization operation, respectively-A ,-M ,-RA ,-R using four kinds of methods, and (1) carries out normalizing with the maximum of whole matrix to each element of matrix；(2) normalizing is carried out to each element of matrix with the average value of whole matrix；(3) to each element of every a line of matrix, normalizing is carried out with the maximum of the row；(4) to each element of every a line of matrix, normalizing is carried out to it with the average value of the row.It is global yardstick normalizing method that first two is used, and latter two uses local scale normalizing method.It is to be noted that, if there is for making the situation that the maximum or average of normalization operation are 0, such as one its outer chord length matrix of convex shaped wheel profile is 0 matrix, then the row of the matrix or matrix has met the consistency of scaling, it is not necessary to perform normalization operation.It should also be noted that we are with same scaling method for normalizing, and internally outer chord length characteristic vector ICM and OCM are normalized operation respectively.

Step 7, step 6 is normalized after interior chord length matrix ICM_NWith outer chord length matrix OCM_NW and 1-w is multiplied by respectively.Here w represents weight coefficient, for adjusting contribution of the inside and outside chord length in target identification；

Step 8, two matrix by rowss are pulled into one-dimensional vector respectively, the new one-dimensional vector of composition that two vectors are combined, the vectorial dimension is N².Referred to as inside and outside chord length characteristic vector (IOCV), can be represented with following form：

Step 9, for test set in each width leaf image, extract every N of leaf image in test set according to step 2 to step 8²Dimension one-dimensional vector.C represents plant leaf blade sample number in test set, and K represents each class sample Leaf picture number, and test set Leaf sum is S=C × K.Used as matrix a line, the one-dimensional vector that all leaf images are extracted combines and constitutes a S × N one-dimensional vector that each width leaf image is extracted so in test set²Matrix, referred to as property data base；

The dimension of step 10, the authentication information in order to further extract target shape and compressive features space, traditional linear discriminent analysis (Linear Discriminant Analysis) feature dimension reduction method is applied to the matrix extracted in step 9 by the present invention, construction dimensionality reduction property data base, idiographic flow block diagram such as accompanying drawing Fig. 2.The dimension reduction space of one (C-1) dimension is obtained, has N per in one-dimensional²Individual value (N=128)；

Step 11, the S × N for obtaining step 9 extraction²Matrix projection in the dimension reduction space of step 10, obtain the matrix an of S × (C-1) size, referred to as dimensionality reduction database；

Step 12, the leaf image T that will be used for test, perform step 2 to step 8, equally extract N²Dimensional vector；

Step 13, the one-dimensional vector that step 12 is extracted projects to the dimension reduction space in step 10, obtain the row vector of one (C-1) dimension；

The each row of step 14, the one-dimensional row vector for obtaining step 13 successively with matrix in dimensionality reduction database carries out Diversity measure, it is pointed out that every a line of matrix represents the width leaf image X in test set in dimensionality reduction database_i, i=1,2 ..., S.Diversity measure formula is as follows：

Here parameter w is weight coefficient, and for adjusting the contribution of interior chord length and outer chord length in target identification, N is that the number of contour line sampled point is 128, M=N/2=64；

Step 15, selection the jth width leaf image minimum with test image T difference values.The minimum calculating of difference value can be represented with equation below：

R_j=min (D_i), i=1,2 ..., S=C × K (8)

Class where calculating the minimum leaf image of selected difference value, and judge whether to belong to same class with the plant leaf blade of test.This recognition methods is referred to as arest neighbors matching (1NN)；

Step 16, w is entered as 0.50, adjusts the value for increasing w, 0.05 is increased every time, until w=1.00.Often adjust and once just repeat step 2 to 15.Report knows highest discrimination and obtains the value of weight w corresponding during highest discrimination.

In order to verify the effect of the inventive method, following experiment has been carried out：

1st, experiment condition：

Confirmatory experiment is carried out on one computer, and the allocation of computer is Intel's dual core processor (2300 megahertzs) and 4096 Mbytes of random access memory (RAM), and programming language is Matlab (7.10 version).

2nd, experimental technique：

Part in figure shown in solid box is replaced the part shown in dotted line frame by experiment using primary image identification framework (as shown in Figure 1).This experiment is carried out on two leaf image storehouses, respectively classical Swedish leaf images storehouse (as shown in Figure 3) and we oneself from the Leaf100 leaf images storehouse (as shown in figs. 4 and 5) of field acquisition.Because this experiment is that the identification for carrying out plant leaf blade is classified, therefore two leaf image storehouses need to be respectively classified into two subsets of training set and test set.There are 25 class plant samples in Swedish blade databases, 75 leaf images are included in every class plant sample, we take preceding 50 leaf images in every class plant sample, i.e., 1250 leaf image composing training collection, remaining 625 leaf images constitute test set.There are 100 class plant samples in Leaf100 blade databases, 12 blade figures are included in every class plant sample, we take preceding 6 leaf images in every class plant sample, i.e., 600 leaf image composing training collection, remaining 600 leaf images constitute test sets；

Training process：Successively to the leaf image gray processing in training set, binaryzation is then carried out, maximum boundary tracking is carried out to the leaf image after binaryzation.To each point of shaped wheel profile uniform sampling 128 for extracting.With the point sampled, to concentrate each point set be initial point, calculates the interior chord length and outer chord length of the string that the point of s segmental arc of rotate counterclockwise is connected, and obtains interior chord length sequenceWith outer chord length sequenceHere s represents the segmental arc number that the line of 2 points of connection is passed through, and is defined as yardstick level, and yardstick level is integer.S is allowed to change to M=N/2, chord length matrix ICM and outer chord length matrix OCM in the interior chord length sequence and outer chord length Sequence composition of different scale level from 1.Internally outer chord length characteristic vector is normalized respectively, it is met translation, rotation and flexible consistency.Two normalized matrix by rowss are pulled into one-dimensional vector, and combines that (dimension is N to one new one-dimensional vector of composition²).The N that each width leaf image is extracted in training set²Dimension one-dimensional vector combines constitutive characteristic database.To extract the authentication information of target shape and the dimension in compressive features space, traditional characteristic dimension reduction method LDA is applied in extracted property data base, obtains dimension reduction space.By matrix projection in property data base to dimension reduction space, the compression dimensionality reduction database of more distinguishing ability is extracted.

Test process：A width leaf image in test set extracts N according to the flow in training process²The one-dimensional characteristic vector of dimension, carries out Diversity measure with each row of matrix in dimensionality reduction database successively, it is pointed out that every a line of matrix all corresponds to the width leaf image in test set.Leaf image corresponding to the row matrix minimum with the width image difference opposite sex is the image most like with leaf image for testing that be identifying.

3rd, the evaluation index of experimental result：

Experimental result calculates discrimination using arest neighbors matching (1NN).

Comprise the following steps that：

M width leaf images in correspondence test set, retest process.With to identify otherness minimum, i.e., whether the class where the most similar leaf image is same plant sample class for class where being relatively used for the leaf image tested every time.If so, then by O_i=1；If it is not, then by O_i=0.Therefore O={ O_iConstitute the one-dimensional vector that a value is for 0 or 1, i=1,2 ..., m

The formula for calculating discrimination is as follows：

Here sum () represents all elements sum in vector O.

4 and the contrast and experiment of prior art：

Table 2 gives to table 3 and extracted on the Leaf100 leaf images storehouse of classical Swedish leaf images storehouse and our oneself collection respectively chord length characteristic vector inside and outside blade profile wire shaped and describe sub, multiple dimensioned distance description and FD description, and the discrimination after application LDA Feature Dimension Reductions.It is noted herein that, LDA being applied to FD description and takes after Fourier transform (C-1) individual coefficient before one-dimensional functions, C represents the species number of applied leaf image storehouse Leaf.C one dimensional fourier transform coefficient of preceding 2 value is taken for the interference for abating the noise, in experiment.

Wherein-M ,-A ,-RM ,-RA, represent four kinds of method for normalizing respectively, and be followed successively by (1) carries out normalizing with the maximum of whole matrix to each element of matrix；(2) normalizing is carried out to each element of matrix with the average value of whole matrix；(3) to each element of every a line of matrix, normalizing is carried out with the maximum of the row；(4) to each element of every a line of matrix, normalizing is carried out to it with the average value of the row.W represents weight parameter, and expression is contribution of the interior chord length matrix in identification, then 1-w then represents contribution of the outer chord length in identification.Given in table under specific certain method for normalizing, reach the size that highest discrimination is interior chord length matrix contribution.

The discrimination (%) that the LDA dimensionality reduction features that table 2 carry out on Swedish blade databases compare

The discrimination (%) that the LDA dimensionality reduction features that table 3 carry out on Leaf100 blade databases compare

Can be observed by table 2, in Swedish blade databases, the method for MDM and IOCV is superior to the method for FD, and-M, IOCV descriptions describes son and is higher by 6.93 percentage points, 4.4 percentage points, 3.83 percentage points and 2 percentage points than MDM respectively in four kinds of method for normalizing of-A ,-RM ,-RA.Can be so that it has been observed that in Leaf100 blade databases, the method for FD is better than MDM methods, in four kinds of method for normalizing of-M ,-A ,-RM ,-RA, the discrimination of FD description describes son and has been higher by 17.5,11.66,8.33 and 8 percentage points than MDM respectively by table 3；And the method that IOCV methods are better than FD, in four kinds of method for normalizing of-M ,-A ,-RM ,-RA, the discrimination of IOCV description describes son and is higher by 3,6,12.34 and 5.17 percentage points than FD respectively.Therefore, compared to MDM and FD description, IOCV description are obtained in that discrimination higher.

Claims (7)

1. a kind of digital image understanding method, including to the method for target image contour line shape facility description, the description side Method describes profile wire shaped using inside and outside chord length characteristic vector (IOCV), it is characterised in that IOCV can extract wheel The convex-concave characteristic of profile shape.

2. inside and outside chord length characteristic vector as claimed in claim 2 a, it is characterised in that target profile curve is shaped as binaryzation Image can be general be expressed as form：

$f(x,y)=\left\{\begin{array}{cc}1,& if(x,y)\∈D\\ 0,& otherwisee\end{array}\right.$

Here x, y represent the transverse and longitudinal coordinate of target image shape.D is target shape distributed areas in the picture.Extract shape The contour line of f simultaneously carries out uniform sampling to it, constitutes an orderly point set for N number of pixel C={ P_i(x_i, y_i), i=1,2 ..., N, it is 128 that N takes even number here, because target profile curve is closure, this In P_N+i=P_i。

3. the orderly point set C of the contour line of bianry image as claimed in claim 2, it is characterised in that make l_siRepresent from profile Point P_i(x_i, y_i) set out, another point P is reached in the counterclockwise direction_i+s(x_i+s, y_i+s) length of the string of segmental arc passed through, can Represented with equation below：

$l_{si}=\sqrt{{\left(x_i-x_{i+s}\right)}^2+{\left(y_i-y_{i+s}\right)}^2}.$

Falling to cut string shape area D is inside and outside according to the string.The part that the string falls in shape area D Length is usedRepresent, be called interior chord length；The length in the outside of shape area D that falls is usedRepresent, be called outer Chord length.

4. chord length in as claimed in claim 3With outer chord lengthCharacterized in that, making x₁=min { x_i, x_i+s, x₂=max { x_i, x_i+s, y₁=min { y_i, y_i+s, y₂=max { y_i, y_i+s, thenWithFollowing form table can be used Show：

$l_{si}^{\left(I\right)}={\∫}_{x_1}^{x_2}{\∫}_{y_1}^{y_2}f(x,y)\mathrm{\δ}(Ax+By-C)dxdy$

$l_{si}^{\left(O\right)}=l_{si}-l_{si}^{\left(I\right)}$

Here A=(y_i+s-y_i)/l_si, B=(x_i+s-x_i)/l_si, C=(x_iy_i+s-y_ix_i+s)/l_si.δ () represents Dirac functions.Need It is noted that Ax+By+C=0 is point P here_iWith point P_i+sConnected normal form of a straight line equation.Can be with from interior chord length formula Find out, interior chord lengthSubstantially shape function f (x, y) is in line segment P_iP_i+sOn projected length.

5. chord length in as claimed in claim 4With outer chord lengthCharacterized in that, fixed s, makes i be changed to from 1 N, we obtain the N sections of string being all corresponding to s unit segmental arc, i.e. every bit on contour line as once starting Point string counterclockwise corresponding to mobile s segmental arc.By calculating each section of inside and outside chord length of string, we obtain interior chord length sequence RowWith outer chord length sequenceIt is pointed out that because contour line is closure, institute Arc length is all corresponded to any one section of string and is respectively s unit length and N-s the two of unit length sections of arcs, to avoid data Redundancy, we allow arc length s values 1,2 ..., M=N/2.Thus two matrixes of M × N are obtained, interior chord length is referred to as Matrix ICM and outer chord length matrix OCM.

$\begin{array}{cc}ICM=\left(\begin{array}{cccc}{l}_{11}^{\left(I\right)}& l_{12}^{\left(I\right)}& ...& l_{1N}^{\left(I\right)}\\ l_{21}^{\left(I\right)}& l_{22}^{\left(I\right)}& ...& l_{2N}^{\left(I\right)}\\ ...& ...& ...& ...\\ l_{M1}^{\left(I\right)}& l_{M2}^{\left(I\right)}& ...& l_{MN}^{\left(I\right)}\end{array}\right)& OCM=\left(\begin{array}{cccc}{l}_{11}^{\left(O\right)}& l_{12}^{\left(O\right)}& ...& l_{1N}^{\left(O\right)}\\ l_{21}^{\left(O\right)}& l_{22}^{\left(O\right)}& ...& l_{2N}^{\left(O\right)}\\ ...& ...& ...& ...\\ l_{M1}^{\left(O\right)}& l_{M2}^{\left(O\right)}& ...& l_{MN}^{\left(O\right)}\end{array}\right)\end{array}$

, it is as claimed in claim 7 in chord length matrix ICM and outer chord length matrix OCM, it is characterised in that internal outer chord length It is normalized respectively, inside and outside chord length characteristic vector ICM and outer chord length matrix OCM met translation, selected and stretch The geometric invariance of contracting.

6. the interior chord length matrix ICM after normalizing as claimed in claim 5_NWith outer chord length matrix OCM_N, its feature exists In, two normalization are pulled into one-dimensional vector by row respectively, inside and outside chord length vector is combined composition contour line convex concave Description of shape characteristic, we term it inside and outside chord length characteristic vector (IOCV).

7. the interior chord length characteristic vector and outer chord length characteristic vector that Euclidean distance as claimed in claim 6 is represented are using diversity factor Contribute different when comparing, profile wire shaped convexo-concave characteristic adjusts interior chord length and outer chord length with weight w and 1-w when merging Contribution difference of the characteristic vector in identification.