CN106023224A - PCNN automatic segmentation method for microscopic image of traditional Chinese medicine - Google Patents

Abstract

The invention discloses a PCNN automatic segmentation method for a microscopic image of a traditional Chinese medicine. The PCNN automatic segmentation method comprises the steps of respectively establishing a CNN automatic binary image dividing algorithm which utilizes cross entropy segmentation criteria; establishing a traditional Chinese medicine microscopic image PCNN multi-value image automatic segmentation algorithm which utilizes maximizing mutual information as a segmentation object and utilizes mutual information entropy difference as a classification criteria, designing a vector pulse coupling neural network model, and realizing automatic segmentation on the microscopic color image of the traditional Chinese medicine through utilizing an index entropy criterion as a segmentation criteria; and establishing a traditional Chinese medicine microscopic image dividing algorithm in a multichannel or three-dimensional PCNN through utilizing a fuzzy index entropy as an optimized segmentation criterion. The PCNN automatic segmentation method has advantages of further improving objectivity, accuracy, repeatability and intelligent degree in quality evaluation of the traditional Chinese medicine, and providing a new approach for modernization of testing and analysis of the traditional Chinese medicine.

Description

A kind of PCNN automatic division method of microimage of Chinese medical herb

Technical field

The invention belongs to biomedical information process field, particularly relate to the PCNN of a kind of microimage of Chinese medical herb Automatic division method.

Background technology

Pulse Coupled Neural Network (PCNN) is according to same on the mammalian brain visual cortex such as cat, monkey Pace pulse provides what phenomenon proposed, has good Biological background, and this model has dynamic variable threshold value, non- The characteristics such as linear modulation coupling, lock-out pulse granting, dynamic pulse granting and space-time summation so that PCNN In signal processing applications, particularly in image procossing is applied, show huge superiority.But it is traditional PCNN model there is also following theoretical not enough and technical disadvantages:

(1) this model is in terms of non-linear modulation coupling and threshold exponent decay, and the decay of its threshold value is repeatedly to become Changing, this changes of threshold can not meet the human eye nonlinear exponent requirement to luminosity response well, and In image (or other signals) after being processed by this threshold value rule, bulk information lies in the activation of neuron In cycle (frequency) or activation phase place, and the image exported does not comprises whole available informations；

(2) a large amount of leak integrators and the existence of some feedback links in PCNN model, although improve mould The bionic approximation ratio of type and the verity of biological treatment information, but this not only adds the complexity of model, Also increase the expense to signal processing time simultaneously；

(3) tradition PCNN model parameter is too much, and (automatically) to parameter sets and optimization can increase perhaps The most difficult；

(4) due to complexity and the particularity of microimage of Chinese medical herb, tradition PCNN model is not suitable for place Reason microscopic structure class image.

Summary of the invention

It is an object of the invention to provide the PCNN automatic division method of a kind of microimage of Chinese medical herb, it is intended to solve Certainly the decay of tradition PCNN model threshold changes repeatedly, it is impossible to meet non-to luminosity response of human eye well Linearized index requirement, the image of output does not comprise whole available informations, and model is complicated, and parameter is too much, no The problem of suitable treatment microscopic structure class image.

The present invention is achieved in that the PCNN automatic division method of a kind of microimage of Chinese medical herb includes:

Microimage of Chinese medical herb is carried out adaptive by step one, introducing two dimension fuzzy collection or ultra-fuzzy sets membership function Should revise, sophisticated image fuzzy entropy or super fuzzy entropy expression formula, build unit link ULPCNN neuron mould Type；

Step 2, on present image, randomly choose a pixel；By conversion visual information computation model Window size, calculate the ceiling capacity of the direction passage of described pixel and its neighbor assignment；

Step 3, ceiling capacity according to each described direction passage determine described visual information computation model Out to out and useful direction, determine ULPCNN neuron according to described out to out and described useful direction Parameter W of model and M, wherein, M is the connection matrix in feed back input territory；W is of coupled connections the company in territory Connect matrix；

Step 4, the feature of foundation microimage of Chinese medical herb, the parameter optimizing ULPCNN neuron models sets Put；

Step 5, ULPCNN neuron models and maximum fuzzy entropy or super model are stuck with paste entropy criterion combine to figure As automatically splitting, extract Chinese crude drug bianry image target, set up the PCNN introducing cross entropy segmentation criterion Microimage of Chinese medical herb partitioning algorithm；

Step 6, with maximum mutual information optimize multivalue image split and carry out image denoising, set up adjacent segmentation Image Mutual information entropy difference minimum classification criterion, chooses Chinese medicine micro-image and sets up based on Minimum mutual information entropy difference PCNN automatic multi-valued targets partitioning algorithm, the multi-valued targets image improved；

Step 7, in coloured image RGB or HIS space, to build PCNN model optimization with push away Extensively；

Step 8, introducing Fuzzy Exponential Entropy segmentation criterion, optimize multichannel or three-dimensional microscopic image PCNN Process model, set up multichannel or three-dimensional PCNN microimage of Chinese medical herb target automatic segmentation algorithm.

Further, the parameter setting of ULPCNN neuron models is optimized method particularly includes:

ULPCNN neuron models chain is optimized in terms of ULPCNN neuron models morphosis and statistics two The parameter connecing input L and feed back input F nonlinear equation arranges, processes the local message adaptive optimization of image Coupling link strength β and optimum utilization output information improve repeatedly the dynamic threshold θ of exponential damping.

Further, to the PCNN model optimization built and popularization method particularly includes:

Introduce the exponential damping dynamic threshold vector improved, set up vector PCNN model；Utilize and improve index Information relativity between dynamic threshold vector and inside neurons active entry vector determines micro-segmentation image Target and background region, completes the automatic segmentation to Chinese crude drug coloured image in conjunction with maximal index entropy criterion.

Further, image denoising method particularly includes:

Step one, by noisy image f (x, y) carries out two dimension Stationary Wavelet Transform, obtains sub-band coefficients respectively: Low frequency coefficient, level detail coefficient, vertical detail coefficient and diagonal detail coefficient；

Step 2, low frequency coefficient to ground floor utilize PCNN to carry out region segmentation；

Step 3, low frequency coefficient is kept constant, to the level detail coefficient of each layer, vertical detail coefficient and Diagonal detail coefficient carries out adjacent region threshold process respectively；

Noise image is processed by step 4, employing Pulse Coupled Neural Network, obtains Entropy sequence En, will En is as edge detection operator；

Step 5, carry out threshold value optimizing, obtain optimum noise-removed threshold value k；

Step 6, according to the edge detection operator En that tries to achieve and optimum noise-removed threshold value k, use improve each to Anisotropic diffusion model carries out denoising to image.

Further, described ULPCNN neuron models also include inspection optimization module, this inspection optimization module For:

The image collected is set up the notable model of image, and the described significance model setting up image includes:

Utilize predetermined over-segmentation algorithm that described image carries out over-segmentation, and template parameter extracts, to whole defeated Enter image, with 8*8 pixel as unit, calculate average gray value and the maximum of each unit of each unit Gray value, obtains at least one region, and in same described region, the color value of each pixel is identical；

Determine color value and the barycenter in each described region；

According to the color value corresponding to regional and the barycenter of regional, set up described significance model.

Further, described described significance model is:

Wherein, S_i1For region R_iThe significance value of middle any pixel point, w (R_j) it is region R_jIn pixel The number of point, D_S(R_i,R_j) be used for characterizing described region R_iWith described region R_jBetween differences in spatial location Metric, D_C(R_i,R_j) be used for characterizing described region R_iWith described region R_jBetween the metric of color distortion, N be described image is carried out over-segmentation after total number in region of obtaining, D_S(R_i,R_j) it is:Center(R_i) it is described region R_iBarycenter, Center(R_j) it is described region R_jBarycenter, when in described image, the coordinate of each pixel all normalizes to [0,1] time；

Further, the Anisotropic Diffusion Model of described improvement is:

$\left\{\begin{array}{c}\frac{\∂I}{\∂t}=div(g\left(\▿I,En\right)\▿I)\\ I(x,y,0)=I_0\end{array}\right.$

Wherein:

$g(\▿I,En)=\frac{1}{1+|G_{\mathrm{\σ}}*\▿I+En|\·k}$

$G_{\mathrm{\σ}}=\frac{1}{2{\mathrm{\π\σ}}^2}\exp (-\frac{x^2+y^2}{2{\mathrm{\σ}}^2})$

σ is scaling function, and I represents that noise image, I0 represent that original image, div represent divergence operator, Represent gradient operator.

Further, the arrangement for correcting grey scale of average gray value includes:

Variable quantity calculating part, it calculates the variable quantity specifying frame of input luminance signal；

Variable quantity storage part, it stores described variable quantity；

Variable quantity comparing section, the variable quantity of its each frame to being stored in described variable quantity storage part compares Judge；

Histogram memory, it creates and stores each brightness electricity of described input luminance signal in units of frame Flat frequency data；

Correcting value calculating part, calculates described input brightness letter according to the frequency data of described histogram memory Number multiple intensity levels in each correcting value data；And,

Inquiry table storage part, by the described correcting value data from described correcting value calculating part, as corresponding to The table data of each intensity level store, and read corresponding with the intensity level of described input luminance signal Level data, exports as gray correction luminance signal；

Described correcting value calculating part includes:

Correction value table storage part, it calculates described school according to the frequency data of described histogram memory On the occasion of data, and store in units of frame；And,

Table selection portion, it is according to the result of determination of the variable quantity of the interframe of described variable quantity comparing section, to described In correction value table storage part, the correcting value meter of storage selects；

The result of determination of the variable quantity of interframe, when described variable quantity is more than setting, before selecting to change The correcting value meter of frame, when described variable quantity is below setting, select the correcting value meter before 1 frame to carry out Gray correction；

Described correcting value calculating part includes:

Correction table, it is for being corrected the correcting value meter of described correction value table storage part；And,

Table correction unit, it carries out school according to the variable quantity from described variable quantity comparing section to described correcting value meter Just；

The result of described judgement, when described variable quantity is below setting, it was predicted that the variable quantity of present frame, Use described correction table to the correcting value meter of described correction value table storage part according to the ratio of interframe variable quantity Example amount is corrected；

Described correcting value calculating part has detection flicker component in units of frame, and determines whether there is flicker component Flicker detection portion,

Do not use the correcting value data of the frame being tested with flicker.

Further, the frontier tracing method of the objects' contour of described bianry image target includes:

Frontier tracing starts, by target scanning, to start suitable when running into impact point from the upper left angle point of gradient image Sequence follow the tracks of, until follow the tracks of subsequent point return to starting point for closed outline or its subsequent point do not have again new after Till continuous point is for non-close line segment；If non-close profile, then trace into after the termination of side need to from Initial point starts to trace into another destination node in the opposite direction；When there is the profile of multiple separation, one by one with Track, for avoiding being absorbed in endless loop, should use background colour to be filled with it after regional processing above is good；

Specifically include:

(x, y) is a boundary point to current point, then next boundary point must (x, in 8 neighborhoods y) at point；

The trend on border can substantially be determined in mutual alignment according to former point and current point, to subsequent point Search time 8 neighborhoods to current point need not carry out calculating and compare, according to former point P and current some C in position again On difference, 5 pixels in edge direction are carried out calculating and compare；

Currently put P (x, y) position encoded for n in 8 neighborhoods of a upper boundary point C, then from current point (x, y) 8 neighborhoods in the position being encoded to n, clockwise the position of mobile 2 pixels is exactly next boundary point Initiating searches position；If not boundary point, then the starting point from search starts according to counterclockwise sequentially searching Rope, next boundary point is just found in search for 5 times altogether；

Current some P is 1 the position encoded of a upper boundary point, then understand, by above-mentioned criterion, the starting point searched for and be Position encoded be 7 point；If not impact point, start sequentially searching position from 7 and be encoded to 0,1,2,3 Point, when for the first time there is its gray scale equal to target area thresholding, this pixel is exactly required lower Boundary's point.

Further, the frontier tracing method of the objects' contour of described bianry image target specifically includes: figure In Xiang, background point value is 0, and impact point is 1, and Pk is kth boundary point, and the initial value of k is 0, and t represents that border terminates The number of point；

(1) the label value label of target area interested；

(2) scanogram the most from top to bottom, finds that point that first pixel is 1 and label value are label Point be starting point P0 on border, and (x, y) is stored in boundary point sequence table, and preset t is 0, position its coordinate Put and be encoded to 0；

(3) determine the original position that next impact point is searched for, then start the most successively from this position Check 8 neighborhood pixels of current border point, the label value of its pixel occurs equal to preset label value when for the first time Time, this pixel is exactly new boundary point Pk (k=k+1), and writes down its position encoded value in 8 neighborhoods；

(4) if new boundary point Pk=P0, i.e. having returned to starting point, this Contour extraction terminates, now boundary point sequence In deposit is exactly the external boundary point coordinates of this target, turn (7)；

(5) if new boundary point Pk ≠ P0, then using Pk as current point, write down that it is position encoded, then turn (3)；

(6) if not finding impact point, illustrating that current point is the destination node of profile, destination node number t adds 1, as Really t=1, then making P0 is current point, and seated position is encoded to 4, turns (3), i.e. opens not closed outline from starting point Beginning to search in the opposite direction, until finding another destination node, if t=2, then turning (7)；

(7) if also needing to follow the tracks of the profile of other targets, (2) are gone back to.

The present invention, under different images segmentation criterion, establishes the PCNN two-value of microimage of Chinese medical herb Image, multivalue image, coloured image and multichannel or the automatic segmentation algorithm of 3-D view target, can enter one Step improves the objectivity of Chinese crude drug quality evaluation, accuracy, repeatability and intelligence degree, for Chinese crude drug Detection provides a kind of new approach with the modernization analyzed.The present invention proposes the frontier tracing of a kind of highly versatile Algorithm, it is possible to move towards according to the position judgment profile of a upper boundary point.When the next boundary point of search, only need The 5 of candidate point judged, just can find the position of next boundary point, thus decrease search Number of times so that the time of frontier tracing is greatly reduced.Algorithm can also be once for the line segment that profile is not closed Scanning obtains its profile information.Experiment shows.Algorithm not only speed is fast, and outline identification is accurate.For The image that object is more complicated, algorithm more can embody its superiority.

Accompanying drawing explanation

Fig. 1 is the PCNN automatic division method flow chart of the microimage of Chinese medical herb that the embodiment of the present invention provides；

Fig. 2 is the ULPCNN neuron models schematic diagram that the embodiment of the present invention provides.

Detailed description of the invention

For the summary of the invention of the present invention, feature and effect can be further appreciated that, hereby enumerate following example, and Accompanying drawing is coordinated to describe in detail as follows.

Refer to Fig. 1 and Fig. 2:

Microimage of Chinese medical herb acquisition pretreatment and Jian Ku:

(1) intend recording according to pharmacopeia be distributed in the Radix Angelicae Sinensis of Gansu Province's different geographical, Radix Codonopsis, Radix Glycyrrhizae, Radix Et Rhizoma Rhei, Hundreds of kind of the Radix Astragali, Bulbus Lilii, Herba Ephedrae, Radix Bupleuri, Radix Isatidis, Fructus Foeniculi, Flos Carthami, Rhizoma Gastrodiae, Bulbus Fritillariae Uninbracteataes etc. are medicinal to be planted Thing is the primary object of research, in the different medical material seasons of growth from gathering genuine medicinal materials sample (or medicine vegetatively Material pollen sample), obtain Chinese medicine morphological image by high-resolution digital camera simultaneously.

(2) sampling, powder on the basis of tentatively identifying through several Chinese medicine cultivations and connoisseur and identify Endization pre-treatment, microsection manufacture etc. process, and obtain respectively finally by scanning electron microscope (or optical microscope) 5-10 visual field micro-image of specimen, the original image of pollen micro-image.

(3) in micro-image acquisition process due to the factor such as light brightness unevenness is even, cause image exposure Not enough or over-exposed image introduces image histogram correction or gray scale nonlinear transformation scheduling algorithm realizes figure Image intensifying；Environmental condition, CCD camera and other senser elements are affected, causes obtaining what image produced Noise jamming, uses the method such as medium filtering, Wiener filtering to reach image filtering purpose respectively.

(4) Large-scale Relational Database management system Oracle is used to build microimage of Chinese medical herb information bank.

The PCNN automatic division method of a kind of microimage of Chinese medical herb includes:

S101, introducing two dimension fuzzy collection or ultra-fuzzy sets membership function carry out self adaptation to microimage of Chinese medical herb Revise, sophisticated image fuzzy entropy or super fuzzy entropy expression formula, build unit link ULPCNN neuron models；

S102, on present image, randomly choose a pixel；By conversion visual information computation model Window size, calculates the ceiling capacity of described pixel and the direction passage of its neighbor assignment；

S103, ceiling capacity according to each described direction passage determine that described visual information computation model is Large scale and useful direction, determine ULPCNN neuron mould according to described out to out and described useful direction Parameter W of type and M, wherein, M is the connection matrix in feed back input territory；W is of coupled connections the connection in territory Matrix；

S104, the feature of foundation microimage of Chinese medical herb, the parameter optimizing ULPCNN neuron models is arranged；

S105, ULPCNN neuron models and maximum fuzzy entropy or super model are stuck with paste entropy criterion combine to image Automatically split, extract Chinese crude drug bianry image target, set up the PCNN introducing cross entropy segmentation criterion Microimage of Chinese medical herb partitioning algorithm；

S106, with maximum mutual information optimize multivalue image split and carry out image denoising, set up adjacent segmentation figure As Mutual information entropy difference minimum classification criterion, choose Chinese medicine micro-image and set up based on Minimum mutual information entropy difference PCNN automatic multi-valued targets partitioning algorithm, the multi-valued targets image improved；

S107, in coloured image RGB or HIS space, to build PCNN model optimization and popularization；

S108, introducing Fuzzy Exponential Entropy segmentation criterion, optimize multichannel or three-dimensional microscopic image PCNN Process model, set up multichannel or three-dimensional PCNN microimage of Chinese medical herb target automatic segmentation algorithm.

Further, the arrangement for correcting grey scale of average gray value includes:

Variable quantity calculating part, it calculates the variable quantity specifying frame of input luminance signal；

Variable quantity storage part, it stores described variable quantity；

Variable quantity comparing section, the variable quantity of its each frame to being stored in described variable quantity storage part compares Judge；

Histogram memory, it creates and stores each brightness electricity of described input luminance signal in units of frame Flat frequency data；

Correcting value calculating part, calculates described input brightness letter according to the frequency data of described histogram memory Number multiple intensity levels in each correcting value data；And,

Inquiry table storage part, by the described correcting value data from described correcting value calculating part, as corresponding to The table data of each intensity level store, and read corresponding with the intensity level of described input luminance signal Level data, exports as gray correction luminance signal,

Described correcting value calculating part includes:

Correction value table storage part, it calculates described school according to the frequency data of described histogram memory On the occasion of data, and store in units of frame；And,

Table selection portion, it is according to the result of determination of the variable quantity of the interframe of described variable quantity comparing section, to described In correction value table storage part, the correcting value meter of storage selects,

The result of determination of the variable quantity of interframe, when described variable quantity is more than setting, before selecting to change The correcting value meter of frame, when described variable quantity is below setting, select the correcting value meter before 1 frame to carry out Gray correction；

Described correcting value calculating part includes:

Correction table, it is for being corrected the correcting value meter of described correction value table storage part；And,

Table correction unit, it carries out school according to the variable quantity from described variable quantity comparing section to described correcting value meter Just,

The result of described judgement, when described variable quantity is below setting, it was predicted that the variable quantity of present frame, Use described correction table to the correcting value meter of described correction value table storage part according to the ratio of interframe variable quantity Example amount is corrected；

Described correcting value calculating part has detection flicker component in units of frame, and determines whether there is flicker component Flicker detection portion,

Do not use the correcting value data of the frame being tested with flicker.

Further, the parameter setting of ULPCNN neuron models is optimized method particularly includes:

ULPCNN neuron models chain is optimized in terms of ULPCNN neuron models morphosis and statistics two The parameter connecing input L and feed back input F nonlinear equation arranges, processes the local message adaptive optimization of image Coupling link strength β and optimum utilization output information improve repeatedly the dynamic threshold θ of exponential damping.

Further, to the PCNN model optimization built and popularization method particularly includes:

Introduce the exponential damping dynamic threshold vector improved, set up vector PCNN model；Utilize and improve index Information relativity between dynamic threshold vector and inside neurons active entry vector determines micro-segmentation image Target and background region, completes the automatic segmentation to Chinese crude drug coloured image in conjunction with maximal index entropy criterion.

Further, image denoising method particularly includes:

Step one, by noisy image f (x, y) carries out two dimension Stationary Wavelet Transform, obtains sub-band coefficients respectively: Low frequency coefficient, level detail coefficient, vertical detail coefficient and diagonal detail coefficient；

Step 2, low frequency coefficient to ground floor utilize PCNN to carry out region segmentation；

Step 3, low frequency coefficient is kept constant, to the level detail coefficient of each layer, vertical detail coefficient and Diagonal detail coefficient carries out adjacent region threshold process respectively；

Noise image is processed by step 4, employing Pulse Coupled Neural Network, obtains Entropy sequence En, will En is as edge detection operator；

Step 5, carry out threshold value optimizing, obtain optimum noise-removed threshold value k；

Step 6, according to the edge detection operator En that tries to achieve and optimum noise-removed threshold value k, use improve each to Anisotropic diffusion model carries out denoising to image.

Further, described ULPCNN neuron models also include inspection optimization module, this inspection optimization module For:

The image collected is set up the notable model of image, and the described significance model setting up image includes:

Utilize predetermined over-segmentation algorithm that described image carries out over-segmentation, and template parameter extracts, to whole defeated Enter image, with 8*8 pixel as unit, calculate average gray value and the maximum of each unit of each unit Gray value, obtains at least one region, and in same described region, the color value of each pixel is identical；

Determine color value and the barycenter in each described region；

According to the color value corresponding to regional and the barycenter of regional, set up described significance model.

Further, described described significance model is:

Wherein, S_i1For region R_iThe significance value of middle any pixel point, w (R_j) it is region R_jIn pixel The number of point, D_S(R_i,R_j) be used for characterizing described region R_iWith described region R_jBetween differences in spatial location Metric, D_C(R_i,R_j) be used for characterizing described region R_iWith described region R_jBetween the metric of color distortion, N be described image is carried out over-segmentation after total number in region of obtaining, D_S(R_i,R_j) it is:Center(R_i) it is described region R_iBarycenter, Center(R_j) it is described region R_jBarycenter, when in described image, the coordinate of each pixel all normalizes to [0,1] time；

Further, the Anisotropic Diffusion Model of described improvement is:

$\left\{\begin{array}{c}\frac{\∂I}{\∂t}=div(g\left(VI,En\right)\▿I)\\ I(x,y,0)=I_0\end{array}\right.$

Wherein:

$g(\▿I,En)=\frac{1}{1+|G_{\mathrm{\σ}}*\▿I+En|\·k}$

$G_{\mathrm{\σ}}=\frac{1}{2{\mathrm{\π\σ}}^2}\exp (-\frac{x^2+y^2}{2{\mathrm{\σ}}^2})$

σ is scaling function, and I represents that noise image, I0 represent that original image, div represent divergence operator, Represent gradient operator.

Build unit link ULPCNN neuron models, build and there is Monotone index ascending threshold function ULPCNN suppresses Capturing Models, as shown in Figure 2.

Further, the frontier tracing method of the objects' contour of described bianry image target includes:

Frontier tracing starts, by target scanning, to start suitable when running into impact point from the upper left angle point of gradient image Sequence follow the tracks of, until follow the tracks of subsequent point return to starting point for closed outline or its subsequent point do not have again new after Till continuous point is for non-close line segment；If non-close profile, then trace into after the termination of side need to from Initial point starts to trace into another destination node in the opposite direction；When there is the profile of multiple separation, one by one with Track, for avoiding being absorbed in endless loop, should use background colour to be filled with it after regional processing above is good；

Specifically include:

(x, y) is a boundary point to current point, then next boundary point must (x, in 8 neighborhoods y) at point；

The trend on border can substantially be determined in mutual alignment according to former point and current point, to subsequent point Search time 8 neighborhoods to current point need not carry out calculating and compare, according to former point P and current some C in position again On difference, 5 pixels in edge direction are carried out calculating and compare；

Currently put P (x, y) position encoded for n in 8 neighborhoods of a upper boundary point C, then from current point (x, y) 8 neighborhoods in the position being encoded to n, clockwise the position of mobile 2 pixels is exactly next boundary point Initiating searches position；If not boundary point, then the starting point from search starts according to counterclockwise sequentially searching Rope, next boundary point is just found in search for 5 times altogether；

Current some P is 1 the position encoded of a upper boundary point, then understand, by above-mentioned criterion, the starting point searched for and be Position encoded be 7 point；If not impact point, start sequentially searching position from 7 and be encoded to 0,1,2,3 Point, when for the first time there is its gray scale equal to target area thresholding, this pixel is exactly required lower Boundary's point.

Further, the frontier tracing method of the objects' contour of described bianry image target specifically includes: figure In Xiang, background point value is 0, and impact point is 1, and Pk is kth boundary point, and the initial value of k is 0, and t represents that border terminates The number of point；

(1) the label value label of target area interested；

(2) scanogram the most from top to bottom, finds that point that first pixel is 1 and label value are label Point be starting point P0 on border, and (x, y) is stored in boundary point sequence table, and preset t is 0, position its coordinate Put and be encoded to 0；

(3) determine the original position that next impact point is searched for, then start the most successively from this position Check 8 neighborhood pixels of current border point, the label value of its pixel occurs equal to preset label value when for the first time Time, this pixel is exactly new boundary point Pk (k=k+1), and writes down its position encoded value in 8 neighborhoods；

(4) if new boundary point Pk=P0, i.e. having returned to starting point, this Contour extraction terminates, now boundary point sequence In deposit is exactly the external boundary point coordinates of this target, turn (7)；

(5) if new boundary point Pk ≠ P0, then using Pk as current point, write down that it is position encoded, then turn (3)；

(6) if not finding impact point, illustrating that current point is the destination node of profile, destination node number t adds 1, as Really t=1, then making P0 is current point, and seated position is encoded to 4, turns (3), i.e. opens not closed outline from starting point Beginning to search in the opposite direction, until finding another destination node, if t=2, then turning (7)；

(7) if also needing to follow the tracks of the profile of other targets, (2) are gone back to.

The present invention, under different images segmentation criterion, establishes the PCNN two-value of microimage of Chinese medical herb Image, multivalue image, coloured image and multichannel or the automatic segmentation algorithm of 3-D view target, can enter one Step improves the objectivity of Chinese crude drug quality evaluation, accuracy, repeatability and intelligence degree, for Chinese crude drug Detection provides a kind of new approach with the modernization analyzed.

The above is only to presently preferred embodiments of the present invention, not makees the present invention any pro forma Limiting, any simple modification made for any of the above embodiments, equivalent are become by every technical spirit according to the present invention Change and modify, belonging in the range of technical solution of the present invention.

Claims (4)

1. the PCNN automatic division method of a microimage of Chinese medical herb, it is characterised in that described Chinese medicine The PCNN automatic division method of material micro-image includes:

Microimage of Chinese medical herb is carried out adaptive by step one, introducing two dimension fuzzy collection or ultra-fuzzy sets membership function Should revise, sophisticated image fuzzy entropy or super fuzzy entropy expression formula, build unit link ULPCNN neuron mould Type；

Step 2, on present image, randomly choose a pixel；By conversion visual information computation model Window size, calculate the ceiling capacity of the direction passage of described pixel and its neighbor assignment；

Step 3, ceiling capacity according to each described direction passage determine described visual information computation model Out to out and useful direction, determine ULPCNN neuron according to described out to out and described useful direction Parameter W of model and M, wherein, M is the connection matrix in feed back input territory；W is of coupled connections the company in territory Connect matrix；

Step 4, the feature of foundation microimage of Chinese medical herb, the parameter optimizing ULPCNN neuron models sets Put；

Step 5, ULPCNN neuron models and maximum fuzzy entropy or super model are stuck with paste entropy criterion combine to figure As automatically splitting, extract Chinese crude drug bianry image target, set up the PCNN introducing cross entropy segmentation criterion Microimage of Chinese medical herb partitioning algorithm；

Step 6, with maximum mutual information optimize multivalue image split and carry out image denoising, set up adjacent segmentation Image Mutual information entropy difference minimum classification criterion, chooses Chinese medicine micro-image and sets up based on Minimum mutual information entropy difference PCNN automatic multi-valued targets partitioning algorithm, the multi-valued targets image improved；

Step 7, in coloured image RGB or HIS space, to build PCNN model optimization with push away Extensively；

Step 8, introducing Fuzzy Exponential Entropy segmentation criterion, optimize multichannel or three-dimensional microscopic image PCNN Process model, set up multichannel or three-dimensional PCNN microimage of Chinese medical herb target automatic segmentation algorithm；

Image denoising method particularly includes:

Step one, by noisy image f (x, y) carries out two dimension Stationary Wavelet Transform, obtains sub-band coefficients respectively: Low frequency coefficient, level detail coefficient, vertical detail coefficient and diagonal detail coefficient；

Step 2, low frequency coefficient to ground floor utilize PCNN to carry out region segmentation；

Step 3, low frequency coefficient is kept constant, to the level detail coefficient of each layer, vertical detail coefficient and Diagonal detail coefficient carries out adjacent region threshold process respectively；

Noise image is processed by step 4, employing Pulse Coupled Neural Network, obtains Entropy sequence En, will En is as edge detection operator；

Step 5, carry out threshold value optimizing, obtain optimum noise-removed threshold value k；

Step 6, according to the edge detection operator En that tries to achieve and optimum noise-removed threshold value k, use improve each to Anisotropic diffusion model carries out denoising to image；

Optimize the parameter setting of ULPCNN neuron models method particularly includes:

ULPCNN neuron models chain is optimized in terms of ULPCNN neuron models morphosis and statistics two The parameter connecing input L and feed back input F nonlinear equation arranges, processes the local message adaptive optimization of image Coupling link strength β and optimum utilization output information improve repeatedly the dynamic threshold θ of exponential damping；

To the PCNN model optimization built and popularization method particularly includes:

Introduce the exponential damping dynamic threshold vector improved, set up vector PCNN model；Utilize and improve index Information relativity between dynamic threshold vector and inside neurons active entry vector determines micro-segmentation image Target and background region, completes the automatic segmentation to Chinese crude drug coloured image in conjunction with maximal index entropy criterion；

Described ULPCNN neuron models also include inspection optimization module, and this inspection optimization module is used for:

The image collected is set up the notable model of image, and the described significance model setting up image includes:

Utilize predetermined over-segmentation algorithm that described image carries out over-segmentation and template parameter extracts, to whole input Image, with 8*8 pixel as unit, calculates the average gray value of each unit and the maximum ash of each unit Angle value, obtains at least one region, and in same described region, the color value of each pixel is identical；

Determine color value and the barycenter in each described region；

According to the color value corresponding to regional and the barycenter of regional, set up described significance model；

Described significance model is:

Wherein, S_i1For region R_iThe significance value of middle any pixel point, w (R_j) it is region R_jIn pixel The number of point, D_S(R_i,R_j) be used for characterizing described region R_iWith described region R_jBetween differences in spatial location Metric, D_C(R_i,R_j) be used for characterizing described region R_iWith described region R_jBetween the metric of color distortion, N be described image is carried out over-segmentation after total number in region of obtaining, D_S(R_i,R_j) it is:Center(R_i) it is described region R_iBarycenter, Center(R_j) it is described region R_jBarycenter, when in described image, the coordinate of each pixel all normalizes to [0,1] time；

The Anisotropic Diffusion Model of described improvement is:

$\left\{\begin{array}{c}\frac{\∂I}{\∂t}=div(g\left(\▿I,En\right)\▿I)\\ I(x,y,0)=I_0\end{array}\right.;$

Wherein:

$g(\▿I,En)=\frac{1}{1+|G_{\mathrm{\σ}}*\▿I+En|\·k};$

$G_{\mathrm{\σ}}=\frac{1}{2{\mathrm{\π\σ}}^2}\exp (-\frac{x^2+y^2}{2{\mathrm{\σ}}^2});$

σ is scaling function, and I represents noise image, I₀Representing original image, div represents divergence operator, table Show gradient operator.

2. the PCNN automatic division method of microimage of Chinese medical herb as claimed in claim 1, it is characterised in that The arrangement for correcting grey scale of described average gray value includes:

Variable quantity calculating part, it calculates the variable quantity specifying frame of input luminance signal；

Variable quantity storage part, it stores described variable quantity；

Variable quantity comparing section, the variable quantity of its each frame to being stored in described variable quantity storage part compares Judge；

Histogram memory, it creates and stores each brightness electricity of described input luminance signal in units of frame Flat frequency data；

Correcting value calculating part, calculates described input brightness letter according to the frequency data of described histogram memory Number multiple intensity levels in each correcting value data；And,

Inquiry table storage part, by the described correcting value data from described correcting value calculating part, as corresponding to The table data of each intensity level store, and read corresponding with the intensity level of described input luminance signal Level data, exports as gray correction luminance signal；

Described correcting value calculating part includes:

Correction value table storage part, it calculates described school according to the frequency data of described histogram memory On the occasion of data, and store in units of frame；And,

Table selection portion, it is according to the result of determination of the variable quantity of the interframe of described variable quantity comparing section, to described In correction value table storage part, the correcting value meter of storage selects；

The result of determination of the variable quantity of interframe, when described variable quantity is more than setting, before selecting to change The correcting value meter of frame, when described variable quantity is below setting, select the correcting value meter before 1 frame to carry out Gray correction；

Described correcting value calculating part includes:

Correction table, it is for being corrected the correcting value meter of described correction value table storage part；And,

Table correction unit, it carries out school according to the variable quantity from described variable quantity comparing section to described correcting value meter Just；

The result of described judgement, when described variable quantity is below setting, it was predicted that the variable quantity of present frame, Use described correction table to the correcting value meter of described correction value table storage part according to the ratio of interframe variable quantity Example amount is corrected；

Described correcting value calculating part has detection flicker component in units of frame, and determines whether there is flicker component Flicker detection portion；

Do not use the correcting value data of the frame being tested with flicker.

3. the PCNN automatic division method of microimage of Chinese medical herb as claimed in claim 1, it is characterised in that The frontier tracing method of the objects' contour of described bianry image target includes:

Frontier tracing starts, by target scanning, to start suitable when running into impact point from the upper left angle point of gradient image Sequence follow the tracks of, until follow the tracks of subsequent point return to starting point for closed outline or its subsequent point do not have again new after Till continuous point is for non-close line segment；If non-close profile, then trace into after the termination of side need to from Initial point starts to trace into another destination node in the opposite direction；When there is the profile of multiple separation, one by one with Track, should use background colour to be filled with it after regional processing above is good；

(x, y) is a boundary point to current point, then next boundary point must (x, in 8 neighborhoods y) at point；

The trend on border is substantially determined in mutual alignment according to former point and current point, is searching subsequent point 8 neighborhoods to current point need not carry out calculating and compare again when seeking, according to former point P and current some C in position 5 pixels in edge direction are carried out calculating and compare by difference；

Currently put P (x, y) position encoded for n in 8 neighborhoods of a upper boundary point C, then from current point (x, y) 8 neighborhoods in the position being encoded to n, clockwise the position of mobile 2 pixels is exactly next boundary point Initiating searches position；If not boundary point, then the starting point from search starts according to counterclockwise sequentially searching Rope, next boundary point is just found in search for 5 times altogether；

Current some P is 1 the position encoded of a upper boundary point, then understand, by above-mentioned criterion, the starting point searched for and be Position encoded be 7 point；If not impact point, start sequentially searching position from 7 and be encoded to 0,1,2,3 Point, when for the first time there is its gray scale equal to target area thresholding, this pixel is exactly required lower Boundary's point.

4. the PCNN automatic division method of microimage of Chinese medical herb as claimed in claim 3, it is characterised in that The frontier tracing method of the objects' contour of described bianry image target specifically includes: background point value in image Being 0, impact point is 1, and Pk is kth boundary point, and the initial value of k is 0, and t represents the number of border destination node；

(1) the label value label of target area interested；

(2) scanogram the most from top to bottom, finds that point that first pixel is 1 and label value are label Point be starting point P0 on border, and (x, y) is stored in boundary point sequence table, and preset t is 0, position its coordinate Put and be encoded to 0；

(3) determine the original position that next impact point is searched for, then start the most successively from this position Check 8 neighborhood pixels of current border point, the label value of its pixel occurs equal to preset label value when for the first time Time, this pixel is exactly new boundary point Pk (k=k+1), and writes down its position encoded value in 8 neighborhoods；

(4) if new boundary point Pk=P0, i.e. having returned to starting point, this Contour extraction terminates, now boundary point sequence In deposit is exactly the external boundary point coordinates of this target, turn (7)；

(5) if new boundary point Pk ≠ P0, then using Pk as current point, write down that it is position encoded, then turn (3)；

(6) if not finding impact point, illustrating that current point is the destination node of profile, destination node number t adds 1, as Really t=1, then making P0 is current point, and seated position is encoded to 4, turns (3), i.e. opens not closed outline from starting point Beginning to search in the opposite direction, until finding another destination node, if t=2, then turning (7)；

(7) if also needing to follow the tracks of the profile of other targets, (2) are gone back to.