CN1438605A - Beer-bottle raised character fetching-identifying hardware system and processing method - Google Patents

Abstract

The hardware system comprises the beer bottle conveyer belt on which is set an operation desk, a glare shield is set above the desk, the video camera is set up around the shield, which is connected to the computer, the rotatable desk positioned on the center of the operation desk. The images of beer bottles are obtained for further processing, detecting and recognizing. Integrated system for collecting and processing images is built in order to detect and recognize the 'B' and the data on the beer bottle so as to classify the bottles and determine whether the bottle is qualified or not.

Description

Beer bottle convexity character extracts and identification hardware system and disposal route

One, affiliated technical field

The invention belongs to the Computer Applied Technology field, relate to beer bottle convexity character and extract and identification hardware system and disposal route, come whether to have on the detection and Identification beer bottle ' B ' word and date of manufacture to judge whether beer bottle is qualified by the theory and the method for utilization computer vision and pattern-recognition.

Two, background technology

Beer bottle is as a kind of container that can bearing certain pressure, and all more common bottle of its manufacturing process and compressive resistance has than big difference.Interrelated data shows, the turnout of China's beer bottle can only reach 20% of annual requirement at present, the cost that adds beer bottle is than higher, therefore no matter from the angle that satisfies productive capacity still from saving the angle of cost, recycle old beer bottle and be absolutely necessary.

For the beer bottle of retrieving, may have following problem: 1. it is too much to recycle number of times.Because beer bottle is a kind of pressure vessel, all can be during each can beer to wherein injecting high pressure.Each pressurization all can produce the inner structure of beer bottle and destroy.Test shows the standard beer bottle that is newly dispatched from the factory, and pressurizeing continuously, body will be broken after 11 times.Therefore, the industry standard regulation, the recovered frequency of beer bottle can not be above 6 times.This number of times is difficult to statistics, and the general provision beer bottle can only be in the use in two years of date of manufacture.2. some common bottle, its shape and outward appearance and beer bottle are approaching, but compressive resistance differs greatly than beer bottle, if use these bottles to take splendid attire beer, will increase the possibility of body fragmentation greatly.

For the standard beer bottle is recycled, China State Bureau of Standardization unites the manufacturing standard GB-4544-1996[30 that relevant beer bottle manufacturing enterprise has formulated beer bottle], wherein clearly stipulate, " every product should be equipped with special marker ' B ' in above 20mm scope at the bottom of the bottle; to show is the special-purpose bottle of splendid attire, and the mark font size is with No. 2 block letter (length * wide about 6mm * 3mm) that is as the criterion.Should in this zone, indicate simultaneously the mark of manufacturing enterprise, the year of production, month, season." this recommended standard rises on June 1st, 1999 and transfer compulsory standard to.

From reclaim bottle, sort out the bottle that meets GB, simultaneously underproof bottle is abandoned, if this work finish by the people because the beer bottle use amount is very big, will employ large quantities of full-time staff undoubtedly and carry out go-on-go work, this is a huge financial burden to enterprise.Utilizing computing machine to finish this work automatically, is very significant.

At present, also there is not similar system both at home and abroad.

Three, summary of the invention

The objective of the invention is under industrial environment, provide a kind of beer bottle convexity character to extract and identification hardware system and disposal route.This method beer bottle is made a video recording and handle obtaining image, character detects, the technology of character recognition, set up the system of an integrated image collection, image processing, to realize that " B " word on the beer bottle and date are carried out detection and Identification, reach the purpose that beer bottle is classified.

The required work of finishing of computing machine is: be partitioned into ' B ' word and the title of an emperor's reign of representing beer bottle from the image that collects, if the title of an emperor's reign that has ' B ' and mark in 2 years, then illustrates it is qualified beer bottle, otherwise is defective bottle.

Realize that technical scheme of the present invention is to solve like this: beer bottle convexity character extracts and the identification hardware system, comprise, one beer bottle travelling belt, be characterized in, the beer bottle travelling belt is provided with an operator's console, and the top of operator's console is provided with a light shield, is provided with a video camera around the light shield, video camera is connected with a computing machine, and the central authorities of operator's console also are provided with a universal stage.

The camera lens of described video camera can be the line array CCD camera.

Realize that above-mentioned beer bottle convexity character extracts and the disposal route of discerning hardware system, may further comprise the steps at least:

1) data acquisition

The data that will constantly be collected by video camera are sent in the computing machine by the PORT COM of data collecting card, and computing machine is the gray level image of a width of cloth 256*2048*8Bits with these data splicings, and image is reduced to the dot matrix of 180*2048;

2) pre-service

Adopt literal location partitioning algorithm, size is calculated maximal value and minimum value respectively in the subwindow of M*N in former figure, and that utilizes these two values to obtain not contain background then comparatively clearly wants sketch map;

3) Character segmentation

The figure that an above step obtains is cut apart;

(4) stroke is extracted

Employing is regarded the gray scale of gray level image as height based on the method for topographic structure (Topographic structures), treats and analysis image with the viewpoint of topographic relief;

Arbitrarily the topographic structure of any can be divided into: ridge, ditch, peak, paddy, saddle, slope, plane etc., with I (x, y) the expression digital picture converts it to regional area continuous model, the local consecutive image signal of remembering be f (x, y); Each facet comprises the space of several pixel coverages, can come match with cubic polynomial:

f(x，y)＝k ₁+k ₂x+k ₃y+k ₄x ²+k ₅xy+k ₆y ²+k ₇x ³+k ₈x ²y+k ₉xy ²+k ₁₀y ³

By this model, can calculate single order, the second derivative of optional position $\frac{\&PartialD;f}{\&PartialD;x},\frac{\&PartialD;f}{\&PartialD;y},\frac{{\&PartialD;}^{2}f}{\&PartialD;x^2},\frac{{\&PartialD;}^{2}f}{{\&PartialD;\mathrm{<figure-callout\; id="2"\; label="y"\; filenames="A0311454100271.png,A0311454100291.png"\; state="\{\{state\}\}">y</figure-callout>}}^2},\frac{{\&PartialD;}^{2}f}{\&PartialD;x\&PartialD;y}$ And have $\frac{{\&PartialD;}^{2}f}{\&PartialD;x\&PartialD;y}=\frac{{\&PartialD;}^{2}f}{\&PartialD;y\&PartialD;x}$

Define the essential characteristic amount of each pixel below:

The gradient vector of f f

‖ f ‖ Grad

ω ⁽¹⁾Vector of unit length when the secondary directional derivative has maximal value.

ω ⁽²⁾With ω ⁽¹⁾The vector of unit length of quadrature.

λ ₁At ω ⁽¹⁾The value of the secondary directional derivative of direction,

λ ₂At ω ⁽²⁾The value of the secondary directional derivative of direction,

f. ω ⁽¹⁾At ω ⁽¹⁾The value of a directional derivative of direction,

f. ω ⁽²⁾At ω ⁽²⁾The value of a directional derivative of direction,

Under the little surface model of match in the above, (x, secondary directional derivative y) is can to obtain f $f_{\mathrm{\&beta;}}^{\&prime;\&prime;}=\left|\sin \mathrm{\&beta;}\cos \mathrm{\&beta;}\right|\&times;H\&times;\left|\frac{\sin \mathrm{\&beta;}}{\cos \mathrm{\&beta;}}\right|$

In the formula, H is He Sai (Hessian) matrix of 2*2, is defined as: $H=\left[\begin{array}{cc}\frac{{\&PartialD;}^{2}f}{\&PartialD;{\mathrm{<figure-callout\; id="2"\; label="x"\; filenames="A0311454100271.png,A0311454100291.png"\; state="\{\{state\}\}">x</figure-callout>}}^2}& \frac{{\&PartialD;}^{2}f}{\&PartialD;x\&PartialD;y}\\ \frac{{\&PartialD;}^{2}f}{\&PartialD;y\&PartialD;x}& \frac{{\&PartialD;}^{2}f}{{\&PartialD;\mathrm{<figure-callout\; id="2"\; label="y"\; filenames="A0311454100271.png,A0311454100291.png"\; state="\{\{state\}\}">y</figure-callout>}}^2}\end{array}\right]$

Two eigenwerts of He Sai matrix are two extreme values of secondary directional derivative, and two proper vectors corresponding with them are that twice directional derivative got the direction vector of extreme value, promptly

H ω ⁽¹⁾＝λ ₁ω ⁽¹⁾

H ω ⁽²⁾＝λ ₁ω ⁽²⁾

So, can obtain all essential characteristic parameters of above-mentioned definition by eigenwert and the characteristic of correspondence vector of asking the He Sai matrix; Further comprehensive these parameters can be judged the topographic structure at this pixel place, and the topographic structure concrete determination methods relevant with the stroke extraction is as follows:

Peak: ‖ f ‖=0, λ ₁＜0, λ ₂＜0

Paddy: ‖ f ‖=0, λ ₁＞0, λ ₂＞0

Ridge: ‖ f ‖ ≠ 0, λ ₁＜0, f ω ⁽¹⁾=0 or

‖ f ‖ ≠ 0, λ ₂＜0, f ω ⁽²⁾=0 or

‖f‖≠0，λ ₁＜0，λ ₂＝0

Saddle: ‖ f ‖=0, λ ₁λ ₂＜0

Planar point: ‖ f ‖=0, λ ₁=0, λ ₂=0

Utilizing topographic structure to carry out stroke then extracts;

(5) character recognition

At first adopt projection-conversion coefficient method to extract feature to character;

If N is a dimension of picture, g (x) and g (y) are respectively the projection of figure on X-axis and Y-axis, they are carried out the Fourier conversion obtain K conversion coefficient.From K conversion coefficient g _k(k=0,1,2 ..., select M representational feature in K-1), make it satisfy between class distance and want in big, the class distance little, the selection of M will decide along with different identification situations, and it is different promptly represent digital M number with English;

Make g _Ki ^(j)Be the k time conversion coefficient of j sample of i class-letter, i be alpha code (i=1 ..., S), S is the total number of word of set of letters, P is a total sample number.w _kBe the mean value of the k time conversion coefficient, then $w_k=\frac{1}{\mathrm{SP}}\underset{i=1}{\overset{S}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{P}{\mathrm{\&Sigma;}}}{g}_{\mathrm{ki}}^{\left(j\right)}$

g _KiBe the mean value of the k time conversion coefficient of i class-letter $g_{\mathrm{ki}}=\frac{1}{P}\underset{j=1}{\overset{P}{\mathrm{\&Sigma;}}}{g}_{\mathrm{ki}}^{\left(j\right)}$

c _kIt is the discrete value of k conversion coefficient $c_k=\frac{2}{\mathrm{SP}(P-1)}\underset{i=1}{\overset{S}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{P}{\mathrm{\&Sigma;}}}\underset{q=i+1}{\overset{P}{\mathrm{\&Sigma;}}}|g_{\mathrm{ki}}^{\left(j\right)}-g_{\mathrm{ki}}^{\left(q\right)}|$

Define the separability function Z of k conversion coefficient _kFor $Z_k=\frac{1}{c_k}\underset{i=1}{\overset{S}{\mathrm{\&Sigma;}}}|g_{\mathrm{ki}}-w_k|$

Z _kBig more, corresponding conversion coefficient g with it _kClassification capacity is preferably arranged more.Z when obtaining different k _kNumerical value, select the Z of M numerical value maximum _k, and with and the k time conversion coefficient g of its correspondence _kFeature as classification;

Obtain after the characteristic of division, utilize BP neural network design Identity System;

Neural network is by being called neuronic independent processing unit and being connected the network that arc connects into and being formed, and network can be divided into several layers; The neuron of input information is formed input layer, and these neurons do not have input and connect arc; The neuron of output information is formed output layer, and it does not have output and connects arc; All the other neurons are the middle layer, and its neuron is called intrerneuron; The middle layer is called hiding layer again; Its connection mode has individual layer connection mode, multilayer connection mode and circular form connection mode;

(6) at last with feature as input, character is as output, for the different neural network of letter and number design, through study and training, just the character recognition on the image can be come out.

The specific algorithm of described Character segmentation may further comprise the steps:

Step 1: adopting the region adaptivity binaryzation is bianry image with image segmentation;

Step 2: eight connected chain code trackings are carried out in the zone, and record is each regional starting point (X down _Min, Y _Min), terminal point (X _Max, Y _Max) etc. information;

The gradation of image value mean value nGrayAverage of the boundary chain code correspondence in gradient map that adds up is if nGrayAverage＜Thresh2 then deletes whole;

Step 3: utilize the geometrical constraint condition of relevant character, reject extraneous areas;

This step comprises following components:

3.1 if (H＞HeightMax) or (H＜HeightMin) or (W＞WidthMax) or (W＜WidthMin2) or ((W＜WidthMin) and ((H＞HeightMax) or (H＜HeightMin)) then this piece are the noise piece, are deleted;

This condition is removed the wide or narrow candidate region of character duration, particularly for character " 1 ", is provided with two minimum thresholding WidthMin2 and Width, satisfies: WidthMin2＜＜desired width＜WidthMin of character 1;

3.2 if (((W/H＞Ratio1) and (Width＜WidthMin3)) or (W/H＜Ratio2)) think then and do not satisfy the scale bar part that this piece is the noise piece, is deleted;

3.3, should delete if the geometric center and the average geometric centre deviation of character are excessive;

Step 4: utilize the geometrical constraint condition of character, with the zone merging of stroke fracture;

This step comprises following components:

4.1 if | Y _Mid(i)-Y _Mid(j) |＜Thresh7 and | X _Mid(i)-X _Mid(j) |＜Thresh7 these two how much pieces so should belong to transverse breakage, with its merging;

4.2 if | Xmid (i)-Xmid (j) |＜Thresh8

And then these two belong to vertical fracture, with its merging;

Repeat 4.1,4.2, till new merging does not take place;

Step 5: utilize the distance of adjacent how much interblocks to carry out the character boundary adjustment.

In system, it is as follows to utilize topographic structure to carry out the concrete steps that stroke extracts:

(1). the character picture that computed segmentation goes out one, second derivative.Their calculating can be passed through difference or local surface fitting method, and we have adopted the method for orthogonal polynomial here;

(2). according to eigenwert and the proper vector of calculating each some place curved surface He Sai matrix, judge the topographic structure type of this point;

(3). extract peak wherein, saddle point, valley point and the planar point that ridge point links to each other, as the stroke point.

The summation X of the input signal that described neuron is accepted _iStill can not reflect due various relations between the neuron input and output, for this reason, also need further to portray this relation, produce a new output with a characteristic function.The general characteristic function can be expressed as:

X _i＝∑w _jx _j+S _i (1a)

x′ _i＝F(X _i-θ _i) (1b)

S wherein _iBe the feedback information of internal state, θ _iBe threshold value, F is the characteristic function of expression neuron activity;

When a plurality of neuron was arranged, available vector form was expressed as:

X′＝F(XW+S) (2)

Wherein importing X is the N n dimensional vector n, and output X ' is the M n dimensional vector n.W is that N * M ties up matrix; Formula (2) means the network connection model input layer N neuron, and output layer has M neuron, and N internal feedback information is arranged, under the effect of characteristic function, by threshold value control output.

Four, description of drawings

Fig. 1 is system hardware structure figure of the present invention;

Fig. 2 is the beer bottle character picture (3 width of cloth) that common camera is gathered;

Fig. 3 is for using the image of beer bottle character picture behind medium filtering that line array CCD is gathered;

Fig. 4 is through pretreated image;

Fig. 5 is the Character segmentation result schematic diagram;

Fig. 6 regards the gray scale of gray level image as height, treats character topographic structure synoptic diagram with analysis image with the viewpoint of topographic relief; Wherein a is a gray level image, and b is the topomap that a fluctuating has order;

Fig. 7 is the result who utilizes the topographic structure stroke to extract, and wherein (a) is the gray-scale map image pattern, (b) is result figure;

The typical link model figure of Fig. 8 neural network;

Fig. 9 is neuronal structure figure.

Five, embodiment

The present invention is described in further detail below in conjunction with accompanying drawing.

Fig. 1 is that beer bottle convexity character extracts and identification hardware system structure synoptic diagram.Comprise that a beer bottle travelling belt 1, beer bottle travelling belt 1 are provided with an operator's console 2, the top of operator's console 2 is provided with a light shield 3, be provided with a video camera 5 around the light shield 3, video camera 5 is connected with a computing machine 6, and the central authorities of operator's console 2 also are provided with a universal stage 4.

The camera lens of video camera 5 can be the line array CCD camera.5.1 computer hardware device system

Wait to know bottle and be transferred to operator's console by travelling belt, wait to know bottle then and on universal stage, revolve and turn around, it is made a video recording, obtain the beer bottle character picture by computer-controlled line array CCD, after the process computer software was handled, beer bottle is divided three classes: a kind of was standard compliant bottle; A kind of is non-compliant bottle; At last a kind of is because some external conditions influence the bottle of refusing to know, then by manually discerning.As shown in Figure 1.2. computer software

(1) image acquisition

Fig. 2 is the beer bottle character pictures of several width of cloth by common camera collection.

By seeing among the figure, the image character that common camera collects is made up of the stroke of shade and brightness irregularities respectively, and background bright secretly inhomogeneous is for identification has brought very big difficulty.In order further to improve picture quality, reduce and disturb, accurately collect the character picture on the beer bottle, the inventor has adopted the line array CCD camera, stamps back of the body printing opacity simultaneously on through the beer bottle that cleans.Have only like this by the light source direct projection to those row just collected.After beer bottle transmits by travelling belt, on pallet, revolve and turn around, in rotary course, the line array CCD camera that is contained in the fixed position constantly sends the data that collect, PORT COM by data collecting card is sent in the main frame, is the gray level image of a width of cloth 256*2048*8Bits with these data splicings in main frame.Because the special marker of beer bottle is beaten in bottle end 20mm scope, in order to reduce follow-up operand, image is reduced to the dot matrix of 180*2048.

Fig. 3 is for using the image of beer bottle character picture behind medium filtering that line array CCD is gathered.

Character on glass is that the abutment surface convex-concave forms, and during optical imagery, the profile of character is the reflected light that relies on its surperficial varying strength, rather than relies on material that the different degrees of absorption of light are formed.Thereby for clear glass, the regional homogeneity of its gradation of image is very poor, and character shape is fracture and discontinuous shape, brings very big difficulty to subsequent treatment, adopts traditional image processing method to be difficult to obtain desirable effect.

(2) pre-service

Examine beer bottle characters in images feature, can notice, near stroke, in the very little zone, exist local brightness maximum value and minimal value.The bottom of beer bottle is thick than top simultaneously, transmittance is low, shows as background and has from top to bottom the gradual process of deepening transition gradually.According to above characteristics, we have designed a kind of new literal location partitioning algorithm, and its basic thought is that size is calculated maximal value and minimum value respectively in the subwindow of M*N in former figure, and that utilizes these two values to obtain not contain background then comparatively clearly wants sketch map.

Shown in Figure 4 is the result images that obtains after the process pre-service.As can be seen, stroke shows especially out, but still the interference that exists some to be caused by impurity and injurious surface mark, when gradual background is removed substantially, also brings the spinoff of stroke chap.But owing to the spacing broad between the character in this problem, can not produce adhesion phenomenon basically, therefore not influence following cutting operation.

(3) Character segmentation

Below, Fig. 4 that an above step obtains is cut apart.Because the original image gray-scale value that collects is less, so we have adopted a kind of dividing method that merges thought based on division.Cut apart and to utilize between the same row the relative information of character to remedy to cut apart the error that causes, so we needn't require separating of stroke and background very accurate.Specific algorithm is:

Step 1: adopting the region adaptivity binaryzation is bianry image with image segmentation.

Step 2: eight connected chain code trackings are carried out in the zone, and record is each regional starting point (X down _Min, Y _Min), terminal point (X _Max, Y _Max) etc. information.

Add up boundary chain code at gradient map I _dMiddle corresponding gradation of image value mean value nGrayAverage is if nGrayAverage＜Thresh2 then deletes whole.

Step 3: utilize the geometrical constraint condition of relevant character, reject extraneous areas.This method comprises following components:

3.1 if (H＞HeightMax) or (H＜HeightMin) or (W＞WidthMax) or (W＜WidthMin2) or ((W＜WidthMin) and ((H＞HeightMax) or (H＜HeightMin)) then this piece are the noise piece, are deleted.

This condition is removed the wide or narrow candidate region of character duration.Particularly for character " 1 ", the inventor is provided with two minimum thresholding WidthMin2 and Width, satisfies: WidthMin2＜＜desired width＜WidthMin of character 1.

3.2 if (((W/H＞Ratio1) and (Width＜WidthMin3)) or (W/H＜Ratio2)) think then and do not satisfy the scale bar part that this piece is the noise piece, should give deletion.

3.3, should delete if the geometric center and the average geometric centre deviation of character are excessive.

Step 4: utilize the geometrical constraint condition of character, with the zone merging of stroke fracture.

4.1 if | Y _Mid(i)-Y _Mid(j) |＜Thresh7 and | X _Mid(i)-X _Mid(j) |＜Thresh7 these two how much pieces so should belong to transverse breakage, so it should be merged.

4.2 if | Xmid (i)-Xmid (j) |＜Thresh8

And then these two belong to vertical fracture, so it should be merged.

Repeat 4.1,4.2, till new merging does not take place.

Step 5: utilize the distance of adjacent how much interblocks to carry out the character boundary adjustment.

Through the processing of above algorithm, reached the purpose of removing the noise piece and repairing the fracture character, utilize step 5 intercharacter distance is determined and to be adjusted simultaneously.So far, our separating character more accurately.Fig. 5 is the result schematic diagram of Character segmentation:

Need to prove: used the priori threshold value in the top step in a large number, in fact this algorithm is also insensitive to the selection of threshold value, and threshold value can be specified in very wide scope, and the variation among a small circle of threshold value is very little to the segmentation result influence; Other it should be noted that: we have obtained more stable and neat character block position by above step, but also are not enough to extract stroke clearly, and this will have influence on final identification.So it is necessary taking more efficiently method to carry out the character extraction.To extract stroke with method more accurately in the next section.

(4) stroke is extracted

The most outstanding characteristics of literal are that brightness power on the stroke is irregular on the beer bottle, if want to extract stroke above it, the algorithm that uses gray scale thresholding or edge extracting simply is not all right.Usually the edge feature of asking for is dissatisfactory to the robustness of optical transform.When under the irradiate light of certain angle and intensity, the edge on the three-dimensional body may not occur, and false edge may occur yet, but also the phenomenon of edge shifting may occur.In beer bottle text detection process,, the image that collects is extracted the edge have very strong instability because literal reflects by illumination fully.

In the present invention, the applicant has adopted the method based on topographic structure (Topographic structures).The basic thought of this method is: regard the gray scale of gray level image as height, treat and analysis image with the viewpoint of topographic relief.As shown in Figure 6, a gray level image, if gray scale is regarded as height, it just is equivalent to the topomap that a fluctuating has order.

The basic thought of this method is: regard the gray scale of gray level image as height, treat and analysis image with the viewpoint of topographic relief.The topographic structure of any can be divided into arbitrarily: ridge, ditch, peak, paddy, saddle, slope, plane etc.They all are the descriptions of the certain relief fabric of gray scale landform, and these all are three-dimensional features.And these features do not change with illumination and reflection, have very strong robustness.

With I (x, y) expression digital picture.We can convert it to regional area continuous model, and under this model, image is made up of many level and smooth facets, is called little surface model.The local consecutive image signal of remembering be f (x, y).Each facet comprises the space of several pixel coverages, can come match with cubic polynomial:

f(x，y)＝k ₁+k ₂x+k ₃y+k ₄x ²+k ₅xy+k ₆y ²+k ₇x ³+k ₈x ²y+k ₉xy ²+k ₁₀y ³

By this model, can calculate single order, the second derivative of optional position $\frac{\&PartialD;f}{\&PartialD;x},\frac{\&PartialD;f}{\&PartialD;y},\frac{{\&PartialD;}^{2}f}{\&PartialD;x^2},\frac{{\&PartialD;}^{2}f}{{\&PartialD;\mathrm{<figure-callout\; id="2"\; label="y"\; filenames="A0311454100271.png,A0311454100291.png"\; state="\{\{state\}\}">y</figure-callout>}}^2},\frac{{\&PartialD;}^{2}f}{\&PartialD;x\&PartialD;y}$ And have $\frac{{\&PartialD;}^{2}f}{\&PartialD;x\&PartialD;y}=\frac{{\&PartialD;}^{2}f}{\&PartialD;y\&PartialD;x}$

Define the essential characteristic amount of each pixel below:

The gradient vector of f f

‖ f ‖ Grad

ω ⁽¹⁾Vector of unit length when the secondary directional derivative has maximal value.

ω ⁽²⁾With ω ⁽¹⁾The vector of unit length of quadrature.

λ ₁At ω ⁽¹⁾The value of the secondary directional derivative of direction,

λ ₂At ω ⁽²⁾The value of the secondary directional derivative of direction,

f. ω ⁽¹⁾At ω ⁽¹⁾The value of a directional derivative of direction,

f. ω ⁽²⁾At ω ⁽²⁾The value of a directional derivative of direction,

Under the little surface model of match in the above, (x, secondary directional derivative y) is can to obtain f $f_{\mathrm{\&beta;}}^{\&prime;\&prime;}=\left|\sin \mathrm{\&beta;}\cos \mathrm{\&beta;}\right|\&times;H\&times;\left|\frac{\sin \mathrm{\&beta;}}{\cos \mathrm{\&beta;}}\right|$

In the formula, H is He Sai (Hessian) matrix of 2*2, is defined as: $H=\left[\begin{array}{cc}\frac{{\&PartialD;}^{2}f}{\&PartialD;x^2}& \frac{{\&PartialD;}^{2}f}{\&PartialD;x\&PartialD;y}\\ \frac{{\&PartialD;}^{2}f}{\&PartialD;y\&PartialD;x}& \frac{{\&PartialD;}^{2}f}{{\&PartialD;\mathrm{<figure-callout\; id="2"\; label="y"\; filenames="A0311454100271.png,A0311454100291.png"\; state="\{\{state\}\}">y</figure-callout>}}^2}\end{array}\right]$

Two eigenwerts of He Sai matrix are two extreme values of secondary directional derivative, and two proper vectors corresponding with them are that twice directional derivative got the direction vector of extreme value, promptly

H ω ⁽¹⁾＝λ ₁ω ⁽¹⁾

H ω ⁽²⁾＝λ ₁ω ⁽²⁾

So, can obtain all essential characteristic parameters of above-mentioned definition by eigenwert and the characteristic of correspondence vector of asking the He Sai matrix.Further comprehensive these parameters can be judged the topographic structure at this pixel place, and the topographic structure concrete determination methods relevant with the stroke extraction is as follows:

Peak: ‖ f ‖=0, λ ₁＜0, λ ₂＜0

Paddy: ‖ f ‖=0, λ ₁＞0, λ ₂＞0

Ridge: ‖ f ‖ ≠ 0, λ ₁＜0, f ω ⁽¹⁾=0 or

‖ f ‖ ≠ 0, λ ₂＜0, f ω ⁽²⁾=0 or

‖f‖≠0，λ ₁＜0，λ ₂＝0

Saddle: ‖ f ‖=0, λ ₁λ ₂＜0

Planar point: ‖ f ‖=0, λ ₁=0, λ ₂=0

In system, it is as follows to utilize topographic structure to carry out the concrete steps that stroke extracts:

(1). the character picture that computed segmentation goes out one, second derivative.Their calculating can be passed through difference or local surface fitting method, and we have adopted the method for orthogonal polynomial here.

(2). according to eigenwert and the proper vector of calculating each some place curved surface He Sai matrix, judge the topographic structure type of this point.

(3). extract peak wherein, saddle point, valley point and the planar point that ridge point links to each other, as the stroke point.

The test findings of extracting the edge in this way is as follows:

Fig. 7 is the result who utilizes the topographic structure stroke to extract, and wherein (a) is gray-scale map image pattern (b) result figure;

By test findings as can be seen, can resist of the variation of beer bottle literal surface to a certain extent, successfully extract the character on the image because of brightness and the generation of angle of incidence of light degree by the method that adopts the basic sketch of landform to express.In this concrete application, effective than classical way.

(5) character recognition

At first adopt projection-conversion coefficient method to extract feature to character.If N is a dimension of picture, g (x) and g (y) are respectively the projection of figure on X-axis and Y-axis, they are carried out the Fourier conversion obtain K conversion coefficient.From K conversion coefficient g _k(k=0,1,2 ..., select M representational feature in K-1), make it satisfy between class distance and want in big, the class distance little, the selection of M will decide along with different identification situations, and it is different promptly represent digital M number with English.

Make g _Ki ^(j)Be the k time conversion coefficient of j sample of i class-letter, i be alpha code (i=1 ..., S), S is the total number of word of set of letters, P is a total sample number.w _kBe the mean value of the k time conversion coefficient, then $w_k=\frac{1}{\mathrm{SP}}\underset{i=1}{\overset{S}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{P}{\mathrm{\&Sigma;}}}{g}_{\mathrm{ki}}^{\left(j\right)}$ g _KiBe the mean value of the k time conversion coefficient of i class-letter $g_{\mathrm{ki}}=\frac{1}{P}\underset{j=1}{\overset{P}{\mathrm{\&Sigma;}}}{g}_{\mathrm{ki}}^{\left(j\right)}$ c _kIt is the discrete value of k conversion coefficient $c_k=\frac{2}{\mathrm{SP}(P-1)}\underset{i=1}{\overset{S}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{P}{\mathrm{\&Sigma;}}}\underset{q=i+1}{\overset{P}{\mathrm{\&Sigma;}}}|g_{\mathrm{ki}}^{\left(j\right)}-g_{\mathrm{ki}}^{\left(q\right)}|$ Define the separability function Z of k conversion coefficient _kFor $Z_k=\frac{1}{c_k}\underset{i=1}{\overset{S}{\mathrm{\&Sigma;}}}|g_{\mathrm{ki}}-w_k|$

Z _kBig more, corresponding conversion coefficient g with it _kClassification capacity is preferably arranged more.Z when obtaining different k _kNumerical value, select the Z of M numerical value maximum _k, and with and the k time conversion coefficient g of its correspondence _kFeature as classification.

Obtain after the characteristic of division, utilize BP neural network design Identity System.

The generation of neural network is from biologically obtaining inspiration, and primary functional elements of some simulation biological neuron is organized the formation network, and this network demonstrates the surprising characteristic close with human brain.

The typical a kind of form of neural network is called link model.It is by being called neuronic independent processing unit and being connected the network that arc connects into and being formed, and network can be divided into several layers.The neuron of input information is formed input layer, and these neurons do not have input and connect arc; The neuron of output information is formed output layer, and it does not have output and connects arc; All the other neurons are the middle layer, and its neuron is called intrerneuron.The middle layer is called hiding layer again.Its connection mode has individual layer connection mode, multilayer connection mode and circular form connection mode.Shown in Figure 8 is a two-layer link model.x ₁, x ₂..., x _nBe input neuron, z ₁, z ₂..., z _pBe intrerneuron, y ₁, y ₂..., y _mBe output neuron.Each connects arc and is connecting two neurons, and with a numerical value W _IjAs weights (strength of joint or memory intensity), as neuron x _iTo z _lOr z _lTo y _i, influence.What positive weights represented that the increase that influences, negative weights represent to influence weakens, and is similar to the enhancing of cynapse and weakens.Current existing artificial neural network, its link model are almost all based on above-mentioned model.

The output of each neuron computes is again its all neuronic inputs of one deck down, is used further to calculate their output.Each neuron all uses their output of identical algorithm computation.One neuronic output is calculated with the neuronic output that is connected to it is connected arc with these weights.Neuron is the basic calculating unit of neural network, generally is the non-linear unit of a plurality of inputs, an output, and an internal feedback and threshold value can be arranged.Fig. 9 is a complete neuron x _iStructure.S wherein _iBe the feedback information of internal state, θ _iBe threshold value, F is the characteristic function of expression neuron activity.The summation X of the input signal that common neuron is accepted _iStill can not reflect due various relations between the neuron input and output.For this reason, also need further to portray this relation, produce a new output with a characteristic function.The general characteristic function can be expressed as:

X _i＝∑w _jx _j+s _i (1a)

x′ _i＝F(X _i-θ _i) (1b)

When a plurality of neuron was arranged, available vector form was expressed as:

X′＝F(XW+S) (2)

Wherein importing X is the N n dimensional vector n, and output X ' is the M n dimensional vector n.W is that N * M ties up matrix.Formula (2) means the network connection model input layer N neuron, and output layer has M neuron, and N internal feedback information is arranged, under the effect of characteristic function, by threshold value control output.

Back-propagation algorithm is called for short the BP algorithm, and it is the algorithm of a supervised training multilayer neural network proposing of people such as Luo Manhatuo.Each training example calculates through twice transmission in network: a propagation calculating forward.From input layer, transmit each layer and after treatment, produces an output, and obtain one and should reality export and the error vector of the difference of required output; To input layer, utilize error vector that weights are successively revised from output layer to backpropagation calculating one time.The BP algorithm has very strong Fundamentals of Mathematics, and it has expanded the usable range of neural network, has produced the example of many application success.The typical structural drawing of BP algorithm is identical with Fig. 8.

BP learning algorithm requirement neuron behavior function can be little.Therefore, there are many functions to can be used as characteristic function.As Sigmoid logic behavior function ${x^{\&prime;}}_j=F\left(X_j\right)=\frac{1}{(1+e^{-X_j})}---\left(3\right)$ X wherein _i=∑ w _Ijx _i, can meet this requirement, its partial derivative is: $\frac{{{\&PartialD;x}^{\&prime;}}_j}{{\&PartialD;x}_j}={x^{\&prime;}}_j(1-{x^{\&prime;}}_j)$

The purpose of training is that the connection arc weights of network are adjusted, and makes that network can produce a required output vector when using an input vector.The scope of training is composed of by the target vector of an input vector and required output, and both are called as " training to " together, and many training are trained the example collection to forming.

When beginning to train, the necessary initialization of whole weights of network is traditionally arranged to be little random number.This can guarantee that not getting maximal value because of weights makes network saturated or abnormal conditions occur.

The training step of BP algorithm is as follows:

(1) concentrates from the training example that to get a training right, the input of input vector as network;

(2) computational grid output vector;

(3) mistake between computational grid output vector and training centering target vector;

(4) again from the output layer backwards calculation to first middle layer, adjust network weight to the direction that reduces errors;

(5) each example in the training set is repeated above-mentioned (1)-(4) step, until the mistake minimum of whole training.

After the study, network is done the identification time spent, only uses (1), (2) step, and (3), (4) go on foot from output layer, the use alternative manner, backwards calculation is till first middle layer always.It calculates presses Delta rule distortion weights adjustment equation

w _Ij(n+1)=w _Ij(n)+η δ _jX ' _j(4) carry out, wherein

w _Ij(n) be the weights of neuron i to the n time change of neuron j;

X ' _jOutput (also available input x in the equation for neuron j _jFor it);

η is the learning rate constant;

δ _jMistake for neuron j.

What the BP algorithm adopted is to expand the Delta rule, investigates the mistake quadratic sum $E=\frac{1}{2}\underset{j}{\mathrm{\&Sigma;}}{({x^{\&prime;}}_j-\stackrel{\&OverBar;}{x_j})}^2---\left(5\right)$

In learning process, carry out in the mode that reduces E as quickly as possible.Generally whether it depends in the weights space and searches for along gradient direction.Make it finally to reach convergence for improving weights or threshold value, use by a certain percentage

Adjust weight w _Ij, w _IjVariable quantity Δ w _IjExpression: ${\mathrm{\&Delta;w}}_{\mathrm{ij}}=-\mathrm{\&eta;}\frac{\&PartialD;E}{\&PartialD;w_{\mathrm{ij}}}$

Because of the mistake quadratic sum by output x ' _iExpression, and x ' _jRepresent (seeing 3 formulas), partial derivative by neuronic non-linear output again

Available following formula calculates $\frac{\&PartialD;E}{\&PartialD;w_{\mathrm{ij}}}=\frac{\&PartialD;E}{{\&PartialD;X}_j}\frac{{\&PartialD;X}_j}{{\&PartialD;w}_{\mathrm{ij}}}$ Again $X_j=\underset{i}{\mathrm{\&Sigma;}}{w}_{\mathrm{ij}}{x}_i,$ So have $\frac{{\&PartialD;X}_j}{{\&PartialD;w}_{\mathrm{ij}}}=\frac{\&PartialD;}{\&PartialD;w_{\mathrm{ij}}}\underset{i}{\mathrm{\&Sigma;}}{w}_{\mathrm{ij}}{x}_i=x_i$ As definition ${\mathrm{\&delta;}}_i=-\frac{\&PartialD;e}{{\&PartialD;X}_j}$ Δ w then _Ij=η δ _jx _iThis expression formula is quite analogous to the expression formula of Delta rule.Now calculate ${\mathrm{\&delta;}}_j=-\frac{\&PartialD;E}{{\&PartialD;X}_j}=-\frac{\&PartialD;E}{{{\&PartialD;x}^{\&prime;}}_j}\frac{{{\&PartialD;x}^{\&prime;}}_j}{\&PartialD;X_j}$ When the j layer is output layer, get by formula (5) and characteristic function $\frac{\&PartialD;E}{{{\&PartialD;x}^{\&prime;}}_j}=-({x^{\&prime;}}_j-\stackrel{\&OverBar;}{x_j}),\frac{{{\&PartialD;x}^{\&prime;}}_j}{\&PartialD;X_j}=f^{\&prime;}\left(X_j\right)$ Output layer j has formula ${\mathrm{\&delta;}}_j=-({x^{\&prime;}}_j-\stackrel{\&OverBar;}{x_j})f\left(X_j\right)---\left(6\right)$ When j is the middle layer, have $\frac{\&PartialD;E}{{{\&PartialD;x}^{\&prime;}}_j}=-\underset{k}{\mathrm{\&Sigma;}}\frac{\&PartialD;E}{{\&PartialD;X}_k}\frac{{\&PartialD;X}_k}{{{\&PartialD;x}^{\&prime;}}_j}=\underset{k}{\mathrm{\&Sigma;}}\frac{\&PartialD;E}{{\&PartialD;X}_k}\left(\frac{\&PartialD;}{{{\&PartialD;x}^{\&prime;}}_j}\underset{m}{\mathrm{\&Sigma;}}{w}_{\mathrm{mk}}{x}_m\right)=\underset{k}{\mathrm{\&Sigma;}}\left(\frac{\&PartialD;E}{{\&PartialD;X}_k}\right)w_{\mathrm{jk}}=\underset{k}{\mathrm{\&Sigma;}}{\mathrm{\&delta;}}_k{w}_{\mathrm{jk}}$ At this moment ${\mathrm{\&delta;}}_j=\left(\underset{k}{\mathrm{\&Sigma;}}{\mathrm{\&delta;}}_k{w}_{\mathrm{jk}}\right)f^{\&prime;}\left(X_j\right)---\left(7\right)$ This explanation will be according to the Delta result of last layer to the neuron Delta rule in middle layer.Therefore, from top output layer, calculate with formula (6), subsequent step is propagated this Delta with formula (7) and is calculated, until bottom. $\frac{{{\&PartialD;x}^{\&prime;}}_j}{{\&PartialD;X}_j}=f^{\&prime;}\left(x_j\right)={x^{\&prime;}}_j(1-{x^{\&prime;}}_j)$

Specifically, the Delta rule that expands in the BP algorithm when the neuron behavior function adopts formula (3) is represented with following two formulas:

Output layer, i.e. x _iDuring for output neuron ${\mathrm{\&delta;}}_i=-({x^{\&prime;}}_j-\stackrel{\&OverBar;}{x_j}){x^{\&prime;}}_j(1-{x^{\&prime;}}_j)$ Middle layer, i.e. x _iDuring for intrerneuron here subscript k refer to the whole neuron index bounds of last layer of neuron j. ${\mathrm{\&delta;}}_j=\left(\underset{k}{\mathrm{\&Sigma;}}{\mathrm{\&delta;}}_k{w}_{\mathrm{jk}}\right){x^{\&prime;}}_j(1-{x^{\&prime;}}_j)$

Above-mentioned formula (7) has been arranged, can use formula (4) to realize adjusting output layer now and connect the arc weights, and then gone back with being connected the propagation downwards of the neuronic δ value of output layer, calculated each middle layer δ value with formula (7) with these weights.The δ value has been arranged, used formula (4) to adjust middle layer numerical value one by one, until input layer.The BP algorithm is to scan forward and the process that combines of scanning backward.

At last with feature as input, character is as output, for the different neural network of letter and number design, through study and training, just the character recognition on the image can be come out.

Claims (6)

1. a beer bottle convexity character extracts and the identification hardware system, comprise, one beer bottle travelling belt [1], it is characterized in that, beer bottle travelling belt [1] is provided with an operator's console [2], and the top of operator's console [2] is provided with a light shield [3], is provided with a video camera [5] around the light shield [3], video camera [5] is connected with a computing machine [6], and the central authorities of operator's console [2] also are provided with a universal stage [4].

2. beer bottle convexity character as claimed in claim 1 extracts and the identification hardware system, it is characterized in that the camera lens of described video camera [5] can be the line array CCD camera.

3. realize that the described beer bottle convexity of claim 1 character extracts and the disposal route of discerning hardware system, it is characterized in that, may further comprise the steps at least:

1) data acquisition

The data that will constantly be collected by video camera are sent in the computing machine by the PORT COM of data collecting card, and computing machine is the gray level image of a width of cloth 256*2048*8Bits with these data splicings, and image is reduced to the dot matrix of 180*2048;

2) pre-service

Adopt literal location partitioning algorithm, size is calculated maximal value and minimum value respectively in the subwindow of M*N in former figure, and that utilizes these two values to obtain not contain background then comparatively clearly wants sketch map;

3) Character segmentation

The figure that an above step obtains is cut apart;

4) stroke is extracted

Employing is regarded the gray scale of gray level image as height based on the method for topographic structure (Topographic structures), treats and analysis image with the viewpoint of topographic relief;

Arbitrarily the topographic structure of any can be divided into: ridge, ditch, peak, paddy, saddle, slope, plane etc., with I (x, y) the expression digital picture converts it to regional area continuous model, the local consecutive image signal of remembering be f (x, y); Each facet comprises the space of several pixel coverages, can come match with cubic polynomial:

f(x，y)＝k ₁+k ₂x+k ₃y+k ₄x ²+k ₅xy+k ₆y ²+k ₇x ³+k ₈x ²y+k ₉xy ²+k ₁₀y ³

By this model, can calculate single order, the second derivative of optional position $\frac{\&PartialD;f}{\&PartialD;x},\frac{\&PartialD;f}{\&PartialD;y},\frac{{\&PartialD;}^{2}f}{\&PartialD;x^2},\frac{{\&PartialD;}^{2}f}{{\&PartialD;y}^2},\frac{{\&PartialD;}^{2}f}{\&PartialD;x\&PartialD;y}$ And have $\frac{{\&PartialD;}^{2}f}{\&PartialD;x\&PartialD;y}=\frac{{\&PartialD;}^{2}f}{\&PartialD;y\&PartialD;x}$ Define the essential characteristic amount of each pixel below:

The gradient vector of f f

‖ f ‖ Grad

ω ⁽¹⁾Vector of unit length when the secondary directional derivative has maximal value.

ω ⁽²⁾With ω ⁽¹⁾The vector of unit length of quadrature.

λ ₁At ω ⁽¹⁾The value of the secondary directional derivative of direction,

λ ₂At ω ⁽²⁾The value of the secondary directional derivative of direction,

f. ω ⁽¹⁾At ω ⁽¹⁾The value of a directional derivative of direction,

f. ω ⁽²⁾At ω ⁽²⁾The value of a directional derivative of direction,

Under the little surface model of match in the above, (x, secondary directional derivative y) is can to obtain f $f_{\mathrm{\&beta;}}^{\&prime;\&prime;}=\left|\sin \mathrm{\&beta;}\cos \mathrm{\&beta;}\right|\&times;H\&times;\left|\frac{\sin \mathrm{\&beta;}}{\cos \mathrm{\&beta;}}\right|$

In the formula, H is He Sai (Hessian) matrix of 2*2, is defined as: $H=\left[\begin{array}{cc}\frac{{\&PartialD;}^{2}f}{\&PartialD;x^2}& \frac{{\&PartialD;}^{2}f}{\&PartialD;x\&PartialD;y}\\ \frac{{\&PartialD;}^{2}f}{\&PartialD;y\&PartialD;x}& \frac{{\&PartialD;}^{2}f}{{\&PartialD;y}^2}\end{array}\right]$

Two eigenwerts of He Sai matrix are two extreme values of secondary directional derivative, and two proper vectors corresponding with them are that twice directional derivative got the direction vector of extreme value, promptly

H ω ⁽¹⁾＝λ ₁ω ⁽¹⁾

H ω ⁽²⁾＝λ ₁ω ⁽²⁾

So, can obtain all essential characteristic parameters of above-mentioned definition by eigenwert and the characteristic of correspondence vector of asking the He Sai matrix; Further comprehensive these parameters can be judged the topographic structure at this pixel place, and the topographic structure concrete determination methods relevant with the stroke extraction is as follows:

Peak: ‖ f ‖=0, λ ₁＜0, λ ₂＜0

Paddy: ‖ f ‖=0, λ ₁＞0, λ ₂＞0

Ridge: ‖ f ‖ ≠ 0, λ ₁＜0, f ω ⁽¹⁾=0 or

‖ f ‖ ≠ 0, λ ₂＜0, f ω ⁽²⁾=0 or

‖f‖≠0，λ ₁＜0，λ ₂＝0

Saddle: ‖ f ‖=0, λ ₁λ ₂＜0

Planar point: ‖ f ‖=0, λ ₁=0, λ ₂=0

Utilizing topographic structure to carry out stroke then extracts;

5) character recognition

At first adopt projection-conversion coefficient method to extract feature to character;

If N is a dimension of picture, g (x) and g (y) are respectively the projection of figure on X-axis and Y-axis, they are carried out the Fourier conversion obtain K conversion coefficient.From K conversion coefficient g _k(k=0,1,2 ..., select M representational feature in K-1), make it satisfy between class distance and want in big, the class distance little, the selection of M will decide along with different identification situations, and it is different promptly represent digital M number with English;

Make g _Ki ^(j)Be the k time conversion coefficient of j sample of i class-letter, i be alpha code (i=1 ..., S), S is the total number of word of set of letters, P is a total sample number.w _kBe the mean value of the k time conversion coefficient, then $w_k=\frac{1}{\mathrm{SP}}\underset{i=1}{\overset{S}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{P}{\mathrm{\&Sigma;}}}{g}_{\mathrm{ki}}^{\left(j\right)}$

g _KiBe the mean value of the k time conversion coefficient of i class-letter $g_{\mathrm{ki}}=\frac{1}{P}\underset{j=1}{\overset{P}{\mathrm{\&Sigma;}}}{g}_{\mathrm{ki}}^{\left(j\right)}$

c _kIt is the discrete value of k conversion coefficient $c_k=\frac{2}{\mathrm{SP}(P-1)}\underset{i=1}{\overset{S}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{P}{\mathrm{\&Sigma;}}}\underset{q=i+1}{\overset{P}{\mathrm{\&Sigma;}}}|g_{\mathrm{ki}}^{\left(j\right)}-g_{\mathrm{ki}}^{\left(q\right)}|$ Define the separability function Z of k conversion coefficient _kFor $Z_k=\frac{1}{c_k}\underset{i=1}{\overset{S}{\mathrm{\&Sigma;}}}|g_{\mathrm{ki}}-w_k|$

Z _kBig more, corresponding conversion coefficient g with it _kClassification capacity is preferably arranged more.Z when obtaining different k _kNumerical value, select the Z of M numerical value maximum _k, and with and the k time conversion coefficient g of its correspondence _kFeature as classification;

Obtain after the characteristic of division, utilize BP neural network design Identity System;

Neural network is by being called neuronic independent processing unit and being connected the network that arc connects into and being formed, and network can be divided into several layers; The neuron of input information is formed input layer, and these neurons do not have input and connect arc; The neuron of output information is formed output layer, and it does not have output and connects arc; All the other neurons are the middle layer, and its neuron is called intrerneuron; The middle layer is called hiding layer again; Its connection mode has individual layer connection mode, multilayer connection mode and circular form connection mode;

6) at last with feature as input, character is as output, for the different neural network of letter and number design, through study and training, just the character recognition on the image can be come out.

4. beer bottle convexity character as claimed in claim 3 extracts the disposal route with the identification hardware system, it is characterized in that the specific algorithm of described Character segmentation may further comprise the steps:

Step 1: adopting the region adaptivity binaryzation is bianry image with image segmentation;

Step 2: eight connected chain code trackings are carried out in the zone, and record is each regional starting point (X down _Min, Y _Min), terminal point (X _Max, Y _Max) etc. information;

The gradation of image value mean value nGrayAverage of the boundary chain code correspondence in gradient map that adds up is if nGrayAverage＜Thresh2 then deletes whole;

Step 3: utilize the geometrical constraint condition of relevant character, reject extraneous areas;

This step comprises following components:

3.1 if (H＞HeightMax) or (H＜HeightMin) or (W＞WidthMax) or (W＜WidthMin2 or ((W＜WidthMin) and ((H＞HeightMax) or (H＜HeightMin)) then this piece are the noise piece, are deleted;

This condition is removed the wide or narrow candidate region of character duration, particularly for character " 1 ", is provided with two minimum thresholding WidthMin2 and Width, satisfies: WidthMin2＜＜desired width＜WidthMin of character 1;

3.2 if (((W/H＞Ratio1) and (Width＜WidthMin3)) or (W/H＜Ratio2)) think then and do not satisfy the scale bar part that this piece is the noise piece, is deleted;

3.3, should delete if the geometric center and the average geometric centre deviation of character are excessive;

Step 4: utilize the geometrical constraint condition of character, with the zone merging of stroke fracture;

This step comprises following components:

4.1 if | Y _Mid(i)-Y _Mid(j) |＜Thresh7 and | X _Mid(i)-X _Mid(j) |＜Thresh7 these two how much pieces so should belong to transverse breakage, with its merging;

4.2 if | Xmid (i)-Xmid (j) |＜Thresh8

And then these two belong to vertical fracture, with its merging;

Repeat 4.1,4.2, till new merging does not take place;

Step 5: utilize the distance of adjacent how much interblocks to carry out the character boundary adjustment.

5. beer bottle convexity character as claimed in claim 3 extracts the disposal route with the identification hardware system, it is characterized in that, in system, it is as follows to utilize topographic structure to carry out the concrete steps that stroke extracts:

(1). the character picture that computed segmentation goes out one, second derivative.Their calculating can be passed through difference or local surface fitting method, and we have adopted the method for orthogonal polynomial here;

(2). according to eigenwert and the proper vector of calculating each some place curved surface He Sai matrix, judge the topographic structure type of this point;

(3). extract peak wherein, saddle point, valley point and the planar point that ridge point links to each other, as the stroke point.

6. beer bottle convexity character as claimed in claim 3 extracts the disposal route with the identification hardware system, it is characterized in that the summation X of the input signal that described neuron is accepted _iStill can not reflect due various relations between the neuron input and output, for this reason, also need further to portray this relation, produce a new output with a characteristic function; The general characteristic function can be expressed as:

X _i＝∑w _jx _j+s _i (1a)

x′ _i＝F(x _i-θ _i) (1b)

S wherein _lBe the feedback information of internal state, θ _iBe threshold value, F is the characteristic function of expression neuron activity;

When a plurality of neuron was arranged, available vector form was expressed as:

X′＝F(XW+S) (2)

Wherein importing X is the N n dimensional vector n, and output X ' is the M n dimensional vector n.W is that N * M ties up matrix; Formula (2) means the network connection model input layer N neuron, and output layer has M neuron, and N internal feedback information is arranged, under the effect of characteristic function, by threshold value control output.

Images