CN101520845B - Layering method of color document images and device thereof - Google Patents
Layering method of color document images and device thereof Download PDFInfo
- Publication number
- CN101520845B CN101520845B CN2008100815577A CN200810081557A CN101520845B CN 101520845 B CN101520845 B CN 101520845B CN 2008100815577 A CN2008100815577 A CN 2008100815577A CN 200810081557 A CN200810081557 A CN 200810081557A CN 101520845 B CN101520845 B CN 101520845B
- Authority
- CN
- China
- Prior art keywords
- characteristic plane
- plane
- picture
- color
- background colour
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
Images
Abstract
The invention provides a layering device of color document images and a method thereof. The layering device of color document images of the invention executes the following processes of: determining background color of the document image; mapping each pixel point in the document image into a RGB space in which the background color is taken as an origin, and constructing a characteristic plane thatreflects the distribution of the pixel points of the document image in the RGB space; based on the density distribution in the characteristic plane, cutting the characteristic plane into more than on e areas; and based on the cutting result, dividing the document image into more than one layers.
Description
Technical field
The present invention relates to the layered approach and the device of color document images, more specifically, relate to the layered approach and the device of color document images with single background colour.
Background technology
Optical character identification (OCR) is a kind of technology of passing through the content of Computer Automatic Recognition latticed form file and picture, present being applied in the fields such as daily life and office by success.
In general, the identification of one width of cloth file and picture comprises printed page analysis and two steps of character recognition, the former is meant image segmentation to each semantic structure, comprises that paragraph, row, word, picture etc., the latter are that information with single character inputs in the recognizer and handles.The result of printed page analysis directly will directly influence final recognition accuracy.
Human eye can only identify tens kinds of gray shade scales on gray level image, but can identify thousands of kinds of colors from coloured image, so the relative gray level image of coloured image, and more information can be provided.In recent ten years, along with development of science and technology such as computing machine and optics, the printing of coloured image, scanning, storage and transmittability are greatly improved, make our document used in everyday also gradually from the greyscale image transitions to the coloured image.Usually in a secondary color document images, different contents (comprise the form, back printed contents of pre-printing, hand-written content, and seal etc.) represent with different colors, so that people's reading.
If we can be decomposed into several layers with color document images according to color characteristic, make every layer corresponding to a certain certain content in the document, respectively each layer handled and discerned then, the identifying of this input document will obtain effective simplification so, and the accuracy rate of identification also can improve greatly.
A kind of sorting technique of directly red, green, blue component in each pixel color being carried out linear calculating is disclosed in non-patent literature 1.This method is some kinds of colors of predefine at first, conclude by experiment then to sum up according to each color component and carry out the linear criterion of classifying of calculating.The advantage of this method is simple, quick, but the scope of application is too narrow, often need different sorting criterions be set at different documents, different scanners.
Non-patent literature 2 and 3 discloses based on the sorting technique of carrying out cluster on the RGB color space.This method is considered as a sample in the RGB color space with the color of each sampled point, and with some clustering algorithms commonly used, for example k-mean algorithm, figure clustering algorithm etc. are classified to it then.This method principle is very directly perceived, but do not consider the generation reason of color distribution in the color document images, and handle with a kind of general three-dimensional data sorting algorithm, be subjected to the limitation of some conditions on using: some algorithm needs to set in advance prioris such as categorical measure, and all kinds of samples meet a certain specific distribution characteristics in the space, and these conditions are ungratified usually.
Patent documentation 1 and 2 discloses based on the method for classifying on luminance component.These class methods at first are transformed into the color space of brightness-colourity form with the input color file and picture from the RGB color space, only carry out the processing of layering then according to luminance component, be similar to gray level image is carried out layering.Usually this class algorithm is divided into literal, line, common picture, the mix colours zone of attributes such as picture partly with luminance picture, and then carries out different processing according to each regional different attribute.Because these algorithms have only been considered monochrome information, do not utilize colouring information, have caused loss of learning, therefore cause its scope of application to be subjected to very big restriction.
Non-patent literature 1:Tony Allen, Nasser Sherkat, Seong Wong, " Use of colourfor hand-filled form analysis and recognition ", Pattern Analysis ﹠amp; Applications, v.8, n.1, pp.163-180,2005.
Non-patent literature 2:J.Zhou and D.Lopresti, " Extracting Text from WWWImages, " Proceedings of the 4th ICDAR, Ulm, Germany, v.1, pp.248-252,1997.
Non-patent literature 3:Sobottka, H.Bunke, and H.Kronenberg, " Identificationof text on colored book and journal covers, " In ICDAR ' 99, v.1, pp.57-60,1999.
Non-patent literature 4:M.Worring and L.Todoran, " Segmentation of colordocuments by line oriented clustering using spatial information, " Proceedingsof the 5
ThICDAR, pp.67-70,1999.
Patent documentation 1: United States Patent (USP) the 7th, 065, No. 254, the invention people is Kristine E.Matthews, name is called " Multilayered image file ".
Patent documentation 2: United States Patent (USP) 5,341, No. 226, the invention people is Jeng-Nan Shiau, name is called " Automatic image segmentation for color documents ".
Summary of the invention
The present invention is in view of the shortcoming and defect of above-mentioned prior art and propose, and its purpose is to provide equipment, method and the program etc. that can carry out layering efficiently and accurately to color document images.
In the desirable color document images of a width of cloth, different contents is designed to different colors, but every layer of all pixel that is comprised all have same color, i.e. the desirable look of this layer.But owing to, will inevitably comprise noise through after the operations such as printing, scanning; And, print or hand-written process in, resulting color is a kind of mixing of paper color and input color, the perhaps mixing of next figure layer color (background colour) and this figure layer color (foreground).Because the influence of these factors, each layer also not only comprises single a kind of color in the image that obtains, but the color space that a series of color is formed distributes.Experiment shows that each layer color of pixel shows as the distribution of a line segment type in rgb space, and the two ends of line segment are respectively foreground and background colour.
Because color document used in everyday generally all is with monochromatic paper, i.e. single background colour.In single background colour document, the pairing figure layer of different content shows as some the line segment type clusters that intersect at this background colour respectively in rgb space.
The present invention is based on the These characteristics of single background colour file and picture, consider the generation reason of color distribution in the color document images, proposed following technical scheme.
The layered approach of invention 1, a kind of color document images is characterized in that this method comprises the steps:
Determine the background colour of described file and picture;
Being mapped to each pixel in the described file and picture with described background colour is in the rgb space of initial point,
Structure has reflected the characteristic plane of the distribution of pixel in described rgb space of described file and picture;
Based on the Density Distribution in the described characteristic plane described characteristic plane is divided into more than one zone;
Based on described segmentation result, described file and picture is divided into more than one figure layer.
The color document images layered approach of invention 2, invention 1, it is characterized in that, in the step on described construction feature plane, consider that simultaneously each pixel in the described file and picture and the distance between the background colour and in the described characteristic plane each put pairing pixel distribution density.
The color document images layered approach of invention 3, invention 2, it is characterized in that, in the step on construction feature plane, each point in the described rgb space is converted to vector in the polar coordinate system, and is in the plane coordinate system of coordinate axis to angle with described polar coordinate system with described DUAL PROBLEMS OF VECTOR MAPPING.
The color document images layered approach of invention 4, invention 1 is characterized in that, in the step of described definite background colour, the pairing color of maximal value is as described background colour in the selection RGB color histogram.
Invention 5, invention 1 color document images layered approach is characterized in that, in the described step of cutting apart characteristic plane, with described characteristic plane be divided into this characteristic plane in the as many zone of number of peaks.
Invention 6, a kind of device that color document images is carried out layering is characterized in that this device comprises:
The background colour determining unit, it determines the background colour of described file and picture;
Map unit, it is mapped to each pixel in the described file and picture with described background colour is in the rgb space of initial point,
Characteristic plane construction unit, its structure have reflected the characteristic plane of the distribution of pixel in described rgb space of described file and picture;
The characteristic plane cutting unit, it is divided into more than one zone based on the Density Distribution in the described characteristic plane with described characteristic plane;
Figure layer division unit, it is divided into more than one figure layer based on described segmentation result with described file and picture.
The device of invention 7, invention 6, it is characterized in that, described characteristic plane construction unit considers that simultaneously each pixel in the described file and picture and the distance between the background colour and in the described characteristic plane each put pairing pixel distribution density, makes up described characteristic plane.
The device of invention 8, invention 7, it is characterized in that, described characteristic plane construction unit is converted to vector in the polar coordinate system with each point in the described rgb space, and is in the plane coordinate system of coordinate axis with described DUAL PROBLEMS OF VECTOR MAPPING to the angle with described polar coordinate system, to make up described characteristic plane.
The device of invention 9, invention 6 is characterized in that, the pairing color of maximal value is as described background colour in the described background colour determining unit selection RGB color histogram.
Invention 10, invention 6 device is characterized in that, described characteristic plane cutting unit with described characteristic plane be divided into this characteristic plane in the as many zone of number of peaks.
Invention 11, a kind of messaging device that makes is characterized in that to the program that color document images carries out layering this program makes messaging device carry out following steps:
Determine the background colour of described file and picture;
Being mapped to each pixel in the described file and picture with described background colour is in the rgb space of initial point,
Structure has reflected the characteristic plane of the distribution of pixel in described rgb space of described file and picture;
Based on the Density Distribution in the described characteristic plane described characteristic plane is divided into more than one zone;
Based on described segmentation result, described file and picture is divided into more than one figure layer.
The program of invention 12, invention 11 is characterized in that this program makes messaging device:
In the step on described construction feature plane, consider that simultaneously each pixel in the described file and picture and the distance between the background colour and in the described characteristic plane each put pairing pixel distribution density.
The program of invention 13, invention 12 is characterized in that this program makes messaging device:
In the step on construction feature plane, each point in the described rgb space is converted to vector in the polar coordinate system, and is in the plane coordinate system of coordinate axis to angle with described polar coordinate system with described DUAL PROBLEMS OF VECTOR MAPPING.
The color document images layered approach of invention 14, invention 11 is characterized in that this program makes messaging device:
In the step of described definite background colour, the pairing color of maximal value is as described background colour in the selection RGB color histogram.
The program of invention 15, invention 11 is characterized in that this program makes messaging device:
In the described step of cutting apart characteristic plane, with described characteristic plane be divided into this characteristic plane in the as many zone of number of peaks.
Invent 16, stored the computer-readable medium of inventing program any in 11～16.
According to the present invention, considered the generation reason of color distribution and the line style feature of distribution in the color document images, can improve the accuracy of color document images layering significantly.
Description of drawings
Fig. 1 is the synoptic diagram of the hierarchy of color document images;
Fig. 2 is the synoptic diagram of the distribution of color document images in the RGB color space;
Fig. 3 is the schematic block diagram of the color document images decker of first embodiment of the invention;
Fig. 4 shows an example of the distribution of pixel in the rgb space that with the background colour is initial point in the color document images;
Fig. 5 is an example of the constructed characteristic plane of characteristic plane construction unit;
Fig. 6 is the general flowchart of the performed processing of the characteristic plane cutting unit among first embodiment;
Fig. 7 is the synoptic diagram based on isocontour non-supervision formula sorting technique that adopts among first embodiment;
Fig. 8 is the one-dimensional data shown in Figure 7 resulting tree structure of classifying;
The synoptic diagram of Fig. 9 for the data area of input being cut apart according to classification tree shown in Figure 8;
Figure 10 is a characteristic plane shown in Figure 5 contour map corresponding to first height;
Figure 11 is a characteristic plane shown in Figure 5 corresponding to the contour map of first and second height;
Figure 12 for according to level line shown in Figure 11 to the characteristic plane shown in Figure 5 resulting tree structure of classifying;
The synoptic diagram of Figure 13 for characteristic plane shown in Figure 5 being cut apart according to the ground floor of tree structure shown in Figure 12;
The synoptic diagram of Figure 14 for characteristic plane segmentation result shown in Figure 13 further being cut apart according to the second layer of tree structure shown in Figure 12;
Figure 15 is the schematic flow diagram of the performed processing of the characteristic plane cutting unit among second embodiment;
Figure 16 is the synoptic diagram of among second embodiment one-dimensional data zone being cut apart;
Figure 17 is the synoptic diagram that calculates the even depth pond of gained according to characteristic plane shown in Figure 5;
Figure 18 is after spreading according to even depth pond shown in Figure 17, resulting segmentation result after characteristic plane shown in Figure 5 is classified.
Embodiment
Followingly describe layered approach and device in detail according to color document images of the present invention with reference to embodiment and accompanying drawing.
As shown in Figure 1, the color document of input is divided according to content and can be divided into three figure layers: form, literal and seal.Generally speaking these three layers is to represent with three kinds of different colors.The target of the layered approach of color document images of the present invention is decomposed these three figure layers exactly and is come according to colouring information, promptly obtain three width of cloth images shown in Fig. 1 right side.
As shown in Figure 2, the single background of input and comprise the color document images of three figure layers, meet certain distribution in the RGB color space: every layer of all color of pixel is distributed as the cluster of a line style, and two end points of this line style are respectively background colour and this bedding is thought look.Therefore, the color document images of input is distributed as the cluster of three line styles in rgb space, and three line segments have a common end points, i.e. image background look; Another end points of every line segment is the desirable look of this layer.
[embodiment 1]
Fig. 3 shows the structured flowchart of the color document images decker 1 of embodiments of the invention 1.The color document images that this color document images decker 1 input for example obtains by scanner, output is as layering result's more than one figure tomographic image.As shown in Figure 3, color document images decker 1 comprises background colour extraction unit 10, map unit 20, characteristic plane construction unit 30, characteristic plane cutting unit 40 and figure layer division unit 50.Background colour extraction unit 10 is determined the background colour of file and picture.It is in the rgb space of initial point that map unit 20 is mapped to each pixel in the file and picture with this background colour.Characteristic plane construction unit 30 makes up the characteristic plane of the distribution of pixel in described rgb space that has reflected the document image.Characteristic plane cutting unit 40 is divided into more than one zone based on the Density Distribution in the described characteristic plane with described characteristic plane.Figure layer division unit 50 is divided into more than one figure layer based on the segmentation result of characteristic plane cutting unit 40 with described file and picture.
Processing sequence when carrying out layering according to the color document images of 1 pair of input of color document images decker below is described in detail the work of its various piece.
At first, background colour extraction unit 10 is determined the background colour of file and picture.Particularly, background colour extraction unit 10 is set up the color histogram of input picture in the RGB color space, is determined the background colour (R of input picture then by this color histogram
b, G
b, B
b).Compute histograms is promptly calculated in the RGB color space, the pairing pixel quantity of each color in the input picture.Generally speaking, the pixel quantity that background comprised of image is maximum, therefore as an example, can be defined as background colour to the maximum color of pixel quantity.Certainly, this only is an example, also can according to circumstances adopt different background colours to settle the standard.In addition, determine that by color histogram the mode of image background look only is an example in the present embodiment, the invention is not restricted to this, as long as can determine the background colour of color document images, definite mode of background colour does not influence enforcement of the present invention.
After the background colour of having determined color document images, map unit 20 is mapped to each pixel in the file and picture with this background colour (R
b, G
b, B
b) be in the rgb space of initial point.Particularly, map unit 20 at first is the RGB coordinate system C of (0,0,0) with initial point
0Origin translation to background colour (R
b, G
b, B
b), obtain new RGB coordinate system C '.For each the pixel (R in the color document images
i, G
i, B
i), calculate its position (R in coordinate system C ' according to following formula 1
i', G
i', B
i').
R
i′＝R
i-R
b
G
i′＝G
i-G
b (1)
B
i′＝B
i-B
b
Promptly 3 coordinates are respectively R, G, B component values poor of R, G, B component values and the background colour of pixel color.Wherein, i is the index of pixel.
Thus, each pixel in the color document images being mapped to the background colour is in the rgb space of initial point.Fig. 4 shows an example of the distribution of pixel in the rgb space that with the background colour is initial point in the color document images.
Next, characteristic plane construction unit 30 makes up the characteristic plane of the distribution of pixel in rgb space C ' that has reflected the document image.
Particularly, for a bit (R among the coordinate system C '
i', G
i', B
i'), the vector from initial point to this point is V
i=[R
i', G
i', B
i']
T, this point is expressed as (Dis in polar coordinate system
i, α
i, β
i), Dis wherein
iBe vectorial V
iLength, α
iAnd β
iBe respectively the angle in the polar coordinate system.Dis
i, α
iAnd β
iBe respectively calculated as follows
α
i＝arctan(G
i′/R
i′) (2)
Then, be coordinate with α and β, construct a two dimensional surface, each pixel is all corresponding to a point on this plane in the color document images of input.For each point on this plane, computation of characteristic values F (α, β):
F(α，β)＝Dis(α，β)+λDen(α，β) (3)
Wherein (α is that ((α is corresponding to point (α, vectorial normalization distribution density β) β) to Den for α, the maximum length of institute's directed quantity β) corresponding to some β) to Dis.Promptly suppose corresponding to (α, vector set β) is combined into { V
i, i=1,2 ..., N (α, β) }, wherein N (α, β) be the vector quantity, then:
Dis(α，β)＝max{‖V
i‖，i＝1，2，...，N(α，β)} (4)
Den(α，β)＝N(α，β)/max{N(α，β)} (5)
λ is used to control the weight between length and the density for predetermined arithmetic number, for example can be taken as 1.0.
For each some computation of characteristic values, made up the characteristic plane of the distribution of pixel in rgb space C ' that has reflected the document image thus.Fig. 5 shows an example of characteristic plane, and wherein ordinate is the eigenwert size of the point in the characteristic plane.
Need to prove,, can adopt other account form, as long as can reflect the distribution of pixel in rgb space C ' of file and picture for the computing of the eigenwert of each point in the characteristic plane.For example, as another example, (α β) is used as the eigenwert of this point, shown in 6 can only to adopt distribution density Den.
F(α，β)＝Den(α，β) (6)
After this, characteristic plane cutting unit 40 is divided into more than one zone (classification) based on the distribution of the eigenwert in this characteristic plane with characteristic plane.In the present embodiment, employing is carried out this processing based on isocontour non-supervision formula sorting algorithm.Carry out specific description below.
Image, (α β) can be considered as a topographical surface to characteristic plane F, and as shown in Figure 5, wherein each class is all corresponding to mountain peak on the topographical surface.The effect of characteristic plane cutting unit 40 is that whole landform surface segmentation is obtained occupied zone, each mountain peak.Be similar to the watershed algorithm in the image segmentation, watershed divide and ponding basin are the notions in the topography, and the watershed divide is meant the abutment line between each ponding basin, underwater, and they have separated each ponding basin.If conversely with the landform in the watershed divide, then can be considered the watershed divide and separated each mountain peak, also promptly found the watershed divide just to reach the purpose that characteristic plane is cut apart.
Adopt a kind of data qualification algorithm of non-supervision formula in the present embodiment, this algorithm has been used for reference watershed algorithm, by selecting several suitable level lines to come the topographical surface of input is carried out cutting apart of stagewise, segmentation result can come record with a tree structure (being designated hereinafter simply as classification tree), can one by one this topographical surface be divided into the zone of some non-overlapping copies then according to this tree-like result.
Fig. 6 is the particular flow sheet of the characteristic plane cutting unit 40 characteristic plane dividing processing of being carried out.
Taking it by and large, at first input feature vector plane (S61).In step S62, calculate isocontour scope according to the characteristic plane of importing, wherein maximal value and minimum value are respectively the maximal value and the minimum value of each point value on this characteristic plane, and the initialization classification tree.Then, in step S63, calculate corresponding level line according to above-mentioned isocontour minimum value.In step S64, can judgement upgrade classification tree, if the pairing closed curve CP of a certain node comprises the level line CC of two or more current sealing in the current classification tree
1, CC
2..., and CC
N, then be judged as and upgrade classification tree.If judged result is sure, then enter step S65, upgrade classification tree, extend N child node from the CP corresponding node, correspond respectively to CC
1To CC
NEnter S67 then, judge whether to satisfy termination condition, whether promptly current isocontour height exceeds the maximal value of each point value on this characteristic plane.If satisfy, then output category is set (S68) and is finished this processing.If do not satisfy termination condition, then enter step S66, increase isocontour height, get back to step S64, can judgement upgrade classification tree.On the other hand, if the judged result of S64 negate, then enter step S67, judge whether to satisfy termination condition.At last, in step S69, the characteristic plane of input is cut apart according to this classification tree.
For ease of explanation and be convenient to understand the present invention, below describe whole process in detail with the watershed divide data instance of one dimension earlier.It should be noted that, in fact characteristic plane to be classified is a two dimensional surface, and its height is opposite with meaning in the watershed algorithm: treat in the grouped data each mountain peak (high-land) among the present invention corresponding to a class, and in the watershed divide each ponding basin (physical features is low) corresponding to a class.Therefore, in based on isocontour sorting algorithm, level line height the highest of initial setting up reduces the level line height then gradually; And in the present embodiment, minimum level line height of initial setting up improves the level line height then gradually.
Fig. 7, Fig. 8 and Fig. 9 are the examples when being applied to one-dimensional data based on isocontour watershed algorithm.Shown in Fig. 7 (a), the input data are the altitude information H of an one dimension, and each position x comprises three low ebbs altogether as can be seen to a height H (x) should be arranged in the field of definition, i.e. three ponding basins.Fig. 7 (b) has a level line of representing with horizontal dotted line, highly is h
1Statistics is lower than this isocontour region S R1={x|H (x)＜h
1, surpass h if reduce the level line height again
1, just SR1 can be divided into two line segment r1 and r2, promptly according to this level line h
1Can be lower than this isocontour Region Segmentation is two sub regions, with r1 and r2 represent respectively.Similarly, reduce the level line height gradually, the height of representing with a horizontal dotted line in as Fig. 7 (c) is h
2Level line the time, be lower than this isocontour zone and be divided into two sub regions by this level line again, represent with r3 and r4 respectively.Fig. 8 be shown in Figure 7 with this one-dimensional data with the present invention propose based on the isocontour sorting technique resulting tree structure of classifying.Root node is whole data area; At first, according to article one level line, be lower than isocontour zone and be split into " r1 " and " r2 " this two sub regions; Then, according to the second level line, be lower than isocontour zone " r2 " and be split into " r3 " and " r4 " this two sub regions.
Fig. 9 is the synoptic diagram of the data area of input being cut apart according to classification tree shown in Figure 8.Classification foundation is the arest neighbors criterion, certain some x and certain regional r={x on the zone
i, i=1,2 ..., between the N} apart from d (x r) is defined as:
If x belongs to r, then d (x, r)=0;
If x does not belong to r, then d (x, r)=min{|x-x
i|, i=1,2 ..., N}.
Can classify by twice pair of input data according to classification tree.One shown in Fig. 9 (a), according to two node r1 of ground floor and the r2 of this tree, is divided into R1 and two zones of R2 with whole data area; Its two, shown in Fig. 9 (b), according to the second layer of this tree, promptly two of r2 child node r3 and r4 are decomposed into R3 and R4 two sub regions again with region R 2.So far, whole data area is split into R1, R3 and these three zones of R4, and promptly data are divided into three classes, and all kinds of scopes is respectively R1, R3 and R4.
Characteristic plane with two dimension is that example describes this sorting technique below.Fig. 5 shows an example of characteristic plane.As shown in Figure 5, four peak valleys are arranged on this characteristic plane, correspond respectively to four classes that characteristic plane should be divided.Figure 10 to Figure 14 has shown the process of based on isocontour non-monitoring data sorting technique characteristic plane being cut apart with above-mentioned.
At first, select first height that is fit to and calculate level line, as shown in figure 10, this level line comprises the curve of three sealings, is respectively c1, c2 and c3.Then, select second suitable height and calculate level line, as shown in figure 11, new level line comprises the curve of four sealings, wherein go up in the closed curve that a level line comprised, have only the c2 region to comprise the closed curve of two current level line correspondences, be respectively c4 and c5.
Therefore, can construct as shown in figure 12 the tree structure that characteristic plane is cut apart that is used for according to above-mentioned level line.At last, according to this tree structure characteristic plane is cut apart.Figure 13 has shown the result of cutting apart according to the tree structure ground floor, and this characteristic plane can be broken down into the zone of c1, c2 and three curve correspondences of c3, is respectively C1, C2 and C3; Figure 14 has shown the result of Figure 13 being cut apart once more according to the tree structure second layer, and curve c2 corresponding region C2 is continued to be decomposed into the zone of c4 and two curve correspondences of c5, is respectively C4 and C5.
So far, the full feature plane is split into zone C 1, C2, C4 and the C5 of 4 non-overlapping copies, the input color file and picture can be decomposed into four figure layers according to this segmentation result.
After 40 pairs of characteristic planes of characteristic plane cutting unit are cut apart, send segmentation result to figure layer division unit 50.Figure layer division unit 50 is classified to corresponding figure layer according to the corresponding relation between the point in each pixel and the characteristic plane in this segmentation result and the file and picture with each pixel on the input file and picture.The pixel corresponding with a zone in the characteristic plane segmentation result is classified as a class, and is classified as different classifications corresponding to the pixel of zones of different in the characteristic plane segmentation result.Thus, according to the classification results of the pixel in the file and picture, each class pixel is divided into one deck in the color document images.
As mentioned above, in the first embodiment of the present invention, at first determine the background colour of input file and picture, it is in the rgb space of initial point that each pixel in the document image is mapped to the background colour, structure has reflected the characteristic plane of the distribution of pixel in the rgb space that with the background colour is initial point of the document image, based on the Density Distribution in this characteristic plane this characteristic plane is divided into more than one zone,, file and picture is divided into more than one figure layer based on segmentation result.According to first embodiment, made full use of the color distribution characteristic in the color document images of single background colour, can carry out layering to color document images exactly.
[second embodiment]
The structure of the color document images decker of second embodiment is substantially the same with the color document images decker 1 of above-mentioned first embodiment, comprises that being mapped to each pixel in the file and picture with this background colour is map unit in the rgb space of initial point; Structure has reflected the characteristic plane construction unit of characteristic plane of the distribution of pixel in described rgb space of the document image; Described characteristic plane is divided into the characteristic plane cutting unit in more than one zone based on the Density Distribution in the described characteristic plane; Described file and picture is divided into the figure layer division unit of figure layer more than based on the segmentation result of characteristic plane cutting unit.Difference is the processing that the characteristic plane cutting unit carries out.In a second embodiment, the characteristic plane cutting unit adopts the watershed algorithm based on gradient direction and even depth pond to come characteristic plane is cut apart.
Be that the center describes second embodiment with this difference below, then no longer repeat for the part identical with first embodiment.
Figure 15 is the general flowchart of the processing carried out of the characteristic plane cutting unit 40 among second embodiment.Taking it by and large, in step S151, from characteristic plane construction unit 30 input feature vector planes.In S152, find out the prime area of each mountain peak correspondence according to the characteristic plane of input, in S153, the prime area is spread then, so that the full feature plane is divided into several regions, in S154, this segmentation result is exported to figure layer division unit 50.
For the sake of simplicity, below still the watershed divide data instance with one dimension describe whole process in detail.It should be noted that equally, in the layered approach of color document images of the present invention, treat that grouped data is a two dimensional surface, and its height is opposite with meaning in the watershed algorithm: treat in the grouped data each mountain peak (high-land) among the present invention corresponding to a class, and in the watershed divide each ponding basin (physical features is low) corresponding to a class.
Figure 16 is the example when the method is applied to one-dimensional data.Figure 16 (a) is the one-dimensional data of input.Figure 16 (b) has shown each pairing prime area, ponding basin, at first calculate the valley point of each local minizing point as the ponding basin, search for this place, valley point neighborhood then to obtain the prime area (S152) that a connected domain is used as this ponding basin, make that institute has a respective heights all to be lower than this local maximum in the connected domain to add constant δ.The value of constant δ can rule of thumb be set, and for example, can be made as 0.1 generally speaking.Obtained three valley points in this example altogether,, obtained three the pairing prime area r1 in ponding basin, r2 and r3 altogether through search corresponding to three ponding basins.Point in each prime area is marked, make identical corresponding to the each point label in same ponding basin, and corresponding to the each point label difference in different ponding basin; Then the adjacent not mark point that marks is a little marked, make that the label of this adjacent not mark point is identical with the label that this has marked a little, if a certain mark point has marked a little adjacently with a plurality of simultaneously, then mark (corresponding to S154) according to the highly minimum point of mark.Through diffusion, obtained concrete region R 1, R2 and the R3 in three ponding basins, shown in Figure 16 (c).
About watershed algorithm based on gradient direction and even depth pond, for example can be with reference to following document:
Feng-Yang?Hsieh，and?Kuo-Chin?Fan，“An?unsupervised?watershedclassifier?based?on?gravity-space?image”，8th?IASTED?Inter.Conf.on?Signaland?Image?Processing?2006.
Adopting the True Data of two dimension below is that example describes this sorting technique.Fig. 5 carries out the background colour estimation and shines upon resulting characteristic plane a width of cloth real image.As shown in the figure, four peak valleys are arranged on the characteristic plane, correspond respectively to four classes that characteristic plane should be divided.Figure 17 and Figure 18 have shown the process of based on the watershed algorithm in gradient direction and even depth pond characteristic plane being cut apart with above-mentioned.
At first, calculate on this characteristic plane each local minizing point and search for this place, valley point neighborhood then to obtain the prime area that a connected domain is used as this ponding basin, as shown in figure 17 as the valley point in ponding basin.Obtain the prime area in four ponding basins altogether, represent with CDP1, CDP2, CDP3 and CDP4 respectively.
Then, the prime area in each ponding basin is spread.The first step marks the point in each prime area, make identical corresponding to the each point label in same ponding basin, and corresponding to the each point label difference in different ponding basin; Second step, the adjacent not mark point that marks is a little marked, make that the label of this adjacent mark point is identical with the label that this has marked a little,, then mark according to the highly minimum point of mark if an a certain mark while has marked a little adjacently with a plurality of.Through diffusion, obtained concrete zone C 1, C2, C3 and the R4 in four ponding basins, as shown in figure 18.
Next same with above-mentioned first embodiment, figure layer division unit 50 is classified to each pixel on the input file and picture on the figure layer of correspondence according to the corresponding relation between the point in each pixel in the segmentation result of this characteristic plane cutting unit 40 and the file and picture and the characteristic plane.The pixel corresponding with a zone in the characteristic plane segmentation result is classified as a class, and is classified as different classifications corresponding to the pixel of zones of different in the characteristic plane segmentation result.Thus, according to the classification results of the pixel in the file and picture, each class pixel is divided into one deck in the color document images.
According to second embodiment, the same background colour of at first determining the input file and picture with above-mentioned first embodiment, it is in the rgb space of initial point that each pixel in the document image is mapped to the background colour, structure has reflected the characteristic plane of the distribution of pixel in the rgb space that with the background colour is initial point of the document image, based on the Density Distribution in this characteristic plane this characteristic plane is divided into more than one zone, based on segmentation result, file and picture is divided into more than one figure layer.Similarly, make full use of the color distribution characteristic in the color document images of single background colour, can carry out layering to color document images exactly.
As mentioned above, main points of the present invention are in the layering of color document images, to determine the background colour of described file and picture; Being mapped to each pixel in the described file and picture with described background colour is in the rgb space of initial point; Structure has reflected the characteristic plane of the distribution of pixel in described rgb space of described file and picture; Based on the Density Distribution in the described characteristic plane described characteristic plane is divided into more than one zone; Based on described segmentation result, described file and picture is divided into more than one figure layer.In addition various details are not construed as limiting the invention.
For example, in the first and second above-mentioned embodiment, adopt respectively based on isocontour non-supervision formula sorting algorithm with based on the watershed algorithm in gradient direction and even depth pond to come characteristic plane is cut apart.But the present invention is not limited to these methods.In fact, the concrete grammar that characteristic plane is cut apart not is main points of the present invention, can adopt the whole bag of tricks to come characteristic plane is cut apart.
Again for example, in above-mentioned first and second embodiment, in the process on construction feature plane, by each point in the rgb space being converted to the vector in the polar coordinate system, and be in the plane coordinate system of coordinate axis to angle, thereby construction feature plane with this polar coordinate system with this DUAL PROBLEMS OF VECTOR MAPPING.But the invention is not restricted to this, also can come the construction feature plane, as long as this characteristic plane can reflect the distribution of each pixel in rgb space of file and picture by any other method.
In addition, among superincumbent first and second embodiment, be that the center describes the present invention with the color document images decker.But the invention is not restricted to this color document images decker, the storage medium that also may be embodied as the performed color document images layered approach of above-mentioned color document images decker, makes the messaging device of computing machine etc. carry out the program of this layered approach and write down this program.
In addition, in the above description, each unit is described as independent module.But these unit need not to be separated from each other physically.Can make up these unit by any way in addition, for example, can realize each unit of color document images decker of the present invention by the computing machine of program of operation the invention described above.
Claims (10)
1. the layered approach of a color document images is characterized in that, this method comprises the steps:
Determine the background colour of described file and picture;
Being mapped to each pixel in the described file and picture with described background colour is in the rgb space of initial point,
Structure has reflected the characteristic plane of the distribution of pixel in described rgb space of described file and picture;
Distribution based on the eigenwert of each point in the described characteristic plane is divided into more than one zone with described characteristic plane;
Based on described segmentation result, described file and picture is divided into more than one figure layer.
2. the layered approach of color document images as claimed in claim 1, it is characterized in that, in the step on described construction feature plane, consider that simultaneously each pixel in the described file and picture and the distance between the background colour and in the described characteristic plane each put pairing pixel distribution density.
3. the layered approach of color document images as claimed in claim 2, it is characterized in that, in the step on construction feature plane, each point in the described rgb space is converted to vector in the polar coordinate system, and is in the plane coordinate system of coordinate axis to angle with described polar coordinate system with described DUAL PROBLEMS OF VECTOR MAPPING.
4. the layered approach of color document images as claimed in claim 1 is characterized in that, in the step of described definite background colour, the pairing color of maximal value is as described background colour in the selection RGB color histogram.
5. the layered approach of color document images as claimed in claim 1 is characterized in that, in the described step of cutting apart characteristic plane, with described characteristic plane be divided into this characteristic plane in the as many zone of number of peaks of distribution of eigenwert of each point.
6. the device that color document images is carried out layering is characterized in that, this device comprises:
The background colour determining unit, it determines the background colour of described file and picture;
Map unit, it is mapped to each pixel in the described file and picture with described background colour is in the rgb space of initial point,
Characteristic plane construction unit, its structure have reflected the characteristic plane of the distribution of pixel in described rgb space of described file and picture;
The characteristic plane cutting unit, its distribution based on the eigenwert of the each point in the described characteristic plane is divided into more than one zone with described characteristic plane;
Figure layer division unit, it is divided into more than one figure layer based on described segmentation result with described file and picture.
7. device as claimed in claim 6, it is characterized in that, described characteristic plane construction unit considers that simultaneously each pixel in the described file and picture and the distance between the background colour and in the described characteristic plane each put pairing pixel distribution density, makes up described characteristic plane.
8. device as claimed in claim 7, it is characterized in that, described characteristic plane construction unit is converted to vector in the polar coordinate system with each point in the described rgb space, and be in the plane coordinate system of coordinate axis with described DUAL PROBLEMS OF VECTOR MAPPING to the angle with described polar coordinate system, to make up described characteristic plane.
9. device as claimed in claim 6 is characterized in that, the pairing color of maximal value is as described background colour in the described background colour determining unit selection RGB color histogram.
10. device as claimed in claim 6 is characterized in that, described characteristic plane cutting unit with described characteristic plane be divided into this characteristic plane in the as many zone of number of peaks of distribution of eigenwert of each point.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2008100815577A CN101520845B (en) | 2008-02-29 | 2008-02-29 | Layering method of color document images and device thereof |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2008100815577A CN101520845B (en) | 2008-02-29 | 2008-02-29 | Layering method of color document images and device thereof |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN101520845A CN101520845A (en) | 2009-09-02 |
| CN101520845B true CN101520845B (en) | 2011-11-30 |
Family
ID=41081427
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2008100815577A Expired - Fee Related CN101520845B (en) | 2008-02-29 | 2008-02-29 | Layering method of color document images and device thereof |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN101520845B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102881001A (en) * | 2011-07-13 | 2013-01-16 | 富士通株式会社 | Device and method for converting color image into grey scale image |
| CN103854019B (en) * | 2012-11-29 | 2018-11-30 | 捷讯平和（北京）科技发展有限公司 | Field extracting method and device in image |
| EP2806401A1 (en) * | 2013-05-23 | 2014-11-26 | Thomson Licensing | Method and device for processing a picture |
| CN109034147B (en) * | 2018-09-11 | 2020-08-11 | 上海唯识律简信息科技有限公司 | Optical character recognition optimization method and system based on deep learning and natural language |
| CN109407839B (en) * | 2018-10-18 | 2020-06-30 | 京东方科技集团股份有限公司 | Image adjusting method and device, electronic equipment and computer readable storage medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1374623A (en) * | 2001-03-14 | 2002-10-16 | 富士通株式会社 | Image processing equipment |
| JP2006004425A (en) * | 2004-06-14 | 2006-01-05 | Xerox Corp | Method for digital image segmentation |
| CN101030257A (en) * | 2007-04-13 | 2007-09-05 | 中国传媒大学 | File-image cutting method based on Chinese characteristics |
| US7356183B2 (en) * | 2002-07-01 | 2008-04-08 | Xerox Corporation | Segmentation method and system for multiple raster content (MRC) representation of documents |
-
2008
- 2008-02-29 CN CN2008100815577A patent/CN101520845B/en not_active Expired - Fee Related
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1374623A (en) * | 2001-03-14 | 2002-10-16 | 富士通株式会社 | Image processing equipment |
| US7356183B2 (en) * | 2002-07-01 | 2008-04-08 | Xerox Corporation | Segmentation method and system for multiple raster content (MRC) representation of documents |
| JP2006004425A (en) * | 2004-06-14 | 2006-01-05 | Xerox Corp | Method for digital image segmentation |
| US7376272B2 (en) * | 2004-06-14 | 2008-05-20 | Xerox Corporation | Method for image segmentation to identify regions with constant foreground color |
| CN101030257A (en) * | 2007-04-13 | 2007-09-05 | 中国传媒大学 | File-image cutting method based on Chinese characteristics |
Also Published As
| Publication number | Publication date |
|---|---|
| CN101520845A (en) | 2009-09-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US7437004B2 (en) | Method for processing data structures with networked semantic units | |
| Bruzzone et al. | A multilevel context-based system for classification of very high spatial resolution images | |
| Balsi et al. | Single-tree detection in high-density LiDAR data from UAV-based survey | |
| Nagao et al. | A structural analysis of complex aerial photographs | |
| US7039239B2 (en) | Method for image region classification using unsupervised and supervised learning | |
| CN103098074B (en) | Document page segmentation in optical character recognition | |
| CN101840581B (en) | Method for extracting profile of building from satellite remote sensing image | |
| CN102073748B (en) | Visual keyword based remote sensing image semantic searching method | |
| US20120093411A1 (en) | Active Segmentation for Groups of Images | |
| CN101520845B (en) | Layering method of color document images and device thereof | |
| CN1312625C (en) | Character extracting method from complecate background color image based on run-length adjacent map | |
| CN103577475A (en) | Picture automatic sorting method, picture processing method and devices thereof | |
| CN106875481B (en) | A kind of production method of three-dimensional visualization remote sensing image Surface classification model | |
| CN109272467A (en) | A kind of stratification image partition method based on multi-scale edge clue | |
| CN110147780A (en) | The landform recognition methods of real-time field robot and system based on level landform | |
| Peter | Measures for the generalization of polygonal maps with categorical data | |
| Saraceno et al. | Biophysical regions identification using an artificial neuronal network: A case study in the South Western Atlantic | |
| Čomić et al. | Morse-Smale decompositions for modeling terrain knowledge | |
| Carlinet et al. | A color tree of shapes with illustrations on filtering, simplification, and segmentation | |
| CN103927533A (en) | Intelligent processing method for graphics and text information in early patent document scanning copy | |
| Martı́ et al. | A new approach to outdoor scene description based on learning and top-down segmentation | |
| JPH09167233A (en) | Image processing method and device therefor | |
| Turtinen et al. | Visual training and classification of textured scene images | |
| Cousty et al. | Watersheds, minimum spanning forests, and the drop of water principle | |
| Ebi et al. | Automatic Data Acquisition from Topographic Maps Using a Knowledge Based Image Analysis System |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20111130 Termination date: 20180229 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |