CN1206866C - Hierarchical object segmentation method based on dynamic image compression standard - Google Patents

Abstract

The present invention relates to a hierarchical object segmentation method based on the motion picture compression standard seven (MPEG-7), which is arranged based on an MPEG (motion picture experts group), wherein a watershed segmentation technology is combined by a descriptor of the MPEG-7 to provide a picture object segmentation mode. The present invention comprises the steps that a computer is trained to know a plurality of picture objects in advance; then, the characteristics of the picture objects (namely the descriptor defined by the MPEG-7) are extracted to be used as the standard of a database; whether the picture objects obtained by the watershed segmentation technology are the needed objects is judged by the characteristics until a most similar object is found out.

Description

Hierarchy type Object Segmentation method based on dynamic image compression standard

Technical field

The present invention relates to a kind of image partitioning method, particularly a kind of hierarchy type Object Segmentation method based on dynamic image compression standard.

Background technology

In recent years, image processing technique is constantly progressive, and the research of video object segmentation (video objectsegmentation) also has more and more many scholars to drop into.The compression algorithm that resembles MPEG-1/2 is in the past just deleted the redundant data between video, and has proposed different compress technique among the MPEG-4, and this technology is called content-based video compression (content-based video coding).The compress technique of MPEG-4 can be according to demand, video content is divided into several VOP (videoobject planes, video object plane), and these VOP being encoded respectively stores and transmits again,, recombinate, delete or replace required VOP by different application at decoding end.

Present employed video object segmentation method roughly can be divided automatically and semi-automatic two kinds, cut apart automatically mainly be utilize object mobile message (motion information) as cutting apart foundation, utilize mobile message that foreground object is separated in background.But this method must just can produce VOP to liking under the condition that moves.This technology of utilizing mobile message to cut apart automatically, the object that moves for meeting is pretty good method, however its shortcoming but is to handle static object.

For the object that does not move, semi-automatic is common processing method.Semi-automatic mainly is by manual operation and area of computer aided computing, most Semi-Automatic Video Object is cut apart research and is all tried to achieve wherein initial image object by user and artificial interface software with interactive mode, to the necessary about position of predefined object bounds (boundary) of a visual user, then utilize active contour model (active contour model) to carry out postposition again and handle, can obtain desired object.But, though this kind method has remedied the shortcoming that mobile object is not cut apart, but carry out before the real-time video processing, all must be by artificial interface predefined object, certainly this is quite inconvenient, for this situation, be starved of a kind of simple and easy, method solves the problem that this type of image object is cut apart easily.

Summary of the invention

The present invention addresses the above problem to propose a kind of hierarchy type Object Segmentation method based on MPEG-7, and to dynamic or still image, and any image object that is wherein comprised can be cut apart.The present invention mainly uses watershed segmentation and MPEG-7 descriptor comparison techniques.The main concept of this technology from people in order to the picture mosaic notion of amusement with for the recognition mode of object, human brain can be charged to some the feature prior learnings and the impressionization of object in people's brain cell, when needs are pieced together out an object constantly by picture mosaic, can be very fast risk an object.The present invention promptly utilizes this kind pattern, the precondition computer is recognized some image objects, and extract characteristics of objects to set up needed database, employed feature is to capture according to the Object Descriptor (objectdescriptor) that video section in the MPEG-7 standard defines, whether utilize these features to judge whether the image object of being cut apart is correct afterwards again, be required image object.

Hierarchy type Object Segmentation method based on dynamic image compression standard of the present invention comprises the following step:

Set up the step and the selected step of threshold value initial value of a database, this threshold value initial value is selected according to system's initial threshold;

Import an image, and convert this image color to GTG;

Detecting a GTG gradient minimum value, carry out watershed segmentation, is that the basis begins expansion with this GTG gradient minimum value, up to arriving a critical value, is the boundary line with this critical value again, adds one and separates dam, and this image Segmentation is become a plurality of watershed regions;

Under this threshold value initial value benchmark, begin to merge these a plurality of watershed regions, wherein merge during less than a threshold value when the mean difference of the color-values of these a plurality of watershed regions;

With these a plurality of watershed region numberings after merging;

At a plurality of watershed regions, use a comparator and another decision to replace under result's the threshold value benchmark, find out this most similar watershed region, wherein this comparator is to utilize a similarity match-on criterion, should image and database relatively, the difference value that check is passed back, when surpassing a replacement result threshold value, promptly replace the analog result of this image, and then merge outward, handle this image hollow space and toward interior deletion, and hollow space is treated to the area of a gaps and omissions block of this image less than 2% of this image, promptly fills up this gaps and omissions block, computing always repeatedly realizes a saturated conditions to this image;

Reduce this threshold value, and do the corresponding processing of watershed region, repeat abovementioned steps and satisfy a stop condition up to this threshold value; And

Export an image result.

Above-mentioned hierarchy type Object Segmentation method based on dynamic image compression standard is characterized in that the described step of setting up database, is to utilize a known image object as the basis, uses a descriptor to extract the descriptor feature of this known image object.

Above-mentioned hierarchy type Object Segmentation method based on dynamic image compression standard is characterized in that described descriptor comprises a color descriptor, texture description symbol and a profile descriptor.

Above-mentioned hierarchy type Object Segmentation method based on dynamic image compression standard is characterized in that described color description symbol comprises the combination in any of color space, main color, color statistics, color adjustment, color quantization and color layout.

Above-mentioned hierarchy type Object Segmentation method based on dynamic image compression standard is characterized in that described texture description symbol can be by the combination in any of homogeneous texture and edge statistics.

Above-mentioned hierarchy type Object Segmentation method based on dynamic image compression standard is characterized in that described profile descriptor can be by the combination in any of object housing, area format descriptor, profile formal description symbol and 3 dimensional coil geometry.

Above-mentioned hierarchy type Object Segmentation method based on dynamic image compression standard, when it is characterized in that described system initial threshold compares descriptor for importing block and database, the value when wherein the most similar at most descriptor being arranged to database.

Above-mentioned hierarchy type Object Segmentation method based on dynamic image compression standard, when it is characterized in that described comparator relatively, input imagery to as if be combined into the still original at the beginning input red-green-blue color image of the pixel value in block pixel value wherein by block.

2/3rds of the difference that above-mentioned hierarchy type Object Segmentation method based on dynamic image compression standard, the threshold value that it is characterized in that described replacement result are subtracted each other for relatively descriptor sum and the descriptor sum that is equal to analogical object descriptive data.

Above-mentioned hierarchy type Object Segmentation method based on dynamic image compression standard is characterized in that described saturated conditions is meant that the similar value after the comparison can't improve.

Above-mentioned hierarchy type Object Segmentation method based on dynamic image compression standard is characterized in that the corresponding processing of described watershed region, is gained watershed block after the most similar visual subject area of gained corresponds to the reduction threshold value before is numbered.

Above-mentioned hierarchy type Object Segmentation method based on dynamic image compression standard, it is characterized in that described stop condition can by threshold value be 0 and the combination made by oneself of user in select one arbitrarily.

Above-mentioned hierarchy type Object Segmentation method based on dynamic image compression standard is characterized in that described image result is the most similar result in the entire process process.The explanation of relevant detailed content of the present invention and notion conjunction with figs. as after.

Description of drawings

Fig. 1 is a method flow diagram of the present invention;

Fig. 2 A is a texture description symbol processing mode schematic diagram;

Fig. 2 B is an object rectangle housing descriptor schematic diagram;

Fig. 2 C is the range descriptors schematic diagram;

Fig. 2 D is an object external form descriptor schematic diagram;

The watershed Object Segmentation schematic diagram that Fig. 3 A proposes for the present invention;

Fig. 3 B is applied to the split-run test result of image for Fig. 3 A;

Fig. 3 C carries out the result that the zone merges for Fig. 3 B under a critical value;

Fig. 3 D carries out the result that the zone merges for Fig. 3 B under another critical value;

Fig. 4 A, 4C, 4E are respectively the present invention input imagery are carried out the result that different threshold region merge;

Fig. 4 B, 4D, 4F be respectively the present invention and database relatively after the most similar image object;

Fig. 5 A～5C handles schematic diagram for the corresponding watershed region that the present invention proposes;

Appointed area and adjacent area thereof that Fig. 6 A proposes for the present invention;

Fig. 6 B is " receipts " schematic diagram of the regional selection mechanism of the present invention's proposition;

Fig. 6 C is " putting " schematic diagram of the regional selection mechanism of the present invention's proposition;

Fig. 6 D handles schematic diagram for the hollow space that the present invention proposes;

Fig. 7 A, 7B are the former figure of experiment input imagery of the present invention;

Fig. 7 C, 7D are the objects that the present invention extracts with the MPEG-7 descriptor; And

Fig. 7 E, 7F are respectively the present invention's find out and the most close image object of the former figure of input imagery.

Embodiment

The present invention proposes a kind of hierarchy type Object Segmentation method based on the MPEG-7 descriptor, and the method is to utilize MPEG-7 descriptor and watershed segmentation computing to realize the new method that image object is cut apart.Fig. 1 is the system flow chart of hierarchy type Object Segmentation method based on dynamic image compression standard (MPEG-7) proposed by the invention, in order to the basic procedure that is described as follows:

At first, must determine that the image object that exists has used the technology of MPEG-7 descriptor to finish and set up database, next input imagery and conversioning colour (step 100), carry out watershed segmentation (step 110), and selected first making under the threshold value (threshold), merge divided area (step 120), the standard of this threshold value is to adopt adjacent domain (region) color difference, then handle corresponding watershed region (watershed region) (step 130), the recombinant chosen area, and with database relatively (step 140), this must declare earlier the image object that is combined into by the zone wherein pixel (pixel) value in the block still be original input RGB (Red GreenBlue, RGB) visual pixel value, system constantly compares between regional selection mechanism and database repeatedly, can not be when the most similar image object comparing data be better at present up to relatively result, think that promptly the zone chooses reach capacity (step 150), just continue next step and reduce threshold value (step 160).By after reducing threshold value and execution area and merging machine-processed two steps, this moment, watershed region was done corresponding watershed region processing with the most similar present image object again, again doing first forefoot area in the zone of gained again chooses with object and relatively waits step, system will operate always, just finish when threshold value is 0 (step 170).

Below be divided into several parts and introduce main concept in the structure of the present invention in detail, (1) database is set up; (2) initial threshold is selected; (3) image input and color conversion; (4) watershed segmentation; (5) corresponding watershed region is handled; (6) regional selection mechanism and database are relatively worked, and respectively are described as follows:

(1) database is set up

In order to allow computer be familiar with so-called image object, must do the MPEG-7 descriptor to known object earlier handles, the result is deposited in the database, wherein the standard of setting up of descriptor database is the various descriptor that adopts video section institute standard in the MPEG-7 specifications again.This training method is as the cognitive pattern of the mankind to object, human can prior object of cognition, and then impressionization ground charges in the brain, when time identical object occurs once again instantly, human just can know know what this to liking.Wherein the MPEG-7 descriptor includes color, texture, object profile, mobile (containing the motor pattern of video camera and moving of scenery) etc., the particular content of descriptors such as object color that the present invention that below makes brief of the introduction mainly uses and object profile:

Color (color) can be divided into color space (color space), main color (dominatecolor), distribution of color (color layout), color statistics (color histogram) again, adjust color (scalable color) and several concrete descriptions of color quantization (color quantization):

-color space, as RGB, aberration (Component Video, YCrCb), colourity saturation purity (Hue Saturation Value, HSV) and M[3] [3], be to describe the employed colored substrate of image, M[] [] be the transition matrix of substrate with RGB for other form.

-main color is described the main color of a certain object, and the numerical value and the shared percentage of these main colors are done specific descriptions.The usefulness of parameter as a comparison in the time of so then can be for the retrieval of similar object.

-color statistics, the statistical distribution of each color has with reference to property for the retrieval of similar image.-color quantization, the quantification manner of description color range has three kinds of patterns: linear (linear), non-linear (nonlinear), question blank (lookup table).

Texture (texture) can be divided into homogeneous texture (homogeneous texture), edge statistics (edge histogram).Texture description is described directivity, degree of roughness and the regularity of image thus.When being used for describing image, image is divided into 6 zones, press split along radial direction, radius is divided into 5 parts, and so, a semicircle just has been divided into 30 zones, shown in Fig. 2 A, according to corresponding operation function, do matching operation respectively according to radial direction and circumferencial direction, obtain a result.The object profile.Can summarize object housing rectangle (objectbounding box), region shape descriptor (region-based shape descriptor), shape boundary descriptor (contour-based shape descriptor) and four kinds of patterns of 3 dimensional coil geometry (shape 3Ddescriptor) under general situation describes:

-object housing rectangle shown in Fig. 2 B, is clearly described object in the image with the angle between length-width ratio, place relative position and the object main shaft and the reference axis of the minimum rectangle that can comprise object.

-region shape descriptor shown in Fig. 2 C, is described with the zone that object is shared, and this method can be described comparatively complex objects, as trade mark etc.

-shape boundary descriptor shown in Fig. 2 D, is described exact shape, and in curvature space (the curvature scale space) mode of formatting, and tolerable convergent-divergent, rotation, twists and covers.

(2) initial threshold is selected

Because the watershed is under the zone merges, might merge to the situation of required object, so the present invention sets and decides initial threshold in its sole discretion by system, to find out best starting point, and initial threshold value is selected, when comparing descriptor for importing block and database, the threshold value when wherein the most similar at most descriptor being arranged to database, and the standard of threshold value is to utilize the input imagery color difference of adjacent area.

Determining method is to be under 0 condition from zone merging threshold value earlier, takes out each watershed region and compares with database.Threshold value be 0 handle after, the zone of carrying out so-called watershed segmentation in the threshold value of up adjusting a unit again merges.Therefore, block counts will reduce, and the block area can become greatly, and threshold value is 0 identical comparison step before repeating according to this, takes out each watershed region exactly and compares with database.Go down repeatedly, up to handle block counts be 1 o'clock with regard to the selected work of outage threshold.

(3) input imagery and conversioning colour

System's input imagery adopts the RGB chromatic image.And when doing watershed segmentation, handle for convenience, be at the GTG image processing, so can convert tablet pattern to another general GTG figure earlier before doing watershed segmentation again, and then carry out watershed segmentation, this translative mode is to define with reference to the Y coordinate in the YUV hue coordinate.

YUV color coordinate system is NTSC, the basic color format that PAL and SECAM color television standard are adopted, and Y represents its luminance signal, and U and V are its chrominance signal, and the relation of it and RGB is as shown in (1) formula:

Y＝0.299R+0.587G+0.112B

U＝-0.147R-0.289G+0.434B (1)

V＝0.615R-0.515G-0.1B

(4) watershed segmentation

The watershed is to use the difference of GTG value, and the pixel of some uncertain regions is ranged than similar area.It can be considered is a kind of algorithm of area extension.The summary of watershed method as shown in Figure 3A, at first the minimum value with regional shade of gray detects, begin to do area extension from here up to the water surface during, promptly add and prevent to spread unchecked, so watershed method can be separated into zones of different with image to other regional dam to the summit in basin.But, need carry out the zone merging before finishing so cut apart because watershed method is very responsive for the variation of GTG gradient.

Through the image after the watershed processing, have very many zones, shown in Fig. 3 B, generally can utilize some operation methods to reduce number of regions, form as Fig. 3 C and Fig. 3 D.The present invention utilizes the standard of the worthwhile one-tenth threshold value of the color difference of adjacent area, and the color difference between the adjacent area can be merged during less than selected threshold value.Wherein the present invention is defined as color difference:

Wherein, R1, R2 represent adjacent two regional symbols; R1.R, R2.R represent the average color pixel value of R in the zone; R1.G, R2.G represent the average color pixel value of G in the zone; R1.B, R2.B represent the average color pixel value of B in the zone.

For the accelerating system processing time, earlier begin to handle from selected threshold value, when whenever handling a threshold value,, equal 0 up to threshold value if condition does not conform to and just reduces threshold value, or end when satisfying lowest difference that the user the defines opposite sex.

(5) corresponding watershed region is handled

Because the watershed segmentation method is in order to solve the problem of over-segmentation (over-segmentation), can carry out so-called zone usually and merge, but often because the difference of threshold value, institute merges the result who also can difference.But confirmablely be, threshold value is higher, and left block counts is fewer, and the expression of object is simpler, relative, accuracy also can with reduction.The present invention just utilizes this specific character, imports the hierarchy type segmenting system, to accelerate the system handles time.

It is to carry out the result that the zone merges for 45,30 and 15 times that Fig. 4 A, Fig. 4 C and Fig. 4 E are respectively threshold value, and Fig. 4 B, Fig. 4 D and Fig. 4 F compare in result the most similar image object in every stratum to database through system.

Can find easily that from Fig. 4 A, Fig. 4 C and Fig. 4 E the bigger institute of the threshold value block that forms is bigger,, and carry out the block selection mechanism and database compares, can obtain as Fig. 4 B and the database result of similar image object so select through the threshold value of front.

The present invention utilizes the result of Fig. 4 B and Fig. 4 C, utilizes Fig. 5 A to Fig. 5 C that corresponding watershed region handling process is described.After downgrading threshold value, can be as the watershed region of Fig. 5 B, and do corresponding with the image object area of the most similar before Fig. 5 A, find out the corresponding watershed region of Fig. 5 C ash color part, just find out mark (label) at the relative block of Fig. 5 B according to the image object of Fig. 5 A, more thus relatively mark carry out the block selection mechanism.

The block that the principle of this part is based on the watershed is incorporated under the different threshold values and can produces different plant characteristics, utilizes this kind relation to come references object of feed system.Also mention before and can produce required object when merging and adjacent object is merged into one situation, if under this threshold value, find a object the most close, just keep this result, when reducing threshold value next time, as above said continuous situation will be separated because of block merges to some extent, therefore when continuing execution area and choosing with database relatively, make the result more mate image object in (match) database.

In brief, this corresponding watershed region treatment technology is to set about from general orientation earlier, utilize the zone of watershed segmentation to merge the reduction threshold value, and block can be more and more littler, does further correction as a result again, and making the output result is required result.

(6) regional selection mechanism and database are relatively worked

At first, how introducing system carries out the block selection mechanism.Merge divided area before, the present invention can be again indicates not isolabeling to each zones of these figures, shown in Fig. 3 D, the zone is behind mark, can in signature, choose different zone combinations, so the present invention mainly has " receipts ", " hollow space processing " and " putting " three steps in regional selection mechanism.

Before system begins, by cognitive object must be the object that region area links to each other, so after threshold value initial value before is selected, can obtain a zone close with database data, be referred to as " appointed area ", the appointed area is carried out the action of " receipts " and " putting " again with adjacent area thus.The present invention is called " receipts " with the action of " appointed area " and " adjacent area " synthetic new appointed area, some zone is called " putting " and delete wherein from synthetic appointed area, shown in Fig. 6 A to Fig. 6 C, and between " receipts " action, can produce the one and another block of cells of the gaps and omissions of object own, shown in Fig. 6 D, therefore after " putting " handles, can carry out what is called " hollow space processing " again, filling up, and next just can carry out the action of " receipts " than the zonule.

So-called " hollow space processing ", because the relation of system's block selection mechanism, often produce the one and another block of cells of the gaps and omissions of object own, shown in Fig. 6 D, and these independent block of cells do not have decisive influence to whole result, so after " receipts " mechanism, treat as reference point with the image object that finds, and find out wherein block of cells, if block of cells is in this object, and area was less than 2% o'clock of this image object, and block of cells will be merged in this most similar image object, upgraded the most similar visual object data that (update) " appointed area " data are this moment afterwards.Then, introduce database again and relatively work, groundwork is divided into two sports, is respectively " comparator " design and " replacing mechanism "." comparator " compares in order to pictorial data after will cutting apart and database, and the various descriptors of the zone of selecting before utilizing combination and MPEG-7 the difference of relatively passing back (difference) foundation as a comparison of fixed similarity adaptation function (similaritymatching function), this must declare earlier the image object that is combined into by block wherein the pixel value in the block still be original input RGB chromatic image pixel value.

And these similarity adaptation functions define as the MPEG-7 specification, color statistics (descriptor one of them) is wherein arranged, before doing the color statistical match, data with existing A and comparing data B can be done feature extraction, these extracting modes have detailed introduction at the MPEG-7 specifications.And wherein comparator is exactly the similarity match-on criterion (similarity matchingcriteria) that utilizes descriptor all can define, and the similarity of the MPEG-7 color statistics of two groups of data can add suitable weighted value usually and handle.The characteristic value that under hue coordinate HSV, is extracted for example, just utilize (3) formula to represent weighted value:

$w_{i,j}=1-\sqrt{\frac{{(v\left(i\right)-v\left(j\right))}^2+{(s\left(i\right)\·\cosh \left(i\right)-s\left(j\right)\·\cosh \left(j\right))}^2+{(s\left(i\right)\·\sinh \left(i\right)-s\left(j\right)\·\sinh \left(j\right))}^2}{2}}$

W＝[w _i，j]；0≤i＜number_of_cells；0≤j＜number_of_cells (3)

Suppose hist[A] set of the color statistics of expression data A, hist[B] be the color statistics set of B, the weighted value that calculates before utilizing, just can comparing data A and the color statistics similarity of B, calculating formula is shown in (4) formula, and (A, B) value is more little for similar more then dist.

dist(A，B)＝[hist(A)-hist(B)] ^TW[hist(A)-hist(B)] (4)

The comparator design that the present invention is used, utilize exactly MPEG-7 to every kind image video the various descriptors that can use, all there is one group similarity match-on criterion in these descriptors, wherein the similarity match-on criterion will calculate the otherness of two groups of data.And the comparator source is exactly the difference that the similarity match-on criterion is passed back, is used as the foundation that image object is chosen.

Relatively work down in regional selection mechanism and database, as long as comparative result has when surpassing " threshold value of replacing the result ", system can keep this result and get off, that is to say designated blocks and adjacent block will be included in and be the most similar image object, and this process is referred to as " replacing mechanism ", " replaces result's threshold value " and be defined as follows among this:

Wherein, CN is the descriptor sum less than analogical object descriptive data; Total_Number_Descriptor is descriptor sum relatively; SN is the descriptor sum that is equal to analogical object descriptive data.

Because the characteristic of various descriptor institute foundation differs, if must be when all descriptors be all admitted replaceable image object, with regard to then more not objective on the convention.So the present invention's regulation promptly can be replaced when surpassing " threshold value of replacing the result ".

Summing up region in front selection mechanism and database relatively works, can find a best region earlier at the beginning, be called " appointed area ", the zone is carried out " receipts " zone and is chosen thus again, and compare work with database, under " receipts " action, find out one group and limit down similar in appearance to the database image object and above the mechanism of replacement, will change present appointed area into this most similar visual subject area.The appointed area that next step is new is according to this again done the regional selection mechanism of " receipts " again and is gone down, up to the result data that merges the block gained all fail than at present analog result come good the time just stop, and " receipts " saturation condition will be set as saturated this moment.

Next carry out " hollow space processing ", if detected the gaps and omissions zonule, will mend fullly, and that the saturation condition of " receipts " will be set as will be unsaturated, and otherwise, the saturation condition of " receipts " still is saturated.

Then the block from the appointed area of having filled up is eliminated one by one again, the better words of data are just replaced mechanism before result's ratio of being eliminated, be same as the action of " receipts ", the block of " putting " is chosen computing repeatedly always, until just stop when can not find best result, this " puts " saturation condition and also can be set as saturated simultaneously.And the last the most similar image object that stays similarly object related data be all this and " put " down optimum.

System can carry out " receipts ", " hollow space processing " and " putting " step always repeatedly, until " receipts " of being carried out and " putting " saturation condition be all and saturatedly just stop, above so-called saturation condition for the result of " receipts " or " putting " gained can not be than present optimum data good the time just stop.

Reduce threshold value afterwards again, get back to corresponding watershed region and handle, find out the appointed area again, further carry out block and choose, further look for more approaching object.

The present invention is as follows with the graphic result of two examples explanation after the present invention is cut apart:

See also Fig. 7 A and Fig. 7 B, the mother and daughter that the woman that this two figure is respectively the 176*144 pixel schemes original input imagery and 176*144 pixel schemes original input imagery, Fig. 7 C and Fig. 7 D are depicted as and do MPEG-7 descriptor extraction process object in advance, and also there has been the object in the database in result, Fig. 7 E and Fig. 7 F are depicted as the image object that utilizes structure of the present invention to find out, and are respectively woman figure and mother and daughter figure and find out the result of close image object.

For the image object of being imported, method proposed by the invention can find a quite approximate image object, because the characteristic of decision error can not appear because of image in most descriptor in the MPEG-7 descriptor when receipts are sought comparison through rotation, therefore content-based receipts are sought (content-base retrieval) and image object is cut apart the aspect for being applied to, and can use the method for native system to handle.When having the information of image and relative MPEG-7 Object Descriptor in the database, can utilize the MPEG-7 database search comparison image that desire is searched to find out, and can utilize method proposed by the invention simultaneously, come image in the database is done to cut apart, and then find out the object of database image.

Though the present invention discloses as above with aforesaid preferred embodiment; but it is not in order to limit the present invention; the those of ordinary skill of the industry without departing from the spirit and scope of the invention; can carry out variation and retouching that various think ofs easily reach, so protection scope of the present invention should be with being as the criterion that claims define.

Claims (13)

1, a kind of hierarchy type Object Segmentation method based on dynamic image compression standard is characterized in that this method comprises the following step:

Set up the step and the selected step of threshold value initial value of a database, this threshold value initial value is selected according to system's initial threshold;

Import an image, and convert this image color to GTG;

Detecting a GTG gradient minimum value, carry out watershed segmentation, is that the basis begins expansion with this GTG gradient minimum value, up to arriving a critical value, is the boundary line with this critical value again, adds one and separates dam, and this image Segmentation is become a plurality of watershed regions;

Under this threshold value initial value benchmark, begin to merge these a plurality of watershed regions, wherein merge during less than a threshold value when the mean difference of the color-values of these a plurality of watershed regions;

With these a plurality of watershed region numberings after merging;

At a plurality of watershed regions, use a comparator and another decision to replace under result's the threshold value benchmark, find out this most similar watershed region, wherein this comparator is to utilize a similarity match-on criterion, should image and database relatively, the difference value that check is passed back, when surpassing a replacement result threshold value, promptly replace the analog result of this image, and then merge outward, handle this image hollow space and toward interior deletion, and hollow space is treated to the area of a gaps and omissions block of this image less than 2% of this image, promptly fills up this gaps and omissions block, computing always repeatedly realizes a saturated conditions to this image;

Reduce this threshold value, and do the corresponding processing of watershed region, repeat abovementioned steps and satisfy a stop condition up to this threshold value; And

Export an image result.

2, the hierarchy type Object Segmentation method based on dynamic image compression standard as claimed in claim 1, it is characterized in that the described step of setting up database, be to utilize a known image object, use a descriptor to extract the descriptor feature of this known image object as the basis.

3, the hierarchy type Object Segmentation method based on dynamic image compression standard as claimed in claim 2 is characterized in that described descriptor comprises a color descriptor, texture description symbol and a profile descriptor.

4, the hierarchy type Object Segmentation method based on dynamic image compression standard as claimed in claim 3 is characterized in that described color description symbol comprises the combination in any of color space, main color, color statistics, color adjustment, color quantization and color layout.

5, the hierarchy type Object Segmentation method based on dynamic image compression standard as claimed in claim 3 is characterized in that described texture description symbol can be by the combination in any of homogeneous texture and edge statistics.

6, the hierarchy type Object Segmentation method based on dynamic image compression standard as claimed in claim 3 is characterized in that described profile descriptor can be by the combination in any of object housing, area format descriptor, profile formal description symbol and 3 dimensional coil geometry.

7, the hierarchy type Object Segmentation method based on dynamic image compression standard as claimed in claim 1, when it is characterized in that described system initial threshold compares descriptor for importing block and database, the value when wherein the most similar at most descriptor being arranged to database.

8, the hierarchy type Object Segmentation method based on dynamic image compression standard as claimed in claim 1, when it is characterized in that described comparator relatively, input imagery to as if be combined into the still original at the beginning input red-green-blue color image of the pixel value in block pixel value wherein by block.

9, the hierarchy type Object Segmentation method based on dynamic image compression standard as claimed in claim 1 is characterized in that described replacement result's threshold value is: 2/3rds of the difference that descriptor sum relatively and the descriptor sum that is equal to analogical object descriptive data subtract each other.

10, the hierarchy type Object Segmentation method based on dynamic image compression standard as claimed in claim 1 is characterized in that described saturated conditions is meant that the similar value after the comparison can't improve.

11, the hierarchy type Object Segmentation method based on dynamic image compression standard as claimed in claim 1, it is characterized in that the corresponding processing of described watershed region, is that gained watershed block after the most similar visual subject area of gained corresponds to the reduction threshold value before is numbered.

12, the hierarchy type Object Segmentation method based on dynamic image compression standard as claimed in claim 1, it is characterized in that described stop condition can by threshold value be 0 and the combination made by oneself of user in select one arbitrarily.

13, the hierarchy type Object Segmentation method based on dynamic image compression standard as claimed in claim 1 is characterized in that described image result is the most similar result in the entire process process.

Images