CN101246554A - Multi-object Image Segmentation Method Based on Pixel Labeling - Google Patents

Abstract

The invention discloses a multi-target image segmentation method based on pixel marking. The specific steps are as follows: using an image captured by a video imaging sensor as an input signal; performing binary processing on the image; scanning the processed image for the first time; Establish an equivalent array; perform the second scan to obtain information such as target quantity, area, perimeter, and position coordinates; adopt the present invention to carry out target marking and edge marking on the binary image of the full field of view, which has certain versatility. The next step is to calculate the feature quantity of the target and high-end processes such as image understanding provide sufficient conditions; and the present invention is not limited by the number of targets, and can also reduce or increase the area of the processing area at will according to the configuration of hardware resources, with Great flexibility.

Description

Multi-object Image Segmentation Method Based on Pixel Labeling

technical field

The invention belongs to the field of photoelectric detection and tracking measurement, in particular to a description method based on multi-target segmentation, which is used for multi-target marking and feature quantity extraction in imaging video tracking.

Background technique

In image processing, image segmentation is a very important part. Through image segmentation, the target (also called foreground) that the user cares about in the image can be extracted and the necessary data can be provided for future image understanding. The methods to achieve image segmentation mainly include: edge-based image segmentation methods, region-based image segmentation methods, and comprehensive segmentation methods based on the above two types of methods.

One of the purposes of image segmentation is to lay the foundation for the realization of automatic target recognition, so extracting the feature quantity of the target is of vital significance to the system. For this reason, parameters such as the area, perimeter, centroid coordinates, and even the position coordinates of all pixels of the target object after image segmentation need to be obtained simultaneously. In order to realize this function, commonly used methods include run-length connectivity analysis and pixel labeling. The run-length connectivity analysis method is a method of marking the target by analyzing the run-length connectivity obtained by continuous scanning lines. It is necessary to process the image in advance to obtain pixel blocks with the same gray level. The condition is that the same gray level has been formed before scanning. A block of pixels is a run. The pixel labeling method scans the binary image without requiring any prior information, and determines which target the pixel belongs to according to the connectivity between pixels, and finally all pixels have a mark which target the pixel belongs to. Feature amount. Afterwards, the area and centroid coordinates of the target can be obtained by counting the pixels of each target.

Contents of the invention

The technical problem to be solved by the present invention: the present invention expands the traditional pixel marking method. This method makes all pixels not only have the mark of which target they belong to, but also the mark of whether they are target boundary points. After counting these points , the perimeter of the single-pixel-wide boundary of the target can be obtained, avoiding the disadvantage of obtaining inaccurate perimeters by performing statistics on multi-pixel wide boundaries after edge detection; due to the important position of the perimeter in target measurement (such as circularity measurement), this improvement is meaningful.

The technical solution adopted by the present invention to solve the technical problems is: a multi-target image segmentation method based on pixel marking, which is characterized in that it comprises the following steps:

(1) Take the image captured by the video imaging sensor as the input signal (expressed by lpraw);

(2) Obtain a binary image (indicated by lpImg01) through threshold segmentation on the taken image;

Let lpImg01[i][j] represent the pixel value of the pixel at coordinate (i, j), lpImg01[i][j]=0 means point (i, j) is part of the background, lpImg01[i][j]=1 Indicates that the point (i, l) belongs to a part of the target area; at the same time, use Perimeter[m] to represent the perimeter of the m-th target; Area[m] represents the area of the m-th target; Posx[m] represents the m-th target The x position coordinates in the entire image; Posy[m] indicates the y position coordinates of the mth target in the entire image; Object[i][j] indicates which target the pixel of coordinates (i, j) belongs to; Edge [i][j] indicates whether the pixel with coordinates (i, j) is the boundary point of the target; Equ[i][j] indicates that the pixels whose target attributes are i and j belong to the same target; Access[i] indicates the first After the second scan, the pixel whose target attribute is i actually belongs to the Access[i] target;

(3) scan the obtained binarized image for the first time;

Scanning is carried out from left to right and from top to bottom, and the variable TTs is used to mark the number of all different target marks in the first scan. When scanning reaches t, the points at positions T ₁ ~ T ₄ have been scanned Therefore, the relationship between t and these four points must be considered. When the gray level of t is 1, that is, t is the target, the connection relationship between the current scanning point and its 8 neighbors can be expressed as:

(1) When the gray values at positions T ₁ ~ T ₄ are all zero, it means that this point has no connectivity with the previous 4 adjacent pixels, and a new mark is given to the current pixel, that is, TTs plus 1 , and assign TTs to Object[i][j] of the current pixel t, that is, Object[i][j]=TTs;

(2) When there is one and only one gray value at positions T ₁ to T ₄ that is 1, let T _m (m=1, 2, 3, 4&T _m non-zero) be the pixel whose gray value is 1, then Assign the Object value of T _m to the Object value of t, indicating that the target attribute of the current point is the same as the target attribute of the Tm pixel, and belongs to the same target as the Tm pixel;

(3) When more than one of the gray values at positions T ₁ to T ₄ is 1, assign the smallest Object value in Tm (m=1, 2, 3, 4 & Tm is non-zero), that is, Object_min, to the current point t Object (set Object_min to represent the smallest Object value in Tm, and set Object_max to represent the largest Object value in Tm); at the same time, set Equ[Object_min][Object_max]=1 in the equivalence matrix (wherein, Object_min<Object_max), which means The pixels whose Object values are Object_min and Object_max in the first scan actually belong to the same object.

At the same time, for the boundary attribute of the pixel, set the currently scanned pixel as t, as long as at least one grayscale in T _m (m=1～8) is zero, then set Edge[i][j] of t= 1, that is, set the boundary attribute of the pixel whose coordinate position is (i, j) in the image to 1, indicating that the point is a boundary point; otherwise, set Edge[i][j]=0 to indicate a non-boundary point. Therefore, after the first scan, the edge information of the target single-pixel width is obtained.

(4) carry out planning equivalent array processing to the discrete equivalent matrix obtained after the first scan;

After the first scan, all target attributes belonging to the same target are marked as TT ₁ , TT ₂ , ..., TTn, where TT ₁ <TT ₂ <...<TTn, the equivalent array has the following three possibilities :

(1) All target attributes already have Equ[TT ₁ ][TT _i ]=1, i=2, 3,..., n, this situation does not need to be further processed;

(2) Equ[TT ₁ ][TT _i ]=1, Equ[TT _i ][TT _j ]=1, then Equ[TT ₁ ][TT _j ]=1;

(3) Equ[TT ₁ ][T _j ]=1, Equ[TT _i ][TT _j ]=1, then Equ[TT ₁ ][TT _i ]=1;

After the above algorithm processing, the expected requirements can be fully met, and full preparations have been made for the second scan; at this time, all the different marks belonging to the same target have established an equivalence relationship with the smallest mark belonging to this target.

(5) Then perform the second scan; obtain target information such as the area, perimeter, and position coordinates of the desired target;

After the previous first scan and the planning of the equivalent array, the second scan is performed; this process is completed in two steps. Step 1: According to the equivalent array Equ after the previous processing, all pixels belonging to the same target, the target The attribute value Object should be the same, and it should be assigned as the smallest mark belonging to this target after the first scan, and record the number of the target pointed to by the smallest mark in the current target attribute from top to bottom and from left to right in the image Target. Step 2 is to do the final finishing work. According to the Access array, assign the target attribute Object of the pixel to the real target serial number, that is, for the target attribute value Object of all pixels, if Access[Object[i][j]]>0, then Object[i][j]=Access[Object[i][j]]; otherwise, the original value of Object remains unchanged; at the same time, according to the interaction between Object[i][j] and Edge[i][j] Relationships, statistics of the target characteristic parameters such as the actual area, perimeter and centroid coordinates of each target, and outline the target edge.

Compared with the prior art, the present invention has the following advantages:

1. Using the method based on 8-neighborhood pixel marking, the edge information of the single-pixel width of the target is obtained; thus the perimeter of the single-pixel-wide boundary of the target is obtained, which avoids inaccurate statistics of multi-pixel wide boundaries after edge detection the disadvantage of the perimeter;

2. The present invention can accurately provide information such as the target area and position coordinates, which lays a foundation for further image analysis and processing. The target feature extraction is accurate, and the method is simple, real-time and practical;

3. The number of objects counted by the present invention can be as many as desired, only limited by the hardware resources in the real-time system.

Description of drawings

Fig. 1 is the flow chart of pixel labeling algorithm in the present invention;

Fig. 2 is the scan relationship diagram based on 8 neighborhood pixel marks in the present invention;

In the figure: t represents the currently scanned pixel, and T ₁ ~ T ₈ represent the adjacent pixels of t pixel in 8 connected regions;

Fig. 3 is a flow chart of the second scanning step 1 in the present invention.

Detailed ways

The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments.

The invention realizes multi-target marking, obtains characteristic signals such as target circumference, area, position coordinates and the like while obtaining a single-pixel-wide target boundary.

The flow chart of its implementation method is shown in Figure 1, and its specific implementation steps are as follows:

(1) First, the image captured by the video imaging sensor is used as an input signal;

(2) Then the original input video signal is binarized to generate a binary image represented by '0' and '1';

(3) Immediately afterwards, scan the binarized image from left to right and from top to bottom for the first time (the starting point of the image is set at the upper left of the image);

The function of the first scan includes: marking each pixel of the binarized image for the first time and recording the target attribute values of the regions with different target attributes according to the connectivity, and marking the boundary attributes belonging to the target edge. The specific implementation method As follows: assume that the variable TTs is used to mark the number of all different target marks in the first scan, and lpImg01[i][j] represents the pixel value of the coordinate (i, j) pixel. lpImg01[i][j]=0 indicates that the point (i, j) is a part of the background, and lpImg01[i][j]=1 indicates that the point (i, j) belongs to a part of the target area. At the same time, Object[i][j] indicates which target the pixel of coordinate (i, j) belongs to; Edge[i][j] indicates whether the pixel of coordinate (i, j) is the boundary point of the target; Equ[i][ j] indicates that the pixels whose target attribute is i and j belong to the same target; Access[i] indicates that the pixel whose target attribute is i actually belongs to the Access[i] target after the first scan; in order to mark the currently scanned pixel, use The 8-neighborhood pixel marking method needs to check the connectivity between the pixel and the four adjacent pixels scanned before it, as shown in Figure 2; in Figure 2, t represents the currently scanned pixel, and T ₁ ~ T ₈ represent the t pixel in the Neighboring pixels of 8 connected regions. Scanning is carried out from left to right and from top to bottom. When scanning to t, the points at positions T ₁ ~ T ₄ have been scanned, so the relationship between t and these 4 points must be considered. The gray level at t is 1 is when t is the target:

(1) When the gray values at positions T ₁ ~ T ₄ are all zero, it means that this point has no connectivity with the previous 4 adjacent pixels, and a new mark is given to the current pixel, that is, TTs plus 1 , and assign TTs to Object[i][j] of the current pixel t, namely Object[i][j]=TTs;

(2) When there is one and only one gray value at positions T ₁ to T ₄ that is 1, let T _m (m=1, 2, 3, 4&T _m non-zero) be the pixel whose gray value is 1, then Assign the Object value of T _m to the Object value of t, indicating that the target attribute of the current point is the same as the target attribute of the pixel T _m , and belongs to the same target as the pixel Tm;

(3) When more than one of the gray values at positions T ₁ to T ₄ is 1, assign the smallest Object value in T _m (m=1, 2, 3, 4 & T _m is non-zero), that is, Object_min, to the current point Object value of t (here Object_min represents the smallest Object value in T _m , and Object_max represents the largest Object value in T _m ). At the same time, set Equ[Object_min][Object_max]=1 in the equivalence matrix (wherein, Object_min<Object_max), indicating that the pixels whose Object values are Object_min and Object_max in the first scan actually belong to the same object.

At the same time, for the boundary attribute of the pixel, set the currently scanned pixel as t, as long as at least one grayscale in T _m (m=1～8) is zero, then set Edge[i][j] of t= 1, indicating that the point is a boundary point; otherwise, set Edge[i][j]=0 to indicate a non-boundary point. Therefore, after the first scan, the edge information of the target single-pixel width is obtained.

(4) carry out planning equivalent array processing to the discrete equivalent matrix obtained after the first scan;

Planning the equivalent array is the core of the present invention; its function is to organize the discrete Equ matrix obtained after the first scan, and to classify all target attributes of the same target area into one category. Specifically considering the real-time and simplicity of the second scan, after classifying all the target attributes of the same target into one category, it is necessary to ensure that all these target attributes in the Equ matrix are equivalent to the smallest target attribute.

The process of planning an equivalent array is as follows: For all target attributes belonging to the same target, marked as TT ₁ , TT ₂ , ..., TTn, where TT ₁ <TT ₂ <...<TTn, there are three types of equivalent arrays possible:

(1) All target attributes already have Equ[TT ₁ ][TT _i ]=1, i=2, 3,..., n, this situation does not need to be further processed;

(2) Equ[TT ₁ ][TT _i ]=1, Equ[TT _i ][TT _j ]=1, then Equ[TT ₁ ][TT _j ]=1;

(3) If Equ[TT ₁ ][TT _j ]=1 and Equ[TT _i ][TT _j ]=1, it is necessary to make Equ[TT ₁ ][TT _i ]=1.

After the above planning equivalent array processing, all the pixels belonging to the same target are grouped into one target, which is fully prepared for the second scan. At this time, all the different marks belonging to the same target establish an equivalence relationship with the smallest mark belonging to this target.

(5) Carry out the second scan;

The second scan is completed in two steps. Step 1: According to the equivalent array Equ after the previous processing, all pixels belonging to the same target should have the same target attribute value Object, and assign it to belong to this target after the first scan. At the same time, it is recorded that the target pointed to by the minimum mark in the current target attribute is the number target from top to bottom and from left to right in the image; the algorithm implementation of step 1 is shown in Figure 3, and the integer variable ObjectNumber is set to represent the actual value of the image. The number of targets, the Access[] array initialization is all zero; the key of step 1 is to distinguish which pixel mark is the real new target, which pixel mark is just one of several different target attributes of the same target, since the scan is from left to Right, from top to bottom, it is obvious that when a target attribute has no equivalent attribute smaller than it, it must represent the emergence of a real new target, and this attribute is set as the initial attribute of the new target. Therefore, it is only necessary to find whether there is an equivalent attribute smaller than the target attribute of the current pixel in the planned equivalence matrix to determine whether the target attribute of the current pixel is the starting attribute of the new target. At the same time, when scanning encounters a pixel containing the starting attribute of a new object for the first time, it must be the first pixel encountered by the new object in the scanning direction. At this time, the actual number of objects ObjectNumber should be increased by 1, and ObjectNumber The target attribute assigned to the current pixel, namely Access[Object[i][j]]=ObjectNumber. After step 1, the image has a total of ObjectNumber objects, and each object has only one object ID, which is the minimum value among all object representations originally belonging to this object. At the same time, the Access array records which targets these identifiers actually correspond to.

Step 2 is the final finishing work, according to the Access array, assign the target attribute Object of the pixel to the real target serial number. That is, for the target attribute value Object of all pixels, if Access[Object[i][j]]>0, then Object[i][j]=Access[Object[i][j]]; otherwise, Object keeps the original value constant.

(6) Finally obtain target information such as the area, perimeter, and position coordinates of the desired target.

Let Perimeter[m] represent the perimeter of the m-th target; Area[m] represent the area of the m-th target; Posx[m] represent the x-position coordinates of the m-th target in the entire image; Posy[m] Indicates the y position coordinates of the mth target in the entire image; according to the relationship between Object[i][j], the total number of targets TTs and Edge[i][j], the actual area and circumference of each target are counted Target characteristic parameters such as length and centroid coordinates, the specific process is: within the entire image range, respectively perform the following operations on the pixels whose target attribute is TTi (where TTi is greater than zero and less than or equal to TTs): count the pixels whose target attribute is TTi The sum of the pixels is stored in Area[TTi], that is, the area of the current target is obtained; the number of single-pixel-wide boundary points Edge[i][j]=1 is counted, and stored in Perimeter[TTi], That is, the perimeter of the current target is obtained; the row direction count i and the column direction count j when the target attribute is TTi are accumulated respectively, and stored in the temporary variables SPosx[TTi] and SPosy[TTi] respectively, and then SPosx[TTi ] divided by Area[TTi], the result is stored in Posx[TTi], and the position Posx[TTi] in the direction of the current target row is obtained; SPosy[TTi] is divided by Area[TTi], and the result is stored in Posy[TTi] , get the position Posy[TTi] of the current target column direction, and finally, the centroid coordinates of the current target in the whole image are (Posx[TTi], Posy[TTi]).

Claims (6)

1. multiple goal image partition method based on element marking is characterized in that may further comprise the steps:

(1) image that is absorbed with the video imaging sensor is represented with lpraw as input signal;

(2) carry out Threshold Segmentation and obtain binary image; If lpImg01[i] [j] expression binary image coordinate (i, the j) pixel value of pixel, lpImg01[i] and [j]=0 expression point (i j) be the part of background, lpImg01[i] [j]=1 represents that (i j) belongs to the part of target area to point;

(3) binary image to gained carries out the scanning first time;

(4) plan array of equal value, be about to all marks of equal value and be included into equivalent set;

(5) carry out the scanning second time then;

(6) obtain target informations such as required area, girth, position coordinates.

2. a kind of multiple goal image partition method according to claim 1 based on element marking, it is characterized in that: when in the described step (3) the gained binary image being carried out scanning the first time, from left to right scan from top to bottom, the starting point of scanning is positioned at the upper left side of image.

3. a kind of multiple goal image partition method according to claim 1 based on element marking, it is characterized in that: when in the described step (3) the gained binary image being carried out scanning the first time, adopt 8 neighborhood territory pixel labeling methods, the number that indicates all the different target labels in the scanning for the first time with variable TTs, suppose that current scan point is t, in the gray scale of the point of t position is 1 to be the point of t position when being target, and the connected relation of current scan point and its 8 neighborhood can be expressed as:

(1) works as T ₁～T ₄When the gray-scale value of position is zero, represent that this point 4 pixels adjacent with the front do not have connectedness, then give current pixel a new mark, promptly TTs adds 1, and TTs is given the Object[i of current pixel t] [j], i.e. Object[i] [j]=TTs;

(2) work as T ₁～T ₄The gray-scale value of position has and has only one to be at 1 o'clock, establishes T _m(m=1,2,3,4﹠amp; T _mNon-zero) for that gray scale is 1 pixel, then with T _mThe Object value Object value of giving t, represent that current point target attribute is identical with the objective attribute target attribute of Tm pixel, and belong to same target with the Tm pixel;

(3) work as T ₁～T ₄The gray-scale value of position has that to surpass one be at 1 o'clock, then with Tm (m=1,2,3,4﹠amp; The Tm non-zero) Object value minimum in is the Object that Object_min gives current some t, simultaneously at the mid-Equ[Object_min of equivalent matrice] [Object_max]=1, to be that the pixel of Object_min and Object_max is actual belong to same target to the Object value in the expression scanning for the first time; Wherein Object_min represents Object value minimum among the Tm, and Object_max represents Object value maximum among the Tm, and Object_min＜Object_max.

4. a kind of multiple goal image partition method according to claim 1 based on element marking, it is characterized in that: when carrying out scanning the first time in the described step (3), all 8 neighbors that to each gray scale are 8 connected regions of 1 current point are made following statistics: if having at least one to be background dot in 8 pixels, it is 0 that a gray scale is promptly arranged, then current point belongs to the object boundary point, otherwise belongs to the target internal point; Through what obtain after current the processing is single pixel edge.

5. a kind of multiple goal image partition method according to claim 1 based on element marking, it is characterized in that: planning array of equal value is put the resulting discrete equivalent matrice in scanning back for the first time in order in the described step (4), and all objective attribute target attributes of same target area are classified as a class; After planning array manipulation of equal value, all pixels that belong to same target all have been classified as a target.

6. a kind of multiple goal image partition method based on element marking according to claim 1 is characterized in that: the scanning second time of described step (5) can realize by two steps:

Step 1 is according to previous processed array of equal value later, every pixel that belongs to same target, its objective attribute target attribute value should be identical, and assignment belongs to the minimum mark of this target after for scanning for the first time, writes down this object simultaneously and be in the image from top to bottom which target from left to right;

Step 2, do last tailing in work, according to objective attribute target attribute after the scanning for the first time is actual which target that belongs to of pixel of i, give real target sequence number at last with the objective attribute target attribute of pixel, simultaneously according to objective attribute target attribute, and the mutual relationship between the border, add up the target property parameters such as real area, girth and center-of-mass coordinate of each target, and sketch the contours of the object edge profile.

Images