JPH1049684A - High-speed moment calculation device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To make it possible to calculate the moment statistics of an object in an image having pixels fast through a simple hardware by calculating and storing the moment along row or to the column of an input image or moment image, and extracting the moment at a key pixel position. SOLUTION: An object pixel set is replaced into a solid shape to form a mask (101). Then, the mask and input image 100 are used to select a key pixel on a solid shape for each object. To hole the statistics of respective objects separate, the moment along the columns of the input image or moment image is calculated by using a mask and stored. Further, to hold the statistics of the respective objects separate, the moment of rows of the input image or moment image is calculated by using the mask. Then, the moment at the key pixel position is extracted (107).

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a high-speed moment calculating apparatus used for real-time visual processing of object detection, identification, or tracking, for example, monitoring, automatic control of vehicles, automatic control of robots, and the like. , A high-speed moment calculation device used in a defense system.

[0002]

It is known that moments based on statistics are becoming increasingly important in application to image processing. In particular, they are expected as simple models of objects. The standard moment equation for discrete is
In Cartesian coordinates:

[0003]

(Equation 1) Here, f [x, y] is an image, p and q are moment orders, and X and Y are image sizes. Since a standard CPU performs calculations serially, it is convenient to implement equation (1) for one-dimensional data scanning along the following coordinate directions.

[0004]

(Equation 2)

[0005]

(Equation 3) The serial processor can determine the moment by determining the sum of each column of the image in equation (2) and the sum of the column sums in equation (3).

[0006] An important consideration in object analysis is that images typically include one or more important objects. Calculations that lead to the use of a window surrounding a large number of objects will result in a combined moment for all objects. Therefore, standard approaches generally perform labeling for the separation of various objects in the image. The moment can be calculated from a table listing the position and value of each pixel associated with the object. This is shown in FIG. 2, where the ship and the plane as objects (109) are labeled 1 and 2 respectively (115). Alternately, a rectangular area is segmented from the screen and moments are calculated on the segmentation window.

The central moment of the object applies to the calculation of the moment using a coordinate system having an origin coincident with the center of gravity (COM) of the object. There are a number of useful object dimensions derived from elliptical images, radius of gyration, distortion, kurtosis, and other higher-dimensional central moments. A straight forward approach to calculating the central moment, therefore, determines the COM and arranges the coordinate system before determining the central moment. COM is typically x
And in the y-direction are calculated as follows:

[0008]

(Equation 4) Where M ₁₀ and M ₀₁ are the primary projection moments,
₀₀ is the 0th moment.

Recently, research has been attempted to find an effective method of adapting an image processing algorithm to a linear processing array (LPA) structure. Typical hardware is SIMD (Single Instruction Multiple Di
mension: single instruction, multi-dimensional) will contain a linear array. Each element receives the same set of instructions from the control unit and processes one column of image data. The obvious advantage is that each column of data is processed simultaneously. This is useful for moments based on calculations where the first scan is repeated for each column of data.

[0010]

The most important problem in the standard moment-based approach is its computational cost (addition of a single scan of length N and multiplication of mN (mN
Is the order of the moment)). This can limit the performance of moment calculations in a simple system, which is expensive in performing multiplications. In all systems,
Calculations can be performed intensively for large images, such as a 640 × 480 pixel frame of standard video. Furthermore, the computational load increases significantly when a combination of 0th, 1st and 2nd order (9 permutations) is required, as well as some sets of 3rd and 4th moments.

The presence of a large number of objects in an image significantly increases the cost of performing a moment calculation. The labeling and segmentation steps involved in the process,
It takes a long time when the object has a considerably complicated shape. Calculating some permutations of the moment order for each of these objects will further increase the load. The introduction of a central moment simply increases the complexity of the system, since each object has to be calculated using a different coordinate system. Similarly, the center of gravity (C
OM) will limit the use of a simpler, difficult division system in equation (4).

Dedicated digital optical hardware is designed to efficiently perform object moment calculations. However, it is difficult to justify the use of such forms, as they tend to be unable to perform other, higher order functions. It would be better to develop techniques for general purpose processors or dedicated hardware, such as a linear processor array (LPA), capable of performing the field of image processing algorithms. There are essentially two issues to consider here in order to make LPA hardware efficient for computed moments. First, the operation of each step of the calculation needs to be performed in parallel as much as possible. For example, a horizontal scan of the computed moment in the x-direction is performed only once. Second, the hardware should process the most appropriate object at each iteration. Performing multiple object schemes, including region segmentation or labeling, is not efficient on LPA due to the fixed window size.

[0013]

According to a first aspect of the present invention, a fast moment calculator for calculating moment statistics of a plurality of objects in an image having a plurality of pixels for use in a digital processing system. In the apparatus, using a means for forming a mask by replacing the object pixel set to a solid shape, using the mask and the input image,
Calculate moments along the input image or moment image columns using a means to select key pixels on a solid shape for each object and a mask to maintain separation between statistics for each object Means for calculating and accumulating moments for a row of the input image or moment image using a mask to maintain a separation between statistics for each object; and calculating moments at key pixel locations. And a means for extracting the moment.

According to a second aspect of the present invention, there is provided a high-speed moment calculating apparatus used in a digital processing system for calculating moment statistics of a plurality of objects in an image having a plurality of pixels. Means for forming a mask by replacing the set with a solid shape; means for selecting key pixels on the solid shape for each object using the mask and the input image; and Means for determining the moment for the input image or sequence of moment images by combining the moments from the scan and the upward scan of the image, and using a mask to maintain a separation between the statistics of each object; To combine the moments from scanning the right of the image and scanning the image to the left, and to maintain a separation between the statistics for each object By and the use of the mask,
A high-speed moment calculating device is provided, which has means for obtaining a moment for a row of an input image or moment image and means for extracting a moment at a key pixel position.

According to a third aspect of the present invention, there is provided a high-speed moment calculating apparatus for calculating moment statistics of a plurality of objects in an image having a plurality of pixels, which is used in a digital processing system. Means for forming a mask by replacing the set with a solid shape; means for selecting key pixels on the solid shape for each object using the mask and the input image; and Means for determining the moments for the input image or sequence of moment images by combining the pixel center moments from the scan and upward scan of the image, and by using a mask to maintain a separation between the statistics of each object. Combining pixel center moments from scanning the image to the right and scanning the image to the left, By using a mask to maintain a moment, a high-speed moment calculating apparatus is obtained, comprising: means for obtaining a moment for a row of an input image or a moment image; and means for extracting a moment at a key pixel position. Can be

According to a fourth aspect of the present invention, there is provided a high-speed moment calculating apparatus used in a digital processing system for calculating moment statistics of a plurality of objects in an image having a plurality of pixels. Means for forming a mask by replacing the set with a solid shape; means for selecting a key pixel on the solid shape for each object using the mask and the input image; and Finding the pixel center moment at each pixel of the scan as a linear combination of the lower order moments and all the moments at the preceding pixel, and scanning the image down and the image up By combining the moments and using masks to maintain the separation between the statistics for each object, the input or moment image Means for determining the moment relative to the current pixel, determining the pixel center moment at each pixel of the scan as a linear combination of the lower order moment at the current pixel and all moments at the preceding pixel, By combining pixel center moments from scanning to the left and scanning the image to the left, and using masks to maintain separation between statistics for each object.
A high-speed moment calculating device is provided, which has means for obtaining a moment for a row of an input image or moment image and means for extracting a moment at a key pixel position.

According to a fifth aspect of the present invention, there is provided a high-speed moment calculating apparatus used in a digital processing system for calculating moment statistics of a plurality of objects in an image having a plurality of pixels. Means for forming the mask by replacing the set with a solid shape, using the mask to maintain a separation between the statistics of each object, and scanning down the image and up the image Determining the maximum value of the pixel center moment;
Finding the maximum value of the pixel center moment from the rightward and leftward scans of the vertical moment image, and using the rightward, downward, leftward, and upward scans to determine the maximum on each object Means for selecting a key pixel on a solid shape using the determined center of gravity by searching for a value and ranking the objects by size and truncating the smallest objects below the threshold. Determining the pixel center moment at each pixel of the scan as a linear combination of the lower order moment at the current pixel and all moments at the previous pixel, scanning down the image and above the image By combining the pixel center moments from scanning to and by using a mask to maintain the separation between the statistics of each object, the sequence of the input or moment images Means for determining the moment against,
Finding the pixel center moment at each pixel in the scan as a linear combination of the lower order moment at the current pixel and all moments at the preceding pixel, scanning the image to the right and Means for determining the moment for a row of the input image or moment image by combining pixel center moments from scanning in the direction and using masks to maintain separation between statistics for each object; key pixel locations And a means for extracting the moment at the moment.

According to a sixth aspect of the present invention, there is provided a high-speed moment calculating apparatus used in a digital processing system for calculating moment statistics of a plurality of objects in an image having a plurality of pixels. Means for forming the mask by replacing the set with a solid shape, using the mask to maintain a separation between the statistics of each object, and scanning down the image and up the image Determining the maximum value of the pixel center moment;
Finding the maximum value of the pixel center moment from the rightward and leftward scans of the vertical moment image, and using the rightward, downward, leftward, and upward scans to determine the maximum on each object Means for selecting a key pixel on a solid shape using the determined center of gravity by searching for a value and ranking the objects by size and truncating the smallest objects below the threshold. By combining the moments from scanning down the image and scanning up the image, and using masks to maintain the separation between statistics for each object, the moments on the input image or the sequence of moment images Means to determine
By combining the moments from scanning the image to the right and scanning the image to the left, and using masks to maintain separation between the statistics for each object, the input or moment image rows are A high-speed moment calculating device is provided, which has means for calculating moments and means for extracting moments at key pixel positions.

According to a seventh aspect of the present invention, the above-mentioned first aspect is provided.
In the high-speed moment calculating device according to any one of the sixth to sixth aspects, each processing element has a processing element (PE) for sequentially processing scan lines, and each PE scans each scanning element until a total moment is received from an adjacent PE. By delaying the processing of the line, a high-speed moment calculator is obtained, further comprising means for performing a horizontal scan on the linear processor array.

According to an eighth aspect of the present invention, the above-described first aspect is provided.
The scanning along the columns is performed along one direction of the diagonal, the scanning along the rows is performed along the opposite direction of the diagonal, and the scanning is performed. Is completed in both directions, a high-speed moment calculating device is obtained, comprising means for combining scanning results of scanning lines adjacent to each other.

The above-mentioned solid shape is, for example, a polygon or another shape.

[0022]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, embodiments of the present invention will be described with reference to the drawings.

The present invention provides an apparatus for efficiently calculating moments of a large number of objects in an image. FIG.
The input image (100) is filtered and thresholded (10) to form a window mask that is typically used to segment objects.
1). The set of pixels that match the object in the mask is 1
(10)
2), the background is set to 0; Points are determined on each solid shape to extract moment information (10
3) These are designated as key pixels (10
4). The input image (100) and the mask are scanned vertically, and if the corresponding mask pixel (bit) has a value of 1, the moment total is increased, otherwise the total is reset to 0 (105). . The mask and vertical moment image are used to calculate the horizontal moment of the pixel whose mask value is equal to one, otherwise the sum is reset to zero (106). When the value of the mask is zero, resetting the moment total to zero prevents information from propagating between objects (105, 106). Storing moment values only for horizontal lines containing key pixels,
Limit required storage (105, 106). The resulting moment image is sampled (103) to determine the moments (108) of various objects at key pixel locations.

Some terms used in the present invention will be explained with reference to FIG. Ship and plane objects (109)
The mask (110) is replaced with a solid shape, and one corner of each rectangle is designated as a key pixel (113). The mask has a value of 0 outside the solid shape (111) and a value of 1 inside the solid shape (112). Further, the key pixel row (114) includes the key pixel (11
It is in the mask image containing 3).

The greatest contribution of the present invention is to increase the efficiency of calculating multiple object moments by simplifying the calculation in three ways. First, this technique utilizes the spatial distance between objects to keep moment statistics separate, which limits the need for complex labeling routines. Second, the binary mask is
Statistics are examined for independent calculations, ignoring complex lookup tables or segmentation. Third
In addition, the data structure used matches the location of the image pixels rather than the object, which reduces the need for prior information and distributing the computation. As a result, moments are calculated faster for multiple objects, on simpler general hardware, than previously done.

A second component of the invention is the technique used to calculate the moment in a series of image scans. The obvious approach is one-way (up or down)
The calculation is performed by vertical moment scanning of (1) and horizontal scanning in one direction (right or left). This requires the use of the mask window of FIG. 1 with an outline parallel to the scan lines to guarantee the information transmitted to the key pixels. We solve this problem by implementing the calculation of the moment using a pair of scans in opposite directions (ie, 105 in FIG. 1).
And 106). Information is conveyed towards the center of the object that allows moment information to be extracted at the center of gravity. This allows the use of irregularly shaped masks to implement windows for each object.

The new scanning approach is implemented in FIG. 1 by replacing the scanning modules (105, 106). As shown in FIG. 3, the input image (200) and the mask (202) are scanned downward,
The moment total is increased if the mask value is 1, otherwise the total is reset to 0 (203). This is repeated in the upward scan (204), and the odd moment is measured (205), if desired.
The result of the two scans is the key pixel row (10
It is added (206) to complete the vertical moment calculation for 5). The vertical moment image (206) and mask (208 ') are scanned to the right, and the moment total is increased if the mask value is 1, otherwise the total is reset to 0 (208). This is repeated 209 in the left direction, and the odd moment is measured 210 if desired. The results of the two scans are added (211) to complete the vertical moment calculation for the key pixel row (212), as was done at (106) in FIG. The position of the key pixel for sampling the moment at (103) in FIG. 1 is now defined (215) as the center of gravity, as shown in FIG.

Allowing the use of irregular object windows by introducing a pair of scans to perform a one-dimensional moment calculation has the following advantages. First, irregularly shaped masks can be easily obtained from filtered images using morphological routines. Second, irregularly shaped object masks can be more tightly packed without duplication of object moment information. As a result, this will increase the generality of the application of the algorithm as well as the execution speed.

The third component of the present invention uses the coordinates to place the pixel at the center, retains the value at the pixel position, and immediately calculates the moment. This corresponds to the moment calculation module (105, 106) in FIG. The central moment for each point on the line area is the pixel n
Is obtained as follows by standardizing the scalar quantity in the equation (2).

[0030]

(Equation 5) Here, 1 and N indicate the first and last pixels, and m is
Moment order. The moment is also decomposed into two data scans and correction elements as follows.

[0031]

(Equation 6) here,

[0032]

(Equation 7) Correction ^{factor, (r-n) m f} [n] is, because non-zero only when m = 0, is applied only to the 0-order. The higher order moment is simply determined by the sum of two identical scans of equation (7), not in opposite directions.

The pixel center moment is obtained by replacing the scanning module (203, 204, 208, 209) and the measurement module (205, 210) in FIG. The scanning module is substituted by the pixel center moment algorithm, shown in FIG. The input image (300) and mask (301) pixels are retrieved from memory (302). If the mask value of the pixel is 0 (303), the pixel counter n and the moment Mm
Are both reset to 0. Otherwise, the pixel counter n is incremented and the moment total is
It is calculated using equation (7) (305). To get the pixel center moment of equation (6), the second scan must be scaled by (-1) ^m which produces only odd scans. This corresponds to the measurement module (2
05, 210) by replacing the module (309) of FIG.

A major contribution of introducing the pixel center moment calculation is that it is possible to calculate the center moment for many objects in parallel. The center moment for various objects can be easily extracted from the pixel center by sampling the centroid of each pixel (215), as shown in FIG. This eliminates the need to segment the object and calculate the moment using different sets of coordinates, which would increase the computational complexity of the system.

The present invention facilitates an efficient technique for calculating the pixel center moment at each pixel on the object.
A direct calculation using equation (7) to find a moment greater than zero for all pixels in a line region of length N requires multiplication and addition of m (N-1) ^2. . This is considerably more expensive than the standard moment equation multiplication of mN and addition of N. The higher central moment for a given pixel is solved by the already calculated moment. This was done by decomposing the moment as described below.

[0036]

(Equation 8) When deformed,

[0037]

(Equation 9) FIG. 7 shows an implementation for fast calculation of the pixel center moment. The calculation module (400) receives the moment value from the data scan (401) of the previous pixel and outputs the current moment value (402). Each moment is determined as a weighted combination of a previous moment value or a lower order current moment value (403).
It must be noted that the weight structure is different for each moment order. The feedback connection (404) indicates that the moment value for each iteration will be used in the next iteration until the data scan is complete.

This fast moment algorithm contributes beneficially to the overall invention in two ways. First, it allows calculating pixel center moments to be performed more efficiently than the standard moment equations. Second, it can eliminate multiplication in the moment calculation (fixed weights), thus simplifying the required hardware. Both of these aspects make performance superior to conventional implementations.

The present invention introduces approximation techniques to calculate the centroid (COM) of many objects in an image. This is achieved by using a zero or first order pixel-centered image scan. FIG. 8 shows the center-of-gravity module (21) of FIG. 4 corresponding to the key pixel module (103) of FIG.
The details of 5) will be described. As implemented earlier, the input image (500) is scanned down (502) and up (503) using the input mask (501) to keep the statistics of the object separate. You. The point with the maximum is calculated from the two data scans (502, 503). The resulting image (504) produces pixels by pixel measurement at a vertical distance from the boundaries of each object. As input, right (505) and left (506) horizontal scans are used. The next maximum operation (507) will produce a weighted distance measurement from the boundary in all directions. The position of the maximum value on each object is approximately the center of gravity of the object. Since this is the only interesting point, the maximum value of the object is propagated to adjacent mask pixels in four directions (right, bottom, left, top)
The pixel with the minimum value is removed (508). The algorithm then ranks the remaining pixel values (509) representing the centroid of another object. The threshold processing step
Added to remove small or erroneous objects in the image (510) and key pixel locations are output (5).
11).

The centroid approximation plays an important role in the present invention. The centroid pixel location is important for extracting moment center data (107), as seen in FIG. This algorithm identifies objects by centroid without the need for labeling or segmentation. The measurement of the center of gravity also provides a quick and simple measurement for classifying objects by size (weight). In addition, it provides a more advantageous centroid calculation than the standard centroid technique which requires a splitting operation (Equation (4)). Partitioning adds computational cost to simple hardware configurations.

The present invention uses two approaches to effectively adapt the above moment calculation approach to a linear processor array (LPA) structure. The vertical scan of the algorithm can obviously be guided in parallel on hardware where each processing element processes one row of data (203 and 204 in FIG. 3, 502 and 503 in FIG. 8). Horizontal scanning is an entirely different problem because data must be transferred between adjacent processors. This discussion is
Horizontal scanning module (208, 209 in FIG. 3, 5 in FIG. 8)
05, 506). FIG. 9 shows details of the approach to handle fast continuous horizontal scans with one processing element time delay. After the first processing element (PE) has processed pixel 1 on key line 1 (602), the sum is
Is passed to a second PE that processes pixel 2 above (60
2). Meanwhile, the first PE processes pixel 1 of key line 2. In the next iteration, the first PE processes pixel 1 of keyline 3 (604), and the second PE
Process pixel 2 of keyline 2 and add a third PE
Processes pixel 3 of key line 1 (607). This process continues until all scan lines have been propagated across all LPA elements.

In a second approach, the present invention uses diagonal scanning to perform moment calculations on the LPA. This is accomplished by passing the result of the data to a horizontally adjacent processing element so that the image is scanned down / up. FIG. 10 shows this change in the scan direction. Lower and upper scans (700, 70
1) is replaced by oblique lower and upper scans. This corresponds to the vertical scanning module (203, 204 in FIG. 3, FIG. 8).
502, 503). The right and left scans (702, 703) are replaced by inverted lower and upper diagonal scans. This is a modification of the horizontal scanning module of FIGS. One potential problem with this approach is that
Diagonal scan is to form a checkerboard pattern such that the odd and even scans produce independent sums. This is shown in FIG. 11, where the odd scan lines are black and the odd scan lines are white. This problem,
It is solved by averaging the results of the odd and even diagonal scans. FIG. 12 shows the final sum module (21
FIG. 8 shows a modified total module (704) corresponding to 1).
And a modified maximum module (705) for replacing the final maximum module.

This final component of the present invention allows the efficient implementation of central moment elementary calculations on linear processor array structures (LPAs). This could speed up moment calculations without resorting to specialized hardware that cannot perform other high-level operations. The first technique has a cost corresponding to a function of the sum of the row size, the column size, and the number of rows. In contrast, the computation time of the second approach is proportional to the sum of the row size and the column size. Both of these are better than the costly standard implementation which corresponds to a function of the product of the product of the row size and the number of rows and the column size.

Next, a method for calculating moments of a large number of objects in an image will be described as a first example with reference to FIG. In FIG. 13, the input image (800)
Are filtered and thresholded to form a window mask (801). The labeling routine is
The pixels are used to group the objects into objects, and each group is replaced with the smallest rectangle containing it (802). The position of the key pixel is the lower right corner of the rectangle (803). Standard vertical moment scan, mask value 1
Is performed on the input image (800) moment for the pixels of, otherwise reset the sum to zero (80
5). The mask and vertical moment image are used to calculate the horizontal moment of the pixel with a mask value of 1; otherwise, reset the sum to 0 (806). All the calculation results are accumulated in the horizontal row including the key pixel (80
3). The resulting image is sampled at key pixel locations to know the moments of various objects.

Next, with reference to FIG. 14, a method for calculating the center moment using a mask having a blob object will be described as a second example. In FIG. 14, an input image (900) is filtered and thresholded to form a window mask (901). The binary window operation consists of the following A) to C). A) Fill holes in the object. B) Eliminate noise pixels using erosion morphology. C) Spread the image until the object has a concave appearance. The location of the key pixel is found (903) by the centroid routine detailed in FIG. The pixel center vertical moment is calculated for pixels with a mask value of 1; otherwise, the sum is zero (905). The mask and vertical moment image are used to calculate the pixel center horizontal moment of the pixel with the mask value of 1; otherwise, reset the sum to 0 (906). The resulting image is the center of gravity position (903)
Is a sample for finding moments of various objects (908).

Next, referring to FIG. 15, details for executing one-dimensional scanning for finding a pixel center moment will be described as a third example. FIG. 15 illustrates a simple case of two binary line segments of length 3 and 5 pixels (1000). Assume that the mask in this case is the same as the input image. In the downward scan, the first moment _M'1 counter n (1001) generally increases with each pixel of the line segments (100
2). Its value is reset to 0 when the calculation of two independent line segments reaches a non-object pixel. This process is repeated in an upward scan (1003) to obtain a mirror-like result (1003).
4). The pixel center first moment is shifted downward (10
02) is found (1005) by subtracting the upward scan (1004). As a result, when the center pixel is present in each line segment, it can be confirmed that the first central moment is zero.

As mentioned earlier, the present invention calculates the pixel center moment at each pixel by means of the previously calculated pixels. This will be described with reference to FIG. FIG. 16 shows some forms for calculating the moment (900). Dimension greater than zero order (1101)
, The first pixel on the object has a moment value of 0 (1102). This can be confirmed by equation (7). The remaining pixel moments are represented by the previous or current moments (1103). The cost required for processing the line segment N pixel length is added (sum)
The process is shown in the table (1104). The total number of calculations is
Comparable to a single scan standard moment calculation which requires the cost of multiplying m (N-1) and adding (N-1). However, this will translate into a performance advantage for simpler systems with higher computational costs for multiplication than for addition.

Referring now to FIG. 17, a one-dimensional scan used to find the center of gravity (COM) of a number of objects will be described as a fourth example. FIG. 12 shows a simple case of the two binary line segments (1200) described above. Again, the mask is assumed to be the same as the input image. In the downward scan (1201), 0-order moment _M'0 increases at each pixel of the line segment (1202). As before, the value is reset to zero as soon as a non-object pixel is reached. This process is repeated in the upward scan (1203) to obtain a mirror image-like result (1204). The maximum per pixel is calculated from the previous two scans (12
05). If the resulting peak value is found in the one-dimensional case, the maximum is at the center of each line segment (1205). It must be noted that this median is larger for longer segments, so that its quantity can be used for ranking objects by size (1205).

Next, a general calculation method for calculating the pixel center moment will be described with reference to FIG. FIG.
In, an input mask including four blob objects (1301) is shown (1300). The black blob area (1301) represents the position and approximate shape of all objects in one image. In the first moment pass, the input image (not shown) and each row of masks are scanned 1302 downward and upward. This is represented by a vertical line with several parallel, two-way arrows (130
3). In the next moment pass, the resulting image is scanned horizontally to the right and left as indicated by the horizontal line (1305) with parallel double arrows (1).
304). In this case, the scan is the center of gravity of the object (COM)
Is executed on the line containing Since the present invention only extracts the center moment information at the position (107) of the COM as described in FIG. 1, another horizontal scanning line is unnecessary.

Next, referring to FIG. 19 and FIG. 20, the calculated moment is converted to a linear processor array (LPA).
Two calculation methods for adapting to hardware will be described as a final example. FIG. 19 shows a first technique for performing on an input mask containing four objects, as described above. In the first moment pass, the input image (not shown) and each row of masks are scanned downward and upward (1400). These scans can be performed in parallel by the LPA (1401). In the next moment pass (1402), the resulting image is scanned horizontally to the right and left. This scan is indicated by an irregular horizontal line (indicating one processing element processing delay between scan lines) (140
3). FIG. 20 shows the oblique scan line technique for the same mask image. Scanning direction used (1405, 1407)
Make 45 ° to the image mask and form an angle of 90 ° with each other. LP for both first and last scan
A can be completely executed in parallel on A (1404, 140
6). This results in a fast execution of the moment algorithm.

The present invention has thus proposed an efficient method for calculating the moments of a large number of objects in an image. That is, the input image is typically filtered and thresholded to form a window mask used to segment the object. The collection of pixels corresponding to the object in the mask is replaced with a solid shape. Key pixel positions on each object are determined from the mask and the input image. The mask and the input image are scanned vertically and the moment sum is increased if the mask pixel (bit) has a value of one, otherwise the sum is reset to zero. The mask and vertical moment images are used for the calculation of the horizontal moment of a pixel with a mask value of 1, otherwise the sum is reset to zero. Resetting the moment total to 0 when the mask value is 0 prevents the propagation of information between objects. Storing moment values only for horizontal lines with key pixels reduces storage requirements. The resulting moment image is sampled to find the moments of various objects at the key pixel locations.

The greatest contribution of the present invention is that the calculation efficiency of a large number of object moments is improved by simplifying the calculation by three methods. First, this technology
It exploits the spatial distance between objects to reduce the need for complex labeling routines and keep moment statistics separate. Second, the binary mask can be used without complex lookup tables or segmentation,
The statistics are consulted to calculate independently. Third, the data structure used corresponds to image pixel locations rather than objects, which reduces the need for previous information,
Distribute the calculations. As a result, moments can be calculated for a large number of objects with simpler general hardware and faster than previously done.

[0053]

As described above, according to the present invention, there is an effect that moment statistics of an object can be calculated at high speed with simple hardware.

[Brief description of the drawings]

FIG. 1 illustrates a method according to the invention for calculating moment statistics of a large number of objects in a single image.

FIG. 2 is a diagram used to explain the present invention and a conventional method.

FIG. 3 illustrates a method of the invention for decomposing a scan of moments so that the center of gravity (COM) of each object can use object statistics for the sample.

FIG. 4 is a diagram illustrating a key pixel sampling scheme for a moment scan decomposed in FIG. 3;

FIG. 5 is a diagram for explaining a method of the present invention for calculating a pixel center moment for each pixel in a single scan.

FIG. 6 is a diagram for explaining a scan factor applied to an odd moment when the scan is divided into two scans in FIG. 5;

FIG. 7 illustrates a fast algorithm according to the present invention for calculating a pixel center moment with reference to a previously calculated pixel center moment.

FIG. 8 is a diagram for explaining a technique according to the present invention for simultaneously calculating COMs (centroids) of many objects.

FIG. 9 illustrates an approach according to the invention for adapting horizontal moment scanning to linear processor array (LPA) hardware.

FIG. 10 illustrates an approach according to the invention for adapting a moment scan to LPA hardware by changing the scan direction.

FIG. 11 is a diagram for explaining a checkerboard pattern obtained from diagonal scanning according to the present invention.

FIG. 12 illustrates a method of the present invention for correcting the effect of a checkerboard pattern on diagonal scanning.

FIG. 13 illustrates an example according to the method of the present invention for calculating moment statistics of multiple objects in a single image.

FIG. 14 is a diagram illustrating an example according to the method of the present invention for calculating the central moment of a large number of objects in a single image.

FIG. 15 is a diagram for explaining a detailed example of obtaining a first moment using two scans in opposite directions.

FIG. 16 is a diagram illustrating a detailed cost analysis when calculating a pixel center moment with reference to a previously calculated pixel center moment according to the present invention.

FIG. 17 is a diagram for explaining a detailed example when performing COM (centroid) calculation using one-dimensional data scanning according to the present invention.

FIG. 18 is a diagram illustrating a method of the present invention for calculating a pixel center moment on standard hardware.

FIG. 19 illustrates a first approach for adapting pixel center moment calculations to LPA hardware according to the present invention.

FIG. 20 illustrates a second approach for adapting pixel center moment calculations to LPA hardware according to the present invention.

[Explanation of symbols]

 109 Object (ship and airplane) 110 Mask 113 Key pixel

Claims (8)

[Claims] 1. A high-speed moment calculator for calculating moment statistics of a plurality of objects in an image having a plurality of pixels, which is used in a digital processing system, wherein a mask is obtained by replacing a set of object pixels with a solid shape. Means for selecting key pixels on a solid shape for each object using the mask and the input image; and a mask for maintaining separation between statistics for each object. Means for calculating and accumulating moments along columns of the input image or moment image, and using a mask to maintain a separation between statistics for each object, using a mask for the rows of the input image or moment image. A high-speed moment calculating apparatus, comprising: means for calculating and accumulating moments by means of a key; and means for extracting moments at key pixel positions. 2. A high-speed moment calculating apparatus for calculating moment statistics of a plurality of objects in an image having a plurality of pixels, used in a digital processing system, wherein a mask is formed by replacing a set of object pixels with a solid shape. Means for selecting a key pixel on a solid shape for each object using a mask and an input image; and means for scanning the image downward and the image upward. Means for determining the moment for the input image or sequence of moment images by combining the moments and using masks to maintain a separation between the statistics of each object; scanning the image to the right and Combining moments from leftward scans and using masks to maintain separation between statistics for each object , Fast moment calculation apparatus characterized by comprising means for determining the moment for the row of the input image or moment images, and means for extracting a moment at the key pixel position. 3. A high-speed moment calculating apparatus used in a digital processing system for calculating moment statistics of a plurality of objects in an image having a plurality of pixels, wherein a mask is obtained by replacing a set of object pixels with a solid shape. Means for selecting a key pixel on a solid shape for each object using a mask and an input image; and means for scanning the image downward and the image upward. By combining pixel center moments and using masks to maintain separation between statistics for each object,
Means for determining moments for a sequence of input images or moment images; combining pixel center moments from scanning the image to the right and scanning the image to the left; and maintaining a separation between statistics for each object. A high-speed moment calculating apparatus, comprising: means for obtaining a moment for a row of an input image or a moment image by using a mask, and means for extracting a moment at a key pixel position. 4. A high-speed moment calculating apparatus used in a digital processing system for calculating moment statistics of a plurality of objects in an image having a plurality of pixels, wherein a mask is obtained by replacing a set of object pixels with a solid shape. Means for selecting a key pixel on a solid shape for each object using a mask and an input image; and means for selecting a lower-order moment at the current pixel and a preceding pixel. As a linear combination of all moments, finding the pixel center moment at each pixel in the scan, combining the pixel center moment from scanning down the image and scanning up the image, The moment for the input image or sequence of moment images by using a mask to maintain the separation between the statistics of Determining the pixel center moment at each pixel of the scan as a linear combination of the lower order moment at the current pixel and all moments at the preceding pixel, and scanning the image to the right. Means for determining the moment for a row of the input image or moment image by combining pixel center moments from scanning the image to the left and using a mask to maintain a separation between statistics for each object. Means for extracting a moment at a key pixel position. 5. A high-speed moment calculating device for calculating moment statistics of a plurality of objects in an image having a plurality of pixels, which is used in a digital processing system, wherein a mask is formed by replacing a set of object pixels with a solid shape. Means for forming the image, using a mask to maintain the separation between statistics for each object, and determining the maximum of the pixel center moment from scanning down the image and scanning up the image. And finding the maximum value of the pixel center moment from the rightward scanning and leftward scanning of the vertical moment image;
By looking for the maximum value on each object using scans down, left and up, and ranking the objects by size and truncating the smallest objects below the threshold Means for selecting key pixels on the solid shape using the determined center of gravity, and each pixel of the scan as a linear combination of the lower order moment at the current pixel and all moments at the preceding pixel. Determining the pixel center moment at, combining the pixel center moment from scanning down the image and scanning up the image, and using a mask to maintain the separation between statistics for each object. Means to determine the moment for the input image or column of the moment image, and the line between the lower order moment at the current pixel and all moments at the preceding pixel As a combination of determining the pixel center moment at each pixel of the scan, combining the pixel center moment from scanning the image to the right and scanning the image to the left, and separating between statistics for each object. A high-speed moment calculating apparatus comprising: means for obtaining a moment for a row of an input image or a moment image by using a mask to maintain the moment; and means for extracting a moment at a key pixel position. 6. A high-speed moment calculating apparatus used in a digital processing system for calculating moment statistics of a plurality of objects in an image having a plurality of pixels, wherein a mask is obtained by replacing a set of object pixels with a solid shape. Means for forming the image, using a mask to maintain the separation between statistics for each object, and determining the maximum of the pixel center moment from scanning down the image and scanning up the image. And finding the maximum value of the pixel center moment from the rightward scanning and leftward scanning of the vertical moment image;
By looking for the maximum value on each object using scans down, left and up, and ranking the objects by size and truncating the smallest objects below the threshold Means for selecting key pixels on a solid shape using the determined center of gravity, combining moments from scanning down the image and scanning up the image, and separating the statistics between each object. Means for determining moments for the input image or sequence of moment images by using a mask to maintain; combining the moments from scanning the image to the right and scanning the image to the left; Means for determining a moment for a row of the input image or moment image by using a mask to maintain a separation between the statistics of the key pixels; Fast moment calculation apparatus characterized by having a means for extracting moment at the location. 7. The high-speed moment calculating device according to claim 1, wherein each processing element has a processing element (PE) for sequentially processing a scan line, and receives a moment total from an adjacent PE. A high-speed moment calculating apparatus further comprising means for performing horizontal scanning on a linear processor array by delaying processing of each scan line in each PE. 8. The high-speed moment calculating apparatus according to claim 1, wherein scanning along a column is performed along one direction of a diagonal, and scanning along a row is performed along a direction opposite to the diagonal. A high-speed moment calculating device comprising means for combining the scan results of adjacent scan lines after the scan is completed in both directions.