Particle size distribution estimation device and program

Abstract

To measure (estimate) the granularity distribution of a material easily and accurately.SOLUTION: A granularity distribution estimation device comprises: input means (S11) which inputs a photographed image of a material; detection means (S17, S19) which detects a minor axis equivalent part and a major axis equivalent part of each aggregate from the photographed image, and calculates a contour feature amount including at least the length of the minor axis equivalent part; determination means (S25, S27) which determines a granularity section becoming a first candidate on the basis of the closeness in a positional relation between the distribution of a learning contour feature amount of each granularity section obtained in advance on the basis of the learning image and a contour feature amount specified by the detection means; calculation means (S29) which calculates an aggregate area as an index value expressing an amount of each aggregate from the photographed image; and area integration means (S30) which calculates the total area for each granularity section by classifying all or a portion of the aggregate area calculated by the calculation means into the granularity section of the first candidate for each aggregate.SELECTED DRAWING: Figure 9

Description

[Technical Field]

[0001] The present invention relates to a particle size distribution estimation device and a program for estimating the particle size distribution of a material consisting of a large number of aggregates. [Background technology]

[0002] The conventional method for quality control of aggregates is to perform intermittent sampling on a conveyor or similar device and measure the particle size distribution of the material (aggregate) using a hand sieve. However, this method places a heavy burden on the worker who has to sieve by hand, and it takes several hours to calculate the particle size distribution.

[0003] Additionally, techniques for controlling the quality of aggregates using methods other than hand sieving have been proposed. In Japanese Patent No. 6243640 (Patent Document 1), the particle size classification of aggregates is estimated from photographed images of falling aggregates. In Japanese Patent Laid-Open No. 2021-67468 (Patent Document 2), the particle size distribution of soil materials is estimated by machine learning the correlation between image data of the soil materials and analysis result data of the particle size distribution of the soil materials.

[0004] Additionally, in Japanese Patent No. 6764982 (Patent Document 3), the particle size distribution of aggregates is estimated by machine learning the features and particle size distribution of three-dimensional data on the surface of a group of aggregates used for training. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Patent No. 6243640 [Patent Document 2] Patent Publication No. 2021-67468 [Patent Document 3] Patent No. 6764982 Summary of the Invention [Problem to be solved by the invention]

[0006] In the methods of Patent Documents 1 and 2, although it is possible to estimate the gradation class of each aggregate using photographed images of the material, the shape of the aggregate in the depth direction is unknown, and therefore it may not be possible to accurately estimate the gradation class of aggregates that have irregular shapes, such as flat shapes. In particular, when photographing aggregates that are freely falling from a conveyor or other transport unit, it is possible that the size of the aggregate cannot be accurately determined because the position of the aggregate is not constant.

[0007] Furthermore, Patent Document 3 has the problem that the equipment costs are high because three-dimensional data is measured, and the equipment is also complicated, which increases the difficulty of the work and the number of steps, which increases the burden on workers.

[0008] The present invention has been made to solve the above-mentioned problems, and an object of the present invention is to provide a particle size distribution estimation device and a program that can measure (estimate) the particle size distribution of a material simply and accurately. [Means for solving the problem]

[0009] A particle size distribution estimation device according to one aspect of the present invention is a particle size distribution estimation device for estimating the particle size distribution of a material consisting of a large number of aggregates, and includes: input means for inputting a photographed image of the material; detection means for detecting a portion corresponding to a minor axis and a portion corresponding to a major axis of each aggregate from the photographed image and calculating a contour feature including at least the length of the portion corresponding to the minor axis; storage means for storing distribution information indicating the distribution of training contour features for each particle size division obtained in advance based on the training image; determination means for determining a first candidate particle size division for each aggregate based on the proximity of the positional relationship between the distribution of training contour features for each particle size division indicated by the distribution information and the contour feature identified by the detection means; calculation means for calculating an aggregate area as an index value representing the amount of each aggregate from the photographed image; and area accumulation means for calculating a total area for each particle size division by classifying all or part of the aggregate area calculated by the calculation means into the first candidate particle size division for each aggregate.

[0010] Preferably, the detecting means detects an abstracted figure obtained by abstracting the contour shape of each aggregate from the photographed image. In this case, it is desirable that the part corresponding to the minor axis is the minor axis or a short side of the abstracted figure, and the part corresponding to the major axis is the major axis or a long side of the abstracted figure.

[0011] Typically, the abstract shape is an approximate ellipse.

[0012] Preferably, the contour feature amount includes a length of a part corresponding to a minor axis and a length of a part corresponding to a major axis, and the distribution is a two-dimensional Gaussian distribution.

[0013] Preferably, the determination means determines, for each aggregate, a plurality of candidates including a first candidate based on the distance between the distribution of learning contour features for each particle size classification indicated by the distribution information and the contour features, and the area accumulation means sets a weighting coefficient based on the distance and classifies the aggregate area into each of the plurality of candidates.

[0014] It is also desirable that the determination means designates the classification one level above or below the first candidate as the second candidate, and that the area accumulation means set a weighting coefficient normalized according to the distance from the first candidate and the second candidate, and classify the aggregate area into the first candidate and the second candidate particle size classification, respectively.

[0015] A particle size distribution estimation program according to another aspect of the present invention causes a computer to execute the following steps: inputting a photographed image of a material consisting of a large number of aggregates; detecting from the photographed image the portions corresponding to the short diameter and the long diameter of each aggregate and calculating a contour feature including at least the length of the portion corresponding to the short diameter; determining, for each aggregate, a first candidate particle size class based on the proximity of the positional relationship between the distribution of training contour features of each particle size class determined in advance based on the training image and the contour feature; calculating, from the photographed image, the aggregate area as an index value representing the amount of each aggregate; and calculating, for each aggregate, the total area for each particle size class by classifying all or part of the aggregate area into the first candidate particle size class. [Effects of the Invention]

[0016] According to the present invention, it is possible to estimate the particle size distribution of a material with a simple configuration and with higher accuracy than conventional methods. [Brief explanation of the drawings]

[0017] [Figure 1] 1 is a schematic diagram illustrating an overview of a particle size distribution estimation system according to an embodiment of the present invention. [Figure 2] (A) is a schematic diagram for explaining a comparative example of a particle size distribution estimation system according to an embodiment of the present invention, and (B) and (C) are schematic diagrams showing the posture of aggregates freely falling from a conveying section. [Figure 3] FIG. 10 is a diagram showing a method for measuring the size of aggregates in an experiment based on a comparative example of the particle-size distribution estimation system according to the embodiment of the present invention. [Figure 4] FIG. 1A is a functional block diagram showing the functional configuration of a learning device according to an embodiment of the present invention, and FIG. 1B is a functional block diagram showing the functional configuration of a particle size distribution estimation device according to an embodiment of the present invention. [Figure 5] 1A and 1B are diagrams for explaining a method for estimating a gradation classification in an embodiment of the present invention, in which (A) shows an ellipse approximating an aggregate, and (B) shows a Gaussian distribution for each gradation classification. [Figure 6] 1 is a diagram for explaining a method for estimating particle size classification in an embodiment of the present invention. FIG. [Figure 7] 1(A) and 1(B) are diagrams schematically showing the results of estimating particle size distribution. [Figure 8] 10 is a flowchart showing a distribution information generation process according to an embodiment of the present invention. [Figure 9] 1 is a flowchart showing a particle size distribution estimation process according to an embodiment of the present invention. [Figure 10] 1A to 1C are diagrams illustrating an example of an image processing method according to an embodiment of the present invention. [Figure 11] 10 is a graph showing an evaluation result when a two-dimensional Gaussian distribution is used in an embodiment of the present invention. [Figure 12] 10 is a table showing evaluation results when a two-dimensional Gaussian distribution is used in an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0018] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described in detail with reference to the drawings. In the drawings, the same or corresponding parts are designated by the same reference numerals and description thereof will not be repeated.

[0019] (About the overview) First, an overview of a particle size distribution estimation system 1 according to the present embodiment will be described with reference to FIGS.

[0020] As shown in Figure 1, the particle size distribution estimation system 1 includes a photographing mechanism 2 for photographing a material 9 consisting of a large number of aggregates as it flows down, and a particle size distribution estimation device (hereinafter abbreviated as "estimation device") 4 for estimating the particle size distribution of the material 9 by image processing the image data photographed by the photographing mechanism 2. The "particle size distribution" is expressed as the cumulative passing rate, where the mesh size of the sieve used in the sieving test (each opening side of a flat-woven sieve) is used as a particle size division, and the proportion of aggregates that pass through the sieve of each particle size division is used.

[0021] The photographing mechanism 2 includes a conveying section 21 that conveys the material 9 forward and allows the material 9 to fall freely from the front end, a screen 22 provided on the rear side of the path along which the material 9 flows, an illumination section 23 that illuminates the screen 22 from the front, and a camera (photographing means) 24 that is arranged facing the screen 22 and photographs the material 9 flowing down the path. The conveying section 21 is typically a belt conveyor. The camera 24 continuously photographs the material 9 falling freely from the conveying section 21, i.e., the aggregates falling freely. The material 9 may also be allowed to fall freely from a drop means other than the conveying section 21. Other possible drop means include allowing the material 9 to fall freely directly from the outlet of a crusher, classifier (sieving machine), etc.

[0022] The estimation device 4 estimates the particle size distribution of each aggregate by processing the image captured by the camera 24. This makes it possible to estimate the particle size distribution of the material 9 without performing a manual sieve test.

[0023] Here, ideally, as shown in Figure 2(A), the smallest circumscribing square R1 of the aggregate region 91 can be considered as a sieve mesh, and the grain size classification of the aggregate (the original aggregate detected as the aggregate region 91) can be appropriately estimated by determining to which grain size classification the side length L10 of the smallest circumscribing square R1 belongs.

[0024] However, flat aggregate 92 as shown by hatching in Figures 2(B) and (C) is considered to assume a stable posture on conveying section 21 when the long side and the middle side face upward. Figure 2(B) is a top view of aggregate 92 on conveying section 21, and Figure 2(C) is a side view of aggregate 92 on conveying section 21. Note that in these figures, the flat shape of aggregate 92 is exaggerated.

[0025] Therefore, flat aggregates 92 tend to fall with their long side and middle side facing forward (toward camera 24), as shown in Fig. 2(C). Some aggregates fall while rotating, but if the falling aggregate is imaged from the front, the outline shape of the long side and middle side can be easily recognized.

[0026] On the other hand, in manual sieving tests, whether or not an object passes through the sieve mesh depends on the size of the sieve mesh multiplied by the mid-side. This is because even if an object does not pass through a mesh that is long x mid-side, it may pass through a mesh that is long x mid-side. Therefore, when estimating the aggregate particle size distribution using the minimum circumscribing square R1, the estimated particle size distribution tends to be larger than the actual particle size obtained in the sieve test. Furthermore, this tendency is thought to be more pronounced for aggregates with a significant difference between the long and short sides (flat aggregates).

[0027] As an example, we measured the three sides (short side, middle side, long side) of aggregate in the 30-20mm particle size range with vernier calipers, and calculated the side length of the square circumscribing the ellipse formed by the middle side x short side or the long side x middle side (see Figure 3), and conducted an experiment to examine the effect on the accuracy of estimating particle size range due to differences in the direction in which the image was taken.The results of this experiment showed that when the middle side x short side was photographed from the front, the side length of the square was concentrated on the actual sieve opening (30-20mm), whereas when the long side x middle side was photographed from the front, the side length of the square was greater than the actual sieve opening (30-20mm).

[0028] Therefore, when estimating the gradation classification of an aggregate by extracting the minimum circumscribing square R1 of the aggregate region 91 as shown in Figure 2(A), there is a possibility that the aggregate will be classified into a gradation classification larger than the actual classification. Since the larger the size (particle size) of the aggregate, the larger the difference between the long side and the short side tends to be, the larger the gradation classification, the larger the error may become. If the middle side x short side of the aggregate could be photographed by the camera 24, such errors could be reduced, but this is difficult to achieve in practice.

[0029] Therefore, the estimation device 4 in this embodiment determines at least one gradation class as a candidate using a statistical method based on the "contour feature value" representing the contour shape of the aggregate, and estimates the gradation distribution of the material 9 by classifying the "aggregate area" into the gradation class determined as a candidate. This makes it possible to correct the above-mentioned errors.

[0030] The "contour feature amount" in this embodiment includes the minor axis length and major axis length of an approximate ellipse of the aggregate region 91 extracted from the captured image (see FIG. 5(A) described later). The "approximate ellipse" is a typical example of an abstracted figure that abstracts the contour shape of the aggregate region 91. The "abstracted figure" is a figure in which the part corresponding to the minor axis and the part corresponding to the major axis of the aggregate (aggregate region 91) can be expressed by two straight lines (preferably perpendicular to each other). Note that the parts corresponding to the minor axis and the major axis do not depend on the actual short side, middle side, and long side of the aggregate, but represent the short side and long side parts that can be identified from the aggregate region 91.

[0031] The "aggregate area" is an index value representing the amount of each aggregate, and in this embodiment, the projected area of ​​the aggregate, i.e., the area of ​​the aggregate region 91, is calculated as the aggregate area. Note that the contour feature amount and aggregate area can be calculated by converting the number of pixels (pixel) of the captured image into an actual value (mm).

[0032] 1, the estimation system 1 according to this embodiment further includes a learning device 3 so that candidates for particle size classification can be determined by a statistical method. The learning device 3 and the estimation device 4 are configured by an information processing device (such as a general-purpose computer) equipped with a processor such as a CPU (Central Processing Unit) and a non-volatile memory.

[0033] The learning device 3 calculates in advance the distribution of the contour feature values ​​of the aggregates (hereinafter referred to as "learning contour feature values") for each gradation class based on learning images of the aggregates for each gradation class. As a result, distribution information indicating the distribution of the learning contour feature values ​​for each gradation class is generated, and the generated distribution information is stored in the storage device of the estimation device 4.

[0034] The functional configurations and operations of the learning device 3 and the estimation device 4 in this embodiment will be described in detail below.

[0035] In this embodiment, the particle size classification is the smallest classification of less than 5 mm, and six classifications of (1) 5-13 mm, (2) 13-20 mm, (3) 20-25 mm, (4) 25-30 mm, (5) 30-40 mm, and (6) 40-50 mm. The method of setting the particle size classification is not particularly limited.

[0036] (About the learning device) FIG. 4A is a functional block diagram showing the functional configuration of the learning device 3.

[0037] The learning device 3 includes an input unit 31 that inputs (acquires) multiple training image data obtained by continuously photographing a group of aggregates (samples) 9A (Figure 1) divided into particle size classifications, a training image memory unit 32 that stores the multiple acquired training image data for each particle size classification, a contour detection unit 33 that performs image processing on the training image data stored in the training image memory unit 32 to detect an approximate ellipse of each aggregate and calculates training contour features based on the detected approximate ellipses, a calculation unit 34 that calculates the distribution of the training contour features for each particle size classification, and a distribution information memory unit 35 that stores the calculation results by the calculation unit 34 as distribution information D1.

[0038] The learning image storage unit 32 includes, for example, learning files equal in number to the number of granularity divisions, and each learning file stores learning image data of the corresponding granularity division.

[0039] The contour detection unit 33 performs image processing using, for example, deep learning to extract (detect) aggregate regions from the learning image, and calculates an approximation ellipse R2 of the aggregate region 91 as an "abstracted figure" that abstracts the contour shape of the aggregate region 91, as shown in Fig. 5(A). The contour detection unit 33 also calculates the major axis length L1 and the minor axis length L2 of the approximation ellipse R2.

[0040] The calculation unit 34 calculates a two-dimensional Gaussian distribution for each particle size classification using the major axis length L1 and minor axis length L2 of the approximation ellipse R2 as variables. The major axis length L1 and minor axis length L2 of the approximation ellipse R2 used to calculate the Gaussian distribution are "learning contour features."

[0041] 5(B) is an example of a graph in which the two-dimensional Gaussian distribution of the training contour feature values ​​for each granularity class is represented by a probability ellipse. Here, Gaussian distributions G1 to G6 corresponding to the above-mentioned granularity classes (1) to (6) respectively are shown. Information on such Gaussian distributions G1 to G6 is stored as distribution information D1 in the distribution information storage unit 35. The information on the two-dimensional Gaussian distribution stored as distribution information D1 may be information that directly represents the probability distribution, or may be information on a normal distribution obtained by normalizing the probability distribution.

[0042] The functions of the contour detection unit 33 and the calculation unit 34 are realized by the processor executing software. The learning image storage unit 32 and the distribution information storage unit 35 are typically configured with non-volatile memory. The input unit 31 is, for example, an interface that inputs image data from the camera 24 via a wired or wireless connection.

[0043] (Functional configuration of the estimation device) FIG. 4B is a functional block diagram showing the functional configuration of the estimation device 4.

[0044] The estimation device 4 includes a distribution information storage unit 46 that stores the distribution information D2 (D1) generated by the learning device 3, an input unit 41 that inputs (acquires) multiple image data from the camera 24, i.e., multiple image data obtained by continuously photographing the material 9 freely falling from the conveying unit 21, an image storage unit 42 that stores the multiple acquired image data, a contour detection unit 43 that processes the image data stored in the image storage unit 42 to detect an approximate ellipse of each aggregate and calculates contour features based on the detected approximate ellipse, an estimation processing unit 44 that determines at least a first candidate grain size classification for each aggregate based on the contour features and classifies the aggregate area based on the determination result to estimate the grain size distribution of the material 9, and an output unit 45 that outputs the estimation result by the estimation processing unit 44. The distribution information D2 can be received from the learning device 3, for example, via a communication interface (not shown).

[0045] Similar to the contour detection unit 33 of the learning device 3, the contour detection unit 43 extracts (detects) aggregate regions from the captured image by performing image processing using, for example, deep learning, and calculates (detects) an approximate ellipse R2 for each aggregate region 91, as shown in FIG. 5(A). In addition, the contour detection unit 43 calculates the major axis length L1 and the minor axis length L2 of the approximate ellipse R2 as contour feature amounts of the aggregate.

[0046] The estimation processing unit 44 determines candidates for granularity classification by evaluating the proximity of the positional relationship between the calculated contour feature and the Gaussian distributions G1 to G6 of the learning contour feature for each granularity classification indicated by the distribution information D2. Specifically, the estimation processing unit 44 determines the first candidate granularity classification based on the proximity of the distance between the calculated contour feature and the Gaussian distributions G1 to G6. Then, the estimation processing unit 44 classifies all or part of the aggregate area into the first candidate granularity classification for each aggregate, thereby calculating the total area for each granularity classification.

[0047] The proximity of the positional relationship to Gaussian distributions G1 to G6 can typically be evaluated by the "Mahalanobis distance," which quantifies the degree of separation from a data group. Specifically, as shown in Figure 6, the Mahalanobis distance between point P, whose X and Y coordinates are the long axis length L1 and short axis length L2 of a certain aggregate A, and each of Gaussian distributions G1 to G6 (population) is calculated. For example, if point P is closest to Gaussian distribution G3, the above-mentioned gradation classification (3) can be determined as the first candidate. Furthermore, the second candidate, third candidate, ... can be determined in order of decreasing distance.

[0048] After determining the candidates for the gradation classification of aggregate A as described above, the estimation processing unit 44 weights and classifies the aggregate area (area of ​​aggregate region 91) according to the Mahalanobis distance. The following method (classification method 1) can be used as a method for classifying the aggregate area of ​​aggregate A after the candidate determination.

[0049] Classification method 1: The weighting coefficients of the grain size classifications of multiple candidates, including the first candidate, are calculated by normalizing them so that the sum is 1 according to the distance, and the area of ​​the aggregate region 91 of aggregate A is weighted for each of the grain size classifications of the multiple candidates and classified.

[0050] When the above-described processing is completed for all the captured images, the estimation processing unit 44 calculates the total area (mm 2), the particle size distribution of the material 9 can be calculated (estimated) by calculating the area frequency (%) for each particle size division. In this way, the estimation processing unit 44 functions as a determination means for determining candidates for particle size division, and as an area integration means for calculating the total area for each particle size division.

[0051] The output unit 45 outputs the particle size distribution of the material 9 estimated by the estimation processing unit 44. The output unit 45 is configured, for example, by a display unit including a display, and displays the area and area frequency for each particle size classification in a table format as shown in FIG. 7(A). Alternatively, a particle size distribution curve as shown in FIG. 7(B) may be generated and displayed. The particle size distribution curve is a curve obtained by plotting the cumulative values ​​of area frequency in ascending order of particle size classification on a graph with particle size (mm) on the horizontal axis and cumulative passing rate (%) on the vertical axis.

[0052] The functions of the contour detection unit 43 and estimation processing unit 44 shown in Fig. 4(B) are realized by a processor executing software. The distribution information storage unit 46 is typically configured with a non-volatile memory, and the image storage unit 42 is configured with a volatile or non-volatile memory. The input unit 41 is, for example, an interface that inputs image data from the camera 24 via a wired or wireless connection.

[0053] The distribution information D2 (D1) generated by the learning device 3 may be stored on a cloud server (not shown), and in such a case, the estimation device 4 may download the distribution information D2 from the cloud server and store it in the distribution information storage unit 46 when estimating the particle size distribution of the material 9.

[0054] The output unit 45 may also be configured by a communication interface for transmitting and receiving data to and from other computers, or may be configured by a writing device for writing data to a removable recording medium.

[0055] (Distribution information generation method) The distribution information generating method executed by the learning device 3 will be described mainly with reference to Figure 4(A) and Figure 8. Figure 8 is a flowchart showing the distribution information generating process.

[0056] 8, first, aggregate (sample 9A) is fed into conveying section 21 for each gradation classification, and the aggregate (hereinafter also referred to as "learning aggregate") freely flowing down from the end of conveying section 21 is continuously photographed by camera 24 (step S1). As a result, a plurality of image data (learning image data) for each gradation classification is input to input section 31 in chronological order.

[0057] The transport speed of the sample 9A by the transport unit 21 may be the same as that of a typical material transport conveyor, for example, around 1 m / s. The falling speed of the sample 9A as it free-falls from the front end of the transport unit 21 increases at an accelerating rate. If the installation height of the transport unit 21 relative to the installation height of the camera 24 is approximately 1 m, the falling speed of the sample 9A flowing down the space in front of the camera 24 (the flow path) will be approximately 4.4 m / s. When the falling speed increases to 4.4 m / s, the dense aggregate group that had been flowing at 1 m / s up until then spreads out, eliminating or reducing overlap between the aggregates. Therefore, according to this embodiment, the camera 24 can capture the spread aggregate group. The camera 24 continuously captures the falling sample 9A so that each aggregate is photographed at least once.

[0058] The plurality (large number) of image data input to the input unit 31 are recorded in chronological order as learning image data in the files of the corresponding granularity classification in the learning image storage unit 32.

[0059] Once image data is acquired for each granularity classification, the contour detection unit 33 performs image processing on each piece of image data to detect (extract) the region of each aggregate, i.e., aggregate region 91, for each piece of image data (step S3). In this embodiment, the shape of each aggregate is determined by image processing using instance segmentation, and the region of each aggregate is extracted according to the determined shape. Instance segmentation is a type of image processing technology using deep learning, and is a technology that detects the position of each object and classifies the pixels that make up the object.

[0060] By using instance segmentation in image processing, shape information can be output while distinguishing each aggregate. Furthermore, since objects belonging to the same class can be classified as separate objects, it is possible to determine the boundaries of areas where aggregates overlap. Note that although instance segmentation is used in this embodiment, other types of image processing techniques may also be used as long as they can detect each individual aggregate.

[0061] Next, the contour detection unit 33 detects, for each granularity classification, the approximate ellipse R2 of each aggregate region 91 detected in step S3 as the "abstract figure" of each training aggregate (step S5). The major axis length L1 and minor axis length L2 of the approximate ellipse R2 are calculated as the contour feature of each training aggregate, i.e., the "training contour feature" (step S7). Depending on the shape of the aggregate, the approximate ellipse R2 may be a perfect circle, in which case the major axis length L1 and minor axis length L2 may be the same length.

[0062] The calculation unit 34 calculates a Gaussian distribution of the training contour feature values ​​obtained for each aggregate in step S7 for each gradation class (step S9). That is, a two-dimensional Gaussian distribution is calculated using the major axis length L1 and minor axis length L2 of the approximation ellipse R2 as variables. This generates distribution information indicating the Gaussian distribution for each gradation class. To reduce variation in the accuracy of the Gaussian distribution, it is desirable to calculate the Gaussian distribution using the same number of training aggregates for each gradation class.

[0063] In this embodiment, an example has been described in which learning device 3 inputs image data directly from camera 24 of imaging mechanism 2, but the present invention is not limited to this example. In other words, input unit 31 in learning device 3 may be configured as a communication interface that receives image data captured in advance from another computer, or a reading device that reads image data from a removable recording medium.

[0064] (Method for estimating particle size distribution) The particle size distribution estimation method executed by the estimation device 4 will be described with reference to the flowchart of Fig. 9. Fig. 9 is a flowchart showing the particle size distribution estimation process.

[0065] Referring to FIG. 9, first, material 9 to be measured, whose particle size distribution is unknown, is fed into conveying section 21, and camera 24 continuously captures images of material 9 flowing freely down from the end of conveying section 21 (step S11). The captured image data is temporarily stored in image storage section 42 in chronological order. The installation height and conveying speed of conveying section 21 during image capture are typically the same as those during learning, and camera 24 captures still images of the dispersed aggregate group. Camera 24 also continuously captures images of falling material 9 so that each aggregate is captured at least once.

[0066] As in step S3 of Figure 8, the contour detection unit 43 first determines the shape of each aggregate for each image data by image processing using instance segmentation, and detects (extracts) the area of ​​each aggregate, i.e., aggregate area 91, according to the determined shape (step S13).

[0067] In this embodiment, for example, at this timing, the side length L10 of the minimum circumscribing square R1 of each aggregate region 91 as shown in FIG. 2(A) is calculated (step S15).

[0068] Thereafter, the contour detection unit 43 detects an approximation ellipse R2 of each aggregate region 91 as an "abstract figure" of each aggregate (step S17), as shown in Fig. 5(A), and calculates the major axis length L1 and the minor axis length L2 of the approximation ellipse R2 as the "contour feature amount" of each aggregate (step S19). In this way, the contour detection unit 43 identifies the contour feature amount of each aggregate through calculation processing using image processing technology.

[0069] If the side length L10 of the minimum circumscribing square R1 calculated in step S15 is less than 5 mm (NO in step S21), the estimation processing unit 44 determines the grain size classification from the side length L10 of the minimum circumscribing square R1 (step S23), since it is considered that the lengths of the long and short sides of the aggregate region 91 do not change significantly (the difference is small).

[0070] On the other hand, if the side length L10 of the minimum circumscribing square R1 of the aggregate region 91 is 5 mm or more (YES in step S21), the Mahalanobis distance from the Gaussian distributions G1 to G6 generated by the above-mentioned method is calculated as the similarity with each particle size division from the diameter lengths L1, L2 (contour feature amounts) of the approximate ellipse R2 (step S25). The estimation processing unit 44 determines the first candidate, second candidate, ... in order of closest distance.

[0071] Next, the estimation processing unit 44 sets a weighting factor according to the calculated Mahalanobis distance (step S27). In this embodiment, according to the above-mentioned classification method 1, the weighting factors of the granularity divisions of multiple candidates, including the first candidate, are calculated by normalizing them according to the distance so that the total sum is 1. As a result, for example, the weighting factors of the six granularity divisions (1) to (6) are calculated as [0.0], [0.1], [0.1], [0.3], [0.5], and [0.0]. In this example, the granularity division (5) is determined to be the first candidate, the granularity division (4) is determined to be the second candidate, and the granularity division (2) is determined to be the third candidate, and the granularity divisions (1) and (6) are not considered candidates.

[0072] At this timing, for example, the estimation processing unit 44 calculates the area of ​​the aggregate region 91 as the "aggregate area" (step S29) and outputs it to the estimation processing unit 44. Note that the method of calculating the aggregate area is not limited to this example, and for example, the area of ​​an approximation ellipse R2 of the aggregate region 91 may be regarded as the aggregate area. In this case, the approximation ellipse R2 may be identified directly (without detecting the aggregate region 91) by image processing the captured image. Furthermore, the calculation of the aggregate area may be simultaneous with or before the calculation of the weighting coefficient by the estimation processing unit 44.

[0073] Based on the results of calculating the weighting coefficient in step S27 and the results of calculating the aggregate area in step S29, the estimation processing unit 44 weights the aggregate area for each aggregate region 91 by the weighting coefficient and classifies it into multiple particle size divisions (step S30).

[0074] Specifically, for example, if the weighting coefficients for the six particle size divisions (1) to (6) calculated for a certain aggregate area 91 are [0.0], [0.1], [0.1], [0.3], [0.5], and [0.0], an area of ​​0.1 times the aggregate area is classified (accumulated) into particle size divisions (2) and (3), an area of ​​0.3 times the aggregate area is classified into particle size division (4), and an area of ​​0.5 times the aggregate area is classified into particle size division (5).

[0075] In this way, when the classification of the aggregate area for all aggregates detected as aggregate regions 91 is completed, the total area (cumulative area value) of each gradation class is obtained, and the area frequency for each gradation class can be calculated from the total area of ​​the aggregates contained in the material 9 (step S31). As described above, when extracting the region of each aggregate using instance segmentation, there is a possibility that small aggregates less than 5 mm will not be detected. Therefore, it is desirable to prepare an image (background image) of the state before the material 9 is allowed to flow, extract the difference from the background image for each captured image, and calculate the sum of these differences as the total area.

[0076] The output unit 45 outputs the calculation result in step S31, that is, the particle size distribution information of the material 9 (step S33). This completes the particle size distribution estimation process.

[0077] As described above, in this embodiment, the grain size distribution of each aggregate is estimated by processing the captured image of the freely falling material 9, so that the grain size distribution of the material 9 can be easily measured without performing a sieving test. Therefore, the estimation device 4 according to this embodiment makes it possible to easily control the quality of the material 9.

[0078] Furthermore, for aggregates in which the side length L10 of the minimum circumscribing square R1 is equal to or greater than a predetermined value (for example, 5 mm), a correction process for the particle size classification is performed using a statistical method, thereby enabling accurate estimation of the particle size distribution of the material 9. For aggregates in which the side length L10 of the minimum circumscribing square R1 is less than a predetermined value (for example, 5 mm), the particle size classification is determined based on the side length L10 of the minimum circumscribing square R1 without using the statistical method, thereby reducing the processing load on the estimation device 4.

[0079] It is desirable that extraction of aggregate region 91 and calculation of contour feature values ​​are performed only once for each aggregate. Therefore, the conveying speed of conveying unit 21 and the photographing speed of camera 24 may be adjusted so that one aggregate does not appear in multiple photographed images, or only aggregates appearing in a partial area of ​​each photographed image may be subjected to image processing, as shown in Fig. 10. Furthermore, if the resolution of the photographed images is low, preprocessing such as trimming or dividing the photographed images may be performed before image processing.

[0080] (Other classification methods for aggregate area) In this embodiment, the aggregate area is classified into a plurality of candidates according to the above-mentioned classification method 1, but other classification methods may also be employed.

[0081] · Classification method 2: The weighting coefficient of the first candidate particle size classification is set to 1 (100%), and the entire aggregate area is classified into the first candidate particle size classification.

[0082] According to classification method 2, it is possible to suppress an increase in error caused by classifying the aggregate area into a particle size classification that is far away (for example, the fifth candidate).

[0083] It is also possible to use both classification method 1 and classification method 2. For example, classification method 1 may be used as the basis, and classification method 2 may be adopted only when the weighting coefficient for the first candidate calculated by classification method 1 is equal to or greater than a predetermined value (for example, 0.6).

[0084] Alternatively, a method may be adopted in which the aggregate area is classified into only two gradation classes, the first and second candidates. In this case, the second candidate may be determined by associating it with the first candidate gradation class, as in classification method 3 shown below.

[0085] Classification method 3: The classification one level above or below the first candidate is designated as the second candidate, and weighting factors are set that are normalized according to the distance from the first candidate and the second candidate, and the aggregate area is weighted and classified according to the particle size classification of the first candidate and the second candidate, respectively.

[0086] Considering the results of an experiment on the effect of differences in the direction in which aggregate is photographed on the accuracy of estimating the gradation classification, one example would be to select the classification one level lower than the first candidate as the second candidate (for example, when the first candidate is equal to or larger than the predetermined gradation classification). Another example would be to select the classification one level higher than the first candidate as the second candidate (for example, when the first candidate is smaller than the predetermined gradation classification). Note that the above-mentioned "predetermined gradation classifications" can each be determined by experiment.

[0087] (Particle size distribution estimation accuracy) The evaluation results for each of classification methods 1 to 3 are shown in the graph of Fig. 11 and the table of Fig. 12. These evaluation results are those obtained when weighting coefficients are calculated based on the distance from a two-dimensional Gaussian distribution using the major axis length L1 and minor axis length L2 of the approximate ellipse as variables, and sieve openings of 5, 13, 20, 25, 30, and 40 correspond to the particle size categories (1) to (6) in the above embodiment, respectively. Fig. 12(A) shows the measured values ​​(area frequency) for each particle size category.

[0088] 11 and 12(B), classification method 1 (adding multiple weights) shows a large difference in particle size category (4) in particular, and the error with the measured data is relatively large. Although classification method 1 tends to judge values ​​to be larger than the measured values ​​overall, it is able to estimate the particle size distribution with a certain degree of accuracy.

[0089] As shown in Figure 12(C), classification method 2 (adding weights in the shortest time) has a smaller error with the measured data than classification method 1, and also has improved accuracy compared to evaluation data without statistical processing. Looking closely, classification method 2 tends to judge the gradation class of relatively large aggregates to be larger than the measured value, and the gradation class of relatively small aggregates to be smaller than the measured value.

[0090] In a method based on classification method 3 (changed weighting rule), as shown in Fig. 12(D), the error from the actual measured value is smaller in all particle size divisions than in classification method 2. Therefore, it was verified that particle size distribution can be estimated with high accuracy by adopting classification method 3. Note that in a method corresponding to the evaluation results in Fig. 12(D), the evaluation results of classification method 2 shown in Fig. 12(C) may be taken into consideration, and a second candidate may be determined according to the particle size division of the first candidate.

[0091] From the above, it was confirmed that although particle size classification is usually expressed as three-dimensional data such as mass ratio, that is, volume, by using any of the above classification methods 1 to 3, particle size distribution can also be estimated using two-dimensional data such as area.

[0092] (Variation) In this embodiment, the major axis length L1 and the minor axis length L2 of the approximate ellipse R2 of each aggregate are used as the "contour feature" and "learning contour feature," and a two-dimensional Gaussian distribution is used. However, this is not a limitation. For example, the minor axis length L2 of the approximate ellipse R2 of each aggregate may be used as the "contour feature" and "learning contour feature," and a one-dimensional Gaussian distribution may be used. Note that the variable of the one-dimensional Gaussian distribution may be either the major axis length L1 or the minor axis length L2.

[0093] Furthermore, in this embodiment, the region of each aggregate (aggregate region 91), the approximation ellipse, and the contour feature amount are detected (calculated) using image processing technology. However, the present invention is not limited to this example, and the detection (calculation) may be performed manually. That is, the contour feature amount of each aggregate may be input via an operation means (not shown) or the like. Therefore, the detection means for detecting the approximation ellipse and calculating the contour feature amount may be implemented manually, and the learning device 3 and the estimation device 4 may be provided with means for specifying the contour feature amount of the approximation ellipse by calculation or input value.

[0094] Furthermore, in this embodiment, an approximate ellipse of the aggregate region 91 is detected as the "abstract figure" of the aggregate. However, any figure other than an approximate ellipse may be used as long as the part corresponding to the minor axis and the part corresponding to the major axis of the aggregate region 91 can be expressed by two straight lines (which are perpendicular to each other). For example, a minimum circumscribing rectangle surrounding the aggregate region 91 may be detected as the "abstract figure" of the aggregate. In this case, the lengths of the minor and major sides of the minimum circumscribing rectangle may be used as the contour feature. In this way, the contour feature includes the minor axis or minor side (part corresponding to the minor axis) and the major axis or major side (part corresponding to the major axis) of a predetermined abstract figure.

[0095] Alternatively, without detecting an abstract figure from the aggregate region 91, the part corresponding to the minor axis and the part corresponding to the major axis of the aggregate region 91 may be detected based on a predetermined rule.

[0096] In addition, although the contour feature amount includes two variables, that is, the part corresponding to the minor axis and the part corresponding to the major axis, in this embodiment, other variables may be included. In other words, the closeness of the positional relationship between the detected contour feature amount including three variables and the three-dimensional Gaussian distribution may be evaluated.

[0097] In this embodiment, for aggregates with small particle sizes, the side length L10 of the minimum circumscribing square R1 of the aggregate region 91 is used as is to determine the particle size classification, but this is not limitative. Therefore, when the estimation device 4 estimates the particle size distribution, the process of detecting the minimum circumscribing square R1 may be omitted.

[0098] Furthermore, in the present embodiment, the learning device 3 and the estimation device 4 have been described as being separate devices, but they may be realized by a common device.

[0099] The estimation method executed by the estimation device 4 can also be provided as a program. Similarly, the estimation method executed by the estimation device 4 can also be provided as a program. Such a program can be provided by being recorded on an optical medium such as a CD-ROM (Compact Disc-ROM) or a computer-readable non-transitory recording medium such as a memory card. The program can also be provided by downloading it via a network.

[0100] The program according to the present invention may execute processing by calling necessary modules in a predetermined sequence at a predetermined timing among program modules provided as part of a computer's operating system (OS). In this case, the program itself does not include the modules, and executes processing in cooperation with the OS. Programs that do not include such modules may also be included in the program according to the present invention.

[0101] Furthermore, the program according to the present invention may be provided as a part of another program. In this case, the program itself does not include the modules included in the other program, and executes processing in cooperation with the other program. Such a program incorporated in another program may also be included in the program according to the present invention.

[0102] The embodiments disclosed herein should be considered to be illustrative in all respects and not restrictive. The scope of the present invention is defined by the claims, not by the above description, and is intended to include all modifications within the meaning and scope of the claims. [Explanation of symbols]

[0103] 1 particle size distribution estimation system, 2 imaging mechanism, 3 learning device, 4 estimation device, 9 material, 9A sample, 31, 41 input unit, 32 learning image memory unit, 33, 43 contour detection unit, 34 calculation unit, 35, 46 distribution information memory unit, 42 image memory unit, 44 estimation processing unit, 45 output unit, 91 aggregate region, D1, D2 distribution information, G1 to G6 Gaussian distribution, R1 minimum circumscribed square, R2 approximation ellipse.

Claims

1. A particle size distribution estimation device for estimating the particle size distribution of a material consisting of a large number of aggregates, an input means for inputting a photographed image of the material; a detection means for detecting a portion corresponding to a minor axis and a portion corresponding to a major axis of each aggregate from the photographed image and calculating a contour feature amount including the length of the portion corresponding to the minor axis and the length of the portion corresponding to the major axis; a storage means for storing distribution information indicating a distribution of learning contour feature values ​​including the lengths of the minor axis equivalents and the lengths of the major axis equivalents for each particle size division, which are calculated in advance based on the learning image; a determining means for determining a first candidate particle size division for each aggregate based on the proximity of a distribution of learning contour feature values ​​for each particle size division indicated by the distribution information to the contour feature value identified by the detecting means; a calculation means for calculating an aggregate area as an index value representing the amount of each aggregate from the photographed image; and an area integrating means for calculating a total area for each particle size class by classifying, for each aggregate, all or part of the aggregate area calculated by the calculating means into the first candidate particle size class.

2. the detecting means detects an abstracted figure obtained by abstracting the contour shape of each aggregate from the photographed image; The particle size distribution estimation device according to claim 1 , wherein the portion corresponding to the minor axis is a short side or a minor axis of the abstracted figure, and the portion corresponding to the major axis is a long side or a major axis of the abstracted figure.

3. The particle size distribution estimation device according to claim 2 , wherein the abstract figure is an approximate ellipse.

4. A particle size distribution estimation device described in any one of claims 1 to 3, wherein the distribution is a two-dimensional Gaussian distribution.

5. A particle size distribution estimation device for estimating the particle size distribution of a material consisting of a large number of aggregates, comprising: an input means for inputting a photographed image of the material; a detection means for detecting a portion corresponding to a minor axis and a portion corresponding to a major axis of each aggregate from the photographed image and calculating a contour feature amount including at least the length of the portion corresponding to the minor axis; a storage means for storing distribution information indicating a distribution of learning contour feature values ​​for each grain size classification, the distribution information being calculated in advance based on the learning image; a determining means for determining a first candidate particle size division for each aggregate based on the proximity of a distribution of learning contour feature values ​​for each particle size division indicated by the distribution information to the contour feature value identified by the detecting means; a calculation means for calculating an aggregate area as an index value representing the amount of each aggregate from the photographed image; and an area integrating means for classifying, for each aggregate, all or part of the aggregate area calculated by the calculating means into the first candidate gradation class, thereby calculating a total area for each gradation class; the determining means determines, for each aggregate, a plurality of candidates including the first candidate according to a distance between the distribution of learning contour feature values ​​for each particle size class indicated by the distribution information and the contour feature value; The particle size distribution estimation device, wherein the area integrating means sets a weighting coefficient according to the distance and classifies the aggregate area into each of the plurality of candidates.

6. A particle size distribution estimation device for estimating the particle size distribution of a material consisting of a large number of aggregates, comprising: an input means for inputting a photographed image of the material; a detection means for detecting a portion corresponding to a minor axis and a portion corresponding to a major axis of each aggregate from the photographed image and calculating a contour feature amount including at least the length of the portion corresponding to the minor axis; a storage means for storing distribution information indicating a distribution of learning contour feature values ​​for each grain size classification, the distribution information being calculated in advance based on the learning image; a determining means for determining a first candidate particle size division for each aggregate based on the proximity of a distribution of learning contour feature values ​​for each particle size division indicated by the distribution information to the contour feature value identified by the detecting means; a calculation means for calculating an aggregate area as an index value representing the amount of each aggregate from the photographed image; and an area integrating means for classifying, for each aggregate, all or part of the aggregate area calculated by the calculating means into the first candidate gradation class, thereby calculating a total area for each gradation class; the determining means determines a division one level above or one level below the first candidate as a second candidate, The particle size distribution estimation device, wherein the area accumulating means sets a normalized weighting coefficient according to the distance from the first candidate and the distance from the second candidate, and classifies the aggregate area into the particle size classification of the first candidate and the second candidate, respectively.

7. A step of inputting a photographed image of a material consisting of a large number of aggregates; detecting a portion corresponding to a minor axis and a portion corresponding to a major axis of each aggregate from the photographed image, and calculating a contour feature amount including the length of the portion corresponding to the minor axis and the length of the portion corresponding to the major axis; determining a first candidate particle size division for each aggregate based on the proximity of a positional relationship between a distribution of training contour feature values ​​including the lengths of the minor axis equivalents and the major axis equivalents in each particle size division, the distribution being determined in advance based on the training image; calculating an aggregate area as an index value representing the amount of each aggregate from the photographed image; and calculating a total area for each aggregate by classifying all or part of the aggregate area into the first candidate particle size class.

Images