JP2012198848A - Object identification device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an object identification device capable of automatically separating areas for each object in identifying each kind of objects using image recognition even when loaves of bread are partly in contact with each other.SOLUTION: Binarization is performed on a color image with two or more loaves of bread in contact with each other therein to separate a binarized image of the two or more loaves of bread, and polygonal approximation is performed on the separated contour lines to obtain the coordinates of apices, P1 to P33, of the approximate polygon. Then, the restriction vector (KH) indicating the curvature is obtained with each apex (P12) based on the relationship with apices nearby (P11-9, P13-14). The borderline is defined based on the length L of a candidate for the borderline connecting a pair of candidates for boundary points, the angles θ defined by the candidates for borderlines and each restriction vector, the area occupancy rate showing the overlap between the circular areas (E12-27) having the borderline candidates as the diameters and the whole image area, and the depth D of the restriction vector. The loaves of bread are separated by the borderlines.

Description

  The present invention relates to an object identification device for identifying an imaged object such as a pan, and more particularly to automatically identifying each object when a plurality of objects are in contact with each other in the same image. The present invention relates to a technique capable of separating each object.

Conventionally, for example, Patent Document 1 proposes a technique for obtaining an outline of an object in order to identify the object included in the image information.
Patent Document 2 proposes a technique for extracting shape features such as contours included in image information and identifying them by comparison with a similar model.
Furthermore, Patent Document 3 proposes a technique for accurately identifying two tablets even if the two tablets are partially in contact with each other when detecting the number of tablets included in the image information. In this case, at least a part of the connecting line connecting two points selected from a large number of points arranged along the inner contour line of the binary image that is in contact with each other to indicate the mass region is the region of the mass region. When passing outside, the two points belong to different tablets, that is, there are two tablets.

JP-A-7-22089 JP 2000-215315 A JP 2006-234519 A

  By the way, in recent years, the number of bakery selling homemade bread has been increasing, and the types of bread produced and sold at such bakery have also been increasing. In such a bakery, customers move the bread they want from the display stand to the tray, and the price is calculated in a state where it is placed on the tray, which takes time and effort. That is, the types of bread are various and diverse, and the appearance is not always the same even if they are the same type. Therefore, it is necessary for the salesperson to accurately identify the type of bread on the tray, and to input it accurately, for example, in the POS register based on the type of bread, which requires skilled work. For this reason, the applicant has been developing a technology that can image the pan on the tray with a CCD camera, automatically recognize the type and number of pans on the tray from this image information, and automate the price calculation. Yes.

  In such bread identification, when a customer places a plurality of breads on the tray from the display stand, it is conceivable that the trays are brought into the price calculation with the individual breads in contact with each other. In this case, even if automatic recognition from the image information is applied, for example, if two pans are in contact with each other to form one outline, even if a feature amount is obtained from the outline, It may be difficult or impossible to discriminate because it does not match what is modeled for the features of individual types of bread. For this reason, the sales staff must move the pans on the tray so that they are not in contact with each other before imaging, and the moving operation and the operation of the POS register are performed while maintaining the sanitary condition. It is thought that it will be time-consuming.

  The present invention has been made in view of such circumstances, and the object of the present invention is to capture each object based on the image even if the objects are imaged in a state where the objects are partially in contact with each other. An object of the present invention is to provide an object identification device that can automatically separate and classify individual objects.

  In order to achieve the above object, in the present invention, a color image captured by an imaging unit for two or more objects placed on the same plane is captured, and the type of the object is determined by recognizing the color image. The following specific items are provided for an object identification device including an image processing means for identification. That is, the image processing means includes a separation processing unit that separates and divides into regions belonging to individual objects when two or more objects included in the color image are in contact with each other. An identification process for the type of the object is executed for each region belonging to each object separated by the processing unit. Then, as the separation processing unit, a function of extracting the image region of the two or more objects from the color image by binarization processing, and a contour line of the image region of the two or more objects from the extracted binary image The bend specified by the relationship between the function that extracts vertices, the function that extracts each vertex constituting the approximate polygon by the polygon approximation process along the extracted contour line, and the vertices located before and after each vertex A function for obtaining the situation as a constricted vector, a length of a boundary line candidate that extracts a pair of vertices as boundary point candidates, and connects the pair of boundary point candidates, and each of the pair of boundary point candidates for the boundary line candidate The function of calculating the angle formed by the constriction vector as a parameter, and the calculated boundary line length is shorter and the angle formed by the constriction vector is closer to 180 degrees. A function of setting a boundary point candidate combination as a pair of boundary points and defining a boundary line; and a function of separating an image area of the two or more objects by a boundary line connecting the pair of boundary points (Claim 1).

  In the case of the present invention, even if two or more objects are imaged in contact with each other, the separation processing unit obtains the contour lines of the image areas of the two or more objects and obtains a polygon approximation along the contour lines. After extracting the vertices and calculating the constriction vector for each extracted vertex, the boundary point candidates are extracted based on the constriction vector, and the length and boundary line of the boundary line candidates connecting the pair of boundary point candidates are extracted. A boundary line is determined based on the angle formed between the candidate and the constriction vector. Then, since the image area in which two or more objects are integrated with the determined boundary line is divided into regions, the type of each object is surely identified based on the image region corresponding to each divided object. It will be. Therefore, for example, in constructing a system that automatically performs price calculation or the like by image identification using a bread as a target object, a salesperson pans on the tray before imaging a plurality of breads (objects) selected by the customer on the tray, for example. Without having to move the pans so that they are not in contact with each other, while maintaining the sanitary condition, the breads in contact are automatically separated (divided) into areas belonging to individual breads. Thus, each bread type can be automatically recognized by image identification. Thereby, in the retail business such as a bakery, it becomes possible to achieve significant labor saving and automation. In the identification process, the type of each object may be identified based on the feature amount relating to the shape of the color image portion corresponding to the image area divided into areas belonging to each object.

  In the object identification device of the invention, as a function of determining the boundary line, a function of extracting a pair of vertices as boundary point candidates by boundary point determination processing, and a boundary by boundary line determination processing based on the extracted boundary point candidates The boundary point determination process, the boundary point candidate is determined by the necking depth represented by the vector amount of each necking vector constituting a pair of vertices being equal to or greater than a predetermined threshold. It can be set as the structure to extract (Claim 2). By doing so, the extraction of boundary point candidates and the determination of boundary lines by extraction of the boundary point candidates can be realized more reliably.

  In this case, the boundary line determination process includes an arithmetic expression for calculating a boundary line score using the length of the boundary line candidate and the angle formed by the constriction vector as parameters. The boundary line can be determined based on (Claim 3). By doing in this way, the boundary line can be determined more reliably.

  Further, as the calculation formula, a boundary score is calculated by adding an area occupancy ratio, which is a ratio of a circular area having the diameter of the boundary line candidate as an overlap with the image area of the two or more objects, to the parameter. (Claim 4). Alternatively, the boundary expression score can be calculated by adding the constriction depth of the constriction vector to the parameter as the arithmetic expression (claim 5). By doing so, the boundary line based on the arithmetic expression can be determined more reliably.

  As described above, according to the object identification device of the present invention, even if two or more objects are imaged in contact with each other, the separation processing unit makes contact with each other in the image area of the two or more objects. It becomes possible to automatically determine the boundary line. Then, since the image region where two or more objects are integrated can be divided into regions with the determined boundary line, the type of each object can be surely determined based on the image region corresponding to each divided object. Can be identified. Therefore, for example, in constructing a system that automatically performs price calculation or the like by image identification using a bread as a target object, a salesperson pans on the tray before imaging a plurality of breads (objects) selected by the customer on the tray, for example. Without the work of moving the foods in a non-contact state with each other, the type of each bread is automatically recognized by image identification after automatically separating (dividing) into individual breads while maintaining a sanitary state. Will be able to. Thereby, in the retail business such as a bakery, it becomes possible to achieve significant labor saving and automation.

  In particular, according to claim 2, as a function of determining the boundary line, a function of extracting a pair of vertices as boundary point candidates by boundary point determination processing, and a boundary line determination process based on the extracted boundary point candidates A boundary line determination function, and as the boundary point determination process, a boundary point candidate is extracted when a necking depth represented by a vector amount of each necking vector constituting a pair of vertices is a predetermined threshold or more By adopting such a configuration, it becomes possible to more reliably realize the extraction of boundary point candidates and the determination of the boundary line by the extraction of boundary point candidates.

  According to a third aspect of the present invention, the boundary line determination process includes an arithmetic expression for calculating a boundary line score using the length of the boundary line candidate and the angle formed by the constriction vector as parameters, and the value of the boundary line score By adopting a configuration in which the boundary line is determined based on the size of the boundary, the boundary line can be determined more reliably.

  Further, according to claim 4, as the calculation formula, a boundary line is added by adding an area occupancy ratio, which is a ratio of a circular area having a diameter of the boundary line candidate to an image area of the two or more objects, to the parameter. By configuring so as to calculate the score, or according to claim 5, it is configured so that the boundary score is calculated by further adding the constriction depth of the constriction vector to the parameter as the arithmetic expression. Thus, the boundary line based on the arithmetic expression can be determined more reliably.

It is a schematic diagram which shows embodiment of the object identification apparatus of this invention. It is a block diagram which shows the image processing apparatus which comprises the object identification apparatus of FIG. It is a flowchart which shows the example of an image process. 4 is a flowchart illustrating an example of a separation process included in the image processing of FIG. 3. FIG. 5A is a plan view showing an example of a color image of two objects (pans) placed in a state where they are partially in contact with each other on the tray, and FIG. It is a top view which shows the binary image containing the two objects (pan) obtained by the digitization process. 6A is a plan view showing the contour line extracted from the binary image in FIG. 5B, and FIG. 6B is obtained from the contour line in FIG. 6A by polygon approximation processing. It is a top view which shows a polygon and each vertex. FIG. 7 is an explanatory plan view for explaining a first process of extracting vertices in a search range in order to obtain a constriction vector for each vertex in FIG. FIG. 8 is a diagram corresponding to FIG. 7 for describing a second process for obtaining a constriction direction vector from vertices extracted by the first process. FIG. 8 is a diagram corresponding to FIG. 7 for describing a third process for obtaining a constriction vector from a constriction direction vector obtained by the second process, and a fourth process for extracting a boundary point candidate thereafter. FIG. 8 is a diagram corresponding to FIG. 7 for describing a fifth process for determining a boundary line from boundary point candidates obtained by the fourth process. FIG. 8 is a diagram corresponding to FIG. 7 for explaining confirmation conditions for final determination of the boundary line obtained by the fifth process. It is explanatory drawing which illustrates about the order for carrying out the separation process of the partial contact between many objects (pan).

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  FIG. 1 shows an object identification device according to an embodiment of the present invention. Reference numeral 2 denotes an imaging means, which is constituted by, for example, a CCD camera, and an object (for example, pan; hereinafter, the object of the present embodiment will be described as pan) 4 placed on the tray 3 is true under illumination from the light source 5. Images are taken from above. The object 4 is an arbitrary number of breads, and in the present embodiment, for convenience of explanation, the following description will be made assuming that the two breads 4a and 4b are placed on the tray 3 in a state of partial contact. Reference numeral 6 denotes an image processing apparatus, and reference numeral 7 denotes, for example, a POS register connected to the image processing apparatus 6 in communication. The image processing device 6 is configured by, for example, a personal computer including a CPU, ROM, RAM, hard disk, communication interface, and appropriate input / display units. The hard disk is installed with an operating system and a program for executing image processing according to the present embodiment, so that image processing described below can be executed. The input unit may be a mouse or a keyboard, for example, and the display unit may be a display such as a liquid crystal panel, but a touch panel 61 that combines the display and a position input device such as a touch pad is adopted. May be. The POS register 7 receives the identification information of the type and number of breads 4a and 4b from the image processing device 6, and executes calculation / input / output such as display of total price, sales management, sales performance management, etc. ing. POS is an abbreviation for point of sale system. Note that the tray 3 is made of a translucent transparent material, and imaging is performed in a state in which the backlight from the backlight light source 8 placed behind (back side) is irradiated. Good. As a result, shadows that may be formed around the pans 4a and 4b can be eliminated as much as possible.

  As shown in FIG. 2, the image processing apparatus 6 includes an image storage unit 62 for storing an image (color digital image) captured and captured from the CCD camera 2, and two pans 4a and 4b. The separation processing unit 63 that separates the individual pans 4a and 4b from the image captured in the state of being brought into contact with each other, that is, separates the boundary lines of the individual pans 4a and 4b and separated from each other. And an identification processing unit 64 for identifying the type of each individual bread 4a, 4b based on the image. Furthermore, the image processing device 6 includes a master data storage unit 65 in which template information combined with feature quantities modeled for each type of bread is stored and registered in advance for identification by the identification processing unit 64; Learning to update the registration contents of the master data storage unit 65 by referring to the registration information in the master data storage unit 65 and performing learning processing on a new feature amount or the like acquired when the identification processing unit 64 executes the identification process. And a processing unit 66. The identification result (bread type and number) by the identification processing unit 64 is output to the POS register 7. Each processing such as the image storage unit 62, the separation processing unit 63, the identification processing unit 64, and the learning processing unit 66 is executed by a predetermined program as described above. It is comprised as what has the function to perform.

  The process executed by the image processing device 6 will be briefly described with reference to FIG. That is, first, an image captured by the CCD camera 2 is captured by inputting an image capture command from the touch panel 61, for example, and recorded in the image storage unit 62 (step S1). An example of this image WI is shown in FIG. In this example, each of the two kinds of breads 4a and 4b is partially in contact with each other. Next, after separating the areas belonging to the pans 4a and 4b by determining the boundary lines of the pans 4a and 4b based on the recorded images (step S2), the pans 4a and 4b The type is identified (step S3), and the identification result is output to the touch panel 61 and / or the POS register 7 (step S4).

  Next, the separation process (step S2) will be described in detail with reference to FIG. Since the image (initial image; color digital image) WI recorded in the image storage unit 62 is captured with the pans 4a and 4b placed on the tray 3, first, background separation processing (step S11) is performed. ) And binarization processing (step S12), the portion of the tray 3, which is the background, is separated, and only the portions of the pans 4a and 4b are cut out. As illustrated in FIG. A binary image BI in which 4b is continuously formed into a lump is obtained. By the above binarization processing, the grayscale information of each pixel appearing in the initial image is binarized into 1 and 0 with a predetermined threshold according to the saturation difference between the background tray 3 and the pans 4a and 4b. As a result, a binary image BI in which the areas of the pans 4a and 4b are displayed in monochrome of either one of white or black is obtained. That is, background separation processing is also performed by binarization processing.

  Next, the contour extraction process (step S13) is performed on the binary image BI, and after obtaining the contour line image CI illustrated in FIG. 6A, the contour line of the contour image CI is targeted. The polygon approximation process (step S14) is executed, and approximated by, for example, counterclockwise polygonal line approximation along the contour line and closing to the original starting point as illustrated in FIG. 6B as a polygon image PI. The polygon and each vertex P1 to P33 constituting this approximate polygon are obtained. Specifically, as the polygon image PI, an image is obtained in which the vertices P1 to P33 and a line segment connecting adjacent vertices are expressed as both areas of the pans 4a and 4b. The vertex coordinates of P1 to P33 are obtained. As the above contour extraction processing and polygon approximation processing, for example, cv Find Contour included in Open CV (Intel product name), which is an open source library for image processing, is used for contour extraction processing, and polyline approximation using the same cv Approx Poly is performed. Each can be used for polygon approximation processing.

  Next, processing based on boundary point determination processing (step S15) and boundary line determination processing (step S16) until the boundary lines of the pans 4a and 4b, which are characteristic processing portions in the present embodiment, are described below. This will be described in detail. As a general procedure, various parameters are extracted for each vertex based on the polygon image PI (specifically, each vertex coordinate), and the boundary score BS is determined for each pair of vertices based on the extracted various parameters. A combination of a pair of vertices having a high boundary line score BS is extracted as a boundary point candidate, and a boundary line is finally determined from the boundary point candidates. The various parameters for obtaining the boundary line score BS include, for example, the length L of the boundary line candidate connecting the pair of boundary point candidates, the constriction depth Di of each of the pair of boundary point candidates, and the boundary line candidate. Occupancy ratio that is the ratio of the angle θi formed by the direction of the constriction vector of each of the pair of boundary point candidates and the circular region having a diameter of the boundary line candidate connecting the pair of boundary point candidates to the binary image BI R.

  First, the constriction vector is obtained by the first process (FIG. 7), the second process (FIG. 8), and the third process (FIG. 9), and the depth Di and the direction of the constriction vector are obtained. In the first process, for each side in the front-rear direction from each vertex, for example, in the counterclockwise direction YA, the negative side (reverse direction) and the positive side (forward direction) are within a predetermined search range in the same direction as the adjacent vertex. Find existing vertices. Specifically, as shown in FIG. 7, for example, a search range (−α to + α) with a predetermined angle α is set on both sides with a direction extending from the vertex P12 to the vertex P11 adjacent to the minus side as the central direction C1. , Vertices P11, P10, and P9 existing in the search range are extracted. Similarly, a search range (-α to + α) with a predetermined angle α is set on both sides of a direction extending from the vertex P12 to the vertex P13 adjacent to the plus side as the center direction C2, and the vertexes P13, P13, P14 is extracted. In the example illustrated in FIG. 7, vertices P19, P18 existing in the negative search range are extracted from the vertex P20, and vertices P21, P22, P23 existing in the positive search range are extracted. Similarly, vertices P26, P25, and P24 existing in the negative search range are extracted from the vertex P27, and vertices P28, P30, and P31 existing in the positive search range are extracted. The search range (−α to + α) may be set so that the immediately preceding vertex when the two vertices are continuously outside the search range falls within the search range. For example, when the vertex P27 is set as the starting point, even if the vertex P29 is outside the search range, there is only one, and the vertices P32 and P33 are continuously outside the search range.

  In the second process, two constriction direction vectors KH1 and KH2 on the negative side and the positive side are obtained for each vertex. That is, for the minus side and the plus side of each target vertex, the same direction as the vector sum obtained by combining the vectors to each vertex extracted from the search range, and the same vector as the vector farthest from the target vertex The constriction direction vector defined by the vector quantity is obtained. Specifically, as shown in FIG. 8, for example, the same as the vector sum VT1 obtained by synthesizing the vector to the vertex P11, the vector to the vertex P10 and the vector to the vertex P9 in the negative search range starting from the vertex P12. A constriction direction vector KH1 having a direction and the same vector quantity as the vector from the vertex P12 to the farthest vertex P9 is obtained. Similarly, the same direction as the vector sum VT2 obtained by combining the vector to the vertex P13 and the vector to the vertex P14 in the plus side search range starting from the vertex P12, and the same as the vector to the vertex P14 farthest from the vertex P12 A constriction direction vector KH2 having a vector quantity is obtained.

  In the example illustrated in FIG. 8, the vertex P20 has the same direction as the vector sum VT1 obtained by combining the vector to the vertex P19 and the vector to the vertex P18 on the minus side starting from the vertex 20 and the vertex P18 farthest from the vertex P20. A constriction direction vector KH1 having the same vector amount as the vector to the vector, and the same direction as the vector sum VT1 obtained by combining the vector to the vertex P21, the vector to the vertex P22, and the vector to the vertex P23 on the plus side, A constriction direction vector KH2 having the same vector amount as the vector from the vertex P20 to the farthest vertex P23 is obtained. Similarly, with respect to the vertex P27, the vertex farthest from the vertex P27 in the same direction as the vector sum VT1 obtained by synthesizing the vector to the vertex P26, the vector to the vertex P25, and the vector to the vertex P24 starting from the vertex 27 While obtaining the constriction direction vector KH1 having the same vector quantity as the vector to P24, the same direction as the vector sum VT1 obtained by combining the vector to the vertex P28, the vector to the vertex P30, and the vector to the vertex P31 on the plus side, A constriction direction vector KH2 having the same vector quantity as the vector from the vertex P27 to the farthest vertex P31 is obtained.

  In the third process, a constriction vector KV is obtained from two constriction direction vectors KH1 and KH2 on the minus side and the plus side for each vertex. Specifically, as shown in FIG. 9, for example, when the vertex P12 is targeted, a triangle having two sides of the two constriction direction vectors KH1 and KH2 starting from the vertex P12 is assumed, and the vertex P12 is defined as A vector represented by the direction in which the bisector of the angle (the included angle) formed by the two constriction direction vectors KH1 and KH2 extends as the starting point and the length to the intersection where the bisector intersects the opposite side of the triangle Is defined as a constriction vector KV. The length of the constriction vector KV is defined as the depth of the constriction vector KV.

  As another example shown in FIG. 9, when the vertex P20 is targeted, a triangle having two sides of the two constriction direction vectors KH1 and KH2 starting from the vertex P20 is assumed, and the vertex P20 is the starting point. Constriction of a vector represented by the direction in which the bisector of the angle (the included angle) formed by the two constriction direction vectors KH1 and KH2 extends and the length to the intersection where the bisector intersects the opposite side of the triangle Obtained as vector KV. When the vertex P27 is targeted, a triangle having two sides of the two constriction direction vectors KH1 and KH2 starting from the vertex P27 is assumed, and the two constriction direction vectors KH1 and KH2 starting from the vertex P27. A vector represented by the direction in which the bisector of the angle (the included angle) extends and the length to the intersection where the bisector intersects the opposite side of the triangle may be obtained as the constriction vector KV.

  By the above first to third processes, the constriction vectors KV, KV,... At each vertex are obtained, and the depth Di and the direction of the constriction vector are obtained. Next, as a fourth process, boundary point candidates are extracted based on the depth Di of the constriction vectors KV, KV,... At these vertices Pi and the angle βi formed by the two constriction direction vectors KH1, KH2. . There are the following two criteria for extraction. In other words, the first determination criterion is an angle βi formed by two necking direction vectors KH1 and KH2 at each vertex (measured in the counterclockwise YA direction from the minus side necking direction vector KH1 to the plus side necking direction vector KH2. Angle) is smaller than a predetermined set angle, and the second criterion is that the depth Di of the constriction vector KV is equal to or greater than a predetermined threshold. As the setting angle, an angle of at least less than 180 degrees is set, preferably an angle of less than 150 degrees, and usually an angle of not more than 135 degrees may be set. As the depth Di of the constriction vector KV is larger than the threshold value, the constriction becomes deeper. In the fourth process, boundary point candidates are extracted based on both the above first and second determination criteria.

In the fifth process, the boundary line score BS is calculated for each boundary line candidate obtained from the boundary point candidates extracted in the fourth process, and the boundary line BS having a large value is determined as the boundary line. As shown for the pair of boundary point candidates P12 and P27 in FIG. 10, the boundary line score BS is a length L (a pair of boundary line candidates connecting the pair of boundary point candidates for each pair of boundary point candidates. The length between the vertex coordinates of the boundary point candidates), the necking depth Dj of each of the pair of boundary point candidates, the angle θj between the boundary line candidate and the direction of the respective necking vectors of the pair of boundary point candidates , Based on each parameter of area occupancy R, which is the ratio of a circular area having a diameter of a boundary line candidate connecting a pair of boundary point candidates to a binary image BI (see FIG. 5B). Calculation is performed according to equation (1). Here, since the boundary point candidates are a pair of opposites, the subscripts j = 1 and 2.
BS = -Nl·logL
+ Nd (logD1 + logD2)
-Ns {log (cosθ1 + 2) + log (cosθ2 + 2)}
+ Nr · logR (1)
Here, Nl, Nd, Ns, and Nr are coefficients for each parameter.

  In the above formula (1), the shorter the boundary line candidate L, the deeper the constriction depth Dj, the larger the angle θj is equal to or close to 180 degrees, or the area occupation ratio R is equal to or close to 100%. As shown, the boundary score BS becomes larger. For example, the area occupancy ratio R of a circular area shown as E12-27 (circle area having a diameter of a line connecting the vertices P12 and P27) in FIG. 10 is 100%, whereas E12-19 (vertex P12, The area occupancy ratio R of the circular area shown as a circular area having a diameter of the line segment connecting P19 is approximately 70%, and the value of the boundary score BS is larger in the former case. The magnitude of the value of the boundary score BS described above mainly depends on the shortness of the length L of the boundary term candidate of the first term and whether the θj of the third term is equal to or close to 180 degrees. It has become. Therefore, the boundary line score BS may be determined using only these two types of parameters as the minimum parameters. Alternatively, the value of the boundary score BS may be determined by further adding a parameter based on the area occupancy R in the fourth term.

  As for the boundary line obtained based on the value of the boundary line score BS in the fifth process, the boundary line finally clearing the following confirming condition is finally determined. That is, (a) a pair of boundary points specifying a boundary line correspond one-to-one, and no two boundary lines are generated from one boundary point (for example, P12 in FIG. 11), (b) two boundary points Boundary lines (for example, a line connecting P16-P29 and a line connecting P12-P27 in FIG. 11) do not intersect each other, (c) a pair of boundary points form constrictions facing each other, in other words, The value of cos θj using θj of (b) is equal to or greater than a predetermined threshold value. (D) Two boundary lines connecting adjacent boundary points (for example, a line connecting P12-P27 in FIG. 11 and P11-P28) Applying the confirming condition that the connecting line) is correct when the boundary line score BS value is as large as possible, and (e) the connecting point between adjacent boundary points is not a boundary line. To the boundary line (in FIG. 11, P12-P27 Confirm the connecting line.

  Finally, a labeling process (step S17; refer to FIG. 4) for setting different values for each area is performed on each pixel in the two areas separated across the boundary line, and the area of the pan 4a, The pan 4b is divided into regions. The region information after the separation processing is output to the identification processing unit 64, and the type of bread 4a, 4b for each region is identified.

  In the above, the case where the separation process is performed for the example in which the two breads 4a and 4b are in contact with each other has been described. Can be executed. That is, as shown in FIG. 12 in which seven pans 4c, 4d, 4e, 4f, 4g, 4h, and 4i are in contact with each other, first, there is a boundary between the blank portions partitioned and formed in a closed state by contact. A line (for example, a boundary line of parts indicated by reference numerals 90 and 91) is repeatedly determined and separated, and then the corresponding boundary line (for example, reference numerals 92 and 93 is set so that the closed blank portion is opened). The boundary line of the part to be shown) is determined and separated, and if there is no blank space in the closed state, it is finally left so that it is separated into individual pans 4c, 4d, 4e, 4f, 4g, 4h, 4i. A boundary line (for example, a boundary line of parts indicated by reference numerals 94 to 99) may be determined and separated.

  According to this embodiment, even if they are of the same type, there is a large individual difference in the baked shape, and although there are many almost circular shapes, there is no perfect circular shape, but from a nearly circular shape to an elliptical shape When a pan having a distorted shape close to a shape is used as the object 4 to be identified or separated, individual objects (pans 4a) based on images captured in a state where a plurality of parts are partially in contact with each other. , 4b) can be automatically separated and divided into four. This avoids a situation in which it is difficult or impossible to identify each object ((pan 4a, 4b) 4 in a state where they are in contact with each other, and each type of individual object ((pan 4a, 4b) 4 is automatically and For example, automatic calculation of price, automatic management of sales performance, etc. can be performed, for example. In constructing an automatic system, the salesperson moves the pans on the tray so that they are not in contact with each other before imaging a plurality of pans (objects) selected by the customer on the tray, for example. In addition, a plurality of breads in contact are automatically separated (divided) into areas belonging to individual breads while maintaining a sanitary condition, and each bread type is automatically recognized by image identification. It can be so made. As a result, in the retail industry, such as a bakery, so that it is possible to achieve a significant labor saving and automation.

<Other embodiments>
In addition, this invention is not limited to the said embodiment, Various other embodiments are included. That is, in the embodiment, the example in which the separation process is performed on the binary image BI has been described. However, the present invention is not limited to this. In this case, as various parameters for obtaining the boundary score BS, in addition to those exemplified in the present embodiment, an edge image is created from the initial image to generate the boundary line. An average value of edge image density values (edge strength) in an area with a predetermined width along the candidate and / or a luminance image is created from the initial image, and an area with a predetermined width along the boundary line candidate is created. You may make it add and use the average value of the density value (luminance) of a luminance image.

2 CCD camera (imaging means)
3 Tray 4 Object 4a, 4b Pan 5 Light source 6 Image processing device (image processing means)
63 Separation processing unit 64 Identification processing unit

Claims (5)

An object identification apparatus provided with an image processing unit that captures a color image captured by an imaging unit with respect to two or more objects placed on the same plane and recognizes the type of the object by recognizing the color image. There,
The image processing unit includes a separation processing unit that separates and classifies the two or more objects included in the color image into regions belonging to individual objects when the two or more objects are in contact with each other, and is separated by the separation processing unit. Configured to execute identification processing for the type of the object for each region belonging to the individual object,
The separation processing unit extracts a region of the two or more objects from the color image by binarization processing, and extracts an outline of the image region of the two or more objects from the extracted binarized image A bending state specified by the relationship between the function to extract the vertices constituting the approximate polygon by the polygon approximation process along the extracted contour line, and the vertices located before and after each vertex. A function to obtain a constriction vector, a length of a boundary line candidate that extracts a pair of vertices as boundary point candidates, and connects the pair of boundary point candidates, and the constriction in each of the pair of boundary point candidates with respect to the boundary line candidate A function for calculating an angle formed by a vector as a parameter, and a pair of boundary points whose calculated boundary line length is shorter and the angle formed by the constriction vector is closer to 180 degrees. And a function for setting the boundary line as a pair of boundary points, and a function for separating the image area of the two or more objects by a boundary line connecting the pair of boundary points. An object identification device. The object identification device according to claim 1,
As a function of determining the boundary line, a function of extracting a pair of vertices as boundary point candidates by boundary point determination processing, and a function of determining a boundary line by boundary line determination processing based on the extracted boundary point candidates are provided. And
Object identification configured to extract boundary point candidates when the necking depth represented by the vector amount of each necking vector constituting a pair of vertices is greater than or equal to a predetermined threshold as the boundary point determination processing apparatus. The object identification device according to claim 2,
The boundary line determination processing includes an arithmetic expression for calculating a boundary line score using the length of the boundary line candidate and the angle formed by the constriction vector as parameters, and the boundary line is determined based on the value of the boundary line score. An object identification device configured to: The object identification device according to claim 3,
The calculation formula is configured to calculate a boundary score by adding an area occupancy ratio, which is a ratio of a circular area having a diameter of the boundary line candidate to an image area of the two or more objects, to a parameter. An object identification device. The object identification device according to claim 4,
The arithmetic expression is configured to calculate a boundary score by adding a constriction depth of the constriction vector to a parameter as well.

Images