JP7696255B2 - Learning support device, learning support method, and learning support program - Google Patents

Description

The disclosed embodiments relate to a learning support device, a learning support method, and a learning support program.

Conventionally, annotation is known as a process for creating training data that serves as positive examples when training an AI (Artificial Intelligence) model using algorithms such as deep learning (see, for example, Patent Document 1).

Among these annotations, in annotations for AI models for image recognition, rectangles called bounding boxes (hereafter referred to as "BBs") that surround each object in the image are created manually using so-called VDT (Visual Display Terminals) work. In addition, metadata indicating the name, attributes, etc. of each object is assigned to each created BB.

JP 2010-009337 A

However, conventional techniques leave room for further improvement in ensuring annotation quality.

For example, to improve the accuracy of an AI model, it is necessary to collect a large number of images and annotate each of these images. However, the VDT work described above requires a huge amount of man-hours and causes great physical and mental fatigue for the workers.

In particular, it is preferable to create the BB described above so as to include as few extraneous areas other than the object as possible in order to improve the quality of the training data. However, this is not easy to do manually, and there is a risk that the quality will be affected by the skill level of the worker.

One aspect of the embodiment has been made in consideration of the above, and aims to provide a learning support device, a learning support method, and a learning support program that can ensure the quality of annotations for an AI model for image recognition.

A learning support device according to one aspect of an embodiment includes a controller that extracts edges of an object in a target image for annotation in creating teacher data , creates a closed-edge rectangle tangent to a closed edge that is an edge extracted as a closed curve from among the extracted edges, calculates a similarity between an input rectangle input by a user for the target image and the closed-edge rectangle, selects the closed-edge rectangle with the largest similarity , and employs the selected closed-edge rectangle as a bounding box for the annotation .

According to one aspect of the embodiment, it is possible to ensure the quality of annotations for an AI model for image recognition.

FIG. 1 is a diagram showing a schematic configuration of a learning support device according to an embodiment. FIG. 2 is a schematic explanatory diagram (part 1) of the learning support method according to the embodiment. FIG. 3 is a schematic explanatory diagram (part 2) of the learning support method according to the embodiment. FIG. 4 is a schematic explanatory diagram (part 3) of the learning support method according to the embodiment. FIG. 5 is a schematic explanatory diagram (part 4) of the learning support method according to the embodiment. FIG. 6 is a block diagram illustrating an example of the configuration of a learning support device according to an embodiment. FIG. 7 is an explanatory diagram of the closed edge rectangle creation process. FIG. 8 is a diagram (part 1) showing a specific example of the most similar rectangle selection process. FIG. 9 is a diagram (part 2) showing a specific example of the most similar rectangle selection process. FIG. 10 is a diagram (part 3) showing a specific example of the most similar rectangle selection process. FIG. 11 is a diagram (part 4) showing a specific example of the most similar rectangle selection process. FIG. 12 is a diagram (part 5) showing a specific example of the most similar rectangle selection process. FIG. 13 is a diagram (part 1) showing a specific example of the screen UI. FIG. 14 is a diagram (part 2) showing a specific example of the screen UI. FIG. 15 is a diagram (part 3) showing a specific example of the screen UI. FIG. 16 is a diagram (part 4) showing a specific example of the screen UI. FIG. 17 is a diagram (part 5) showing a specific example of the screen UI. FIG. 18 is a diagram (part 6) showing a specific example of the screen UI. FIG. 19 is a diagram (part 7) showing a specific example of the screen UI. FIG. 20 is a diagram (part 8) showing a specific example of the screen UI. FIG. 21 is a diagram (part 9) showing a specific example of the screen UI. FIG. 22 is a diagram (part 10) showing a specific example of the screen UI. FIG. 23 is a diagram (part 11) showing a specific example of the screen UI. FIG. 24 is a diagram (part 12) showing a specific example of the screen UI. FIG. 25 is a diagram (part 13) showing a specific example of the screen UI. FIG. 26 is a diagram (part 14) showing a specific example of the screen UI. FIG. 27 shows a processing sequence executed by the learning assistance device according to the embodiment.

Below, embodiments of the learning support device, learning support method, and learning support program disclosed in the present application will be described in detail with reference to the attached drawings. Note that the present invention is not limited to the embodiments described below.

First, an overview of the learning support method according to the embodiment will be described with reference to Figs. 1 to 5. Fig. 1 is a diagram showing a schematic configuration of a learning support device 10 according to the embodiment. Figs. 2 to 5 are diagrams (part 1) to (part 4) showing an overview of the learning support method according to the embodiment.

The learning support device 10 is a computer used by a user U, who is an annotator, when performing annotation for an AI model for image recognition. The learning support device 10 is, for example, a desktop or notebook PC (Personal Computer), a tablet terminal, a smartphone, a server, a workstation, etc.

As shown in FIG. 1, the learning support device 10 according to the embodiment has an HMI (Human Machine Interface) unit 3. The learning support device 10 also has a target image DB 11a and a rectangle information DB 11e.

The HMI unit 3 is a component that provides interface components for the user U. The HMI unit 3 includes an input unit 3a and an output unit 3b.

The input unit 3a is an input device that accepts input operations from the user U, and is realized, for example, by a keyboard, a mouse, a pen tablet, a touch panel, etc. Note that the input unit 3a may also be realized by a software component.

The output unit 3b is an output device that displays and outputs the target image of the annotation and the BB input on the target image, and is realized by a display or the like. Note that the output unit 3b may be configured integrally with the input unit 3a by a touch panel display.

The target image DB11a is a database that stores each image that is the subject of annotation work. The rectangle information DB11e is a database that stores rectangle information, which is information about the position, size, etc. of the BB created on each image during annotation.

The user U performs VDT work via the HMI unit 3 to annotate each image stored in the target image DB 11a, and the rectangle information related to the BB created as a result is stored in the rectangle information DB 11e.

However, existing annotation technologies like this leave room for further improvement in terms of ensuring annotation quality.

For example, to improve the accuracy of an AI model, it is necessary to collect a large number of images and annotate each of these images, but this requires a huge amount of man-hours in VDT work and causes great physical and mental fatigue for the user U.

In particular, it is preferable to create the BB so that it includes as little extraneous area as possible other than the object in order to improve the quality of the training data, but this is not easy to do manually, and there is a risk that the quality will be affected by the level of skill of the user U.

Therefore, in the learning support method according to the embodiment, the edges of the object in the target image for annotation are extracted, and from among the extracted edges, a closed edge rectangle is created that is tangent to a closed edge, which is an edge extracted as a closed curve. The similarity between the input rectangle input to the target image by user U and the closed edge rectangle is calculated, and the closed edge rectangle with the greatest similarity is selected.

2, in the learning support method according to the embodiment, first, the learning support device 10 creates an input rectangle IR for the target image for annotation by an input operation of the user U (step S1). As shown in the figure, the user U specifies the input rectangle IR, for example, by performing an area selection operation with a mouse or the like from a start point P1 to an end point P2 on the target image.

On the other hand, the learning support device 10 extracts edges in advance for the same target image, as shown in FIG. 3 (step S2). For such edge extraction, a known algorithm such as the Canny edge detector may be used, or an original algorithm may be used.

Then, as shown in Figure 4, for the extracted edges that are extracted as closed curves, i.e., closed edges with matching ends (hereafter referred to as "closed edges"), the learning support device 10 automatically creates rectangles tangent to the closed edges (step S3). The rectangles tangent to the closed hedges (hereafter referred to as "closed edge rectangles") are shown by dashed lines in the figure.

Then, as shown in the figure, the learning support device 10 selects, from among the closed edge rectangles, the closed edge rectangle that is most similar to the input rectangle IR as the automatically adjusted rectangle AR (step S4). The method for selecting the most similar closed edge rectangle will be described later using Figure 8 and subsequent figures.

Then, as shown in FIG. 5, the learning support device 10 displays on the output unit 3b the automatically adjusted rectangular AR along with a dialog asking the user U whether or not to adopt the rectangular AR (step S5).

As shown in the figure, if the user U selects "Yes" here, the learning support device 10 uses the adjusted rectangle AR as BB. On the other hand, if the user U selects "No", the learning support device 10 uses the rectangle before adjustment, i.e., the input rectangle IR, as BB.

In this way, the learning support method according to the embodiment extracts the edges of an object in an image to be annotated, creates a closed edge rectangle that is tangent to a closed edge, which is an edge extracted as a closed curve from among the extracted edges, calculates the similarity between the input rectangle IR input to the target image by user U and the closed edge rectangle, and selects the closed edge rectangle with the greatest similarity.

Therefore, according to the learning support method of the embodiment, it is possible to significantly reduce the amount of work required to create a BB for each image collected in large quantities, which in the existing technology required the user U to do one by one based on his/her personal sense and judgment. This also makes it possible to reduce the mental and physical fatigue of the user U during VDT work.

In addition, according to the learning support method of the embodiment, a closed edge rectangle that is automatically created using image analysis technology including edge extraction and that is in contact with the boundary of an object in an image can be used as the BB, thereby reducing the influence of the user U's level of proficiency on the quality of the training data.

In other words, according to the learning support method of the embodiment, it is possible to ensure the quality of annotations for an AI model for image recognition. Below, a more specific description is given of an example configuration of a learning support device 10 to which the learning support method of the embodiment is applied.

Figure 6 is a block diagram showing an example of the configuration of a learning support device 10 according to an embodiment. Note that Figure 6 shows only the components necessary to explain the features of the embodiment, and does not include general components.

In other words, each component shown in FIG. 6 is a functional concept, and does not necessarily have to be physically configured as shown. For example, the specific form of distribution and integration of each block is not limited to that shown, and all or part of it can be functionally or physically distributed and integrated in any unit depending on various loads, usage conditions, etc.

In addition, in the explanation using Figure 6, the explanation of components that have already been explained may be simplified or omitted.

As shown in FIG. 6, the learning support device 10 according to the embodiment includes a memory unit 11 and a control unit 12. The learning support device 10 is also connected to an HMI unit 3 via a wired or wireless connection, or directly.

Since the HMI unit 3 has already been described, a description thereof will be omitted here. The storage unit 11 is realized, for example, by a semiconductor memory element such as a random access memory (RAM) or a flash memory, or a storage device such as a hard disk or an optical disk. In the example of FIG. 6, the storage unit 11 stores a target image DB 11a, extraction algorithm information 11b, edge information 11c, similarity calculation information 11d, and a rectangle information DB 11e.

The target image DB 11a and rectangle information DB 11e have already been explained, so their explanation will be omitted here. The extraction algorithm information 11b is library information of algorithms used in the edge extraction process executed by the edge extraction unit 12ba, which will be described later.

The edge information 11c stores information about the edges extracted by the edge extraction unit 12ba. The similarity calculation information 11d stores algorithms and various parameters that serve as the basis for calculating the similarity between the input rectangle IR and the closed edge rectangle. The similarity calculation information 11d can be selected in advance by the user U, for example, and set in advance as static information.

The control unit 12 is a controller, and is realized, for example, by a CPU (Central Processing Unit) or MPU (Micro Processing Unit) executing various programs (not shown) stored in the memory unit 11 using a RAM as a working area. The control unit 12 can also be realized, for example, by an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

The control unit 12 has an image depiction unit 12a, an image analysis unit 12b, and a rectangle creation unit 12c, and realizes or executes the information processing functions and actions described below.

The image depiction unit 12a outputs the target image of the annotation stored in the target image DB 11a to the output unit 3b. The image depiction unit 12a includes a rectangle drawing unit 12aa. The rectangle drawing unit 12aa draws the input rectangle IR input via the input unit 3a on the target image.

The image analysis unit 12b performs image analysis processing on the target image. The image analysis unit 12b includes an edge extraction unit 12ba. The edge extraction unit 12ba performs edge extraction processing on the target image based on the extraction algorithm information 11b.

In addition to the above-mentioned Canny algorithm, the edge extraction unit 12ba can use known algorithms for edge extraction, such as Sobel and Laplacian.

The edge extraction unit 12ba may perform edge extraction processing using an algorithm arbitrarily selected by the user U from among these publicly known algorithms and original algorithms.

By extracting the boundaries of objects in an image using this type of image analysis processing, it is possible to maintain a certain level of quality without relying on the senses and judgments of the user U. Furthermore, by allowing the algorithm to be selected arbitrarily, it is possible to realize functions that are not dependent on the characteristics of the training data. For example, by using different algorithms depending on the color or shape of the object, or the brightness or saturation of the image, it is possible to extract the edges of the object with high precision.

The rectangle creation unit 12c includes a closed edge rectangle creation unit 12ca and a most similar rectangle selection unit 12cb. The closed edge rectangle creation unit 12ca creates a closed edge rectangle that is in contact with the above-mentioned closed edge based on the edge information 11c extracted by the edge extraction unit 12ba.

Here, FIG. 7 is an explanatory diagram of the closed edge rectangle creation process. As shown in FIG. 7, the closed edge rectangle creation unit 12ca obtains the minimum horizontal position value X1, maximum horizontal position value X2, minimum vertical position value Y1, and maximum vertical position value Y2 of the closed edge Ce from the edge information 11c.

Then, the closed edge rectangle creation unit 12ca creates a closed edge rectangle R using the four coordinates (X1, Y1), (X1, Y2), (X2, Y1), and (X2, Y2) defined by these, as shown in the same figure.

This eliminates the need for user U to adjust BB to match the boundaries of the object one by one, significantly reducing the amount of work required for VDT work and easing the burden on user U.

Returning to the explanation of FIG. 6, the most similar rectangle selection unit 12cb calculates the similarity between each closed edge rectangle R created by the closed edge rectangle creation unit 12ca and the input rectangle IR based on the similarity calculation information 11d, and selects the closed edge rectangle R with the maximum similarity.

The algorithm for calculating the similarity can be selected by the user U as desired, as in the case of edge extraction described above. The most similar rectangle selection unit 12cb uses, for example, aspect ratio, center position, width/length, etc. as elements for measuring the degree of similarity. The most similar rectangle selection unit 12cb then calculates the similarity from the sum of squares of the element differences between the input rectangle IR and each closed edge rectangle R, and selects the closed edge rectangle R with the maximum similarity.

Here, specific examples of the most similar rectangle selection process will be described with reference to Figs. 8 to 12. Figs. 8 to 12 are diagrams (part 1) to (part 5) showing specific examples of the most similar rectangle selection process.

First, as shown in Figure 8, the input rectangle IR has a center position of (4, 3), a width of 8, and a height of 6, and contains closed edges of a star, a circle, and a pentagon.

In such a case, as shown in FIG. 9, when comparing only the aspect ratios, the most similar rectangle selection unit 12cb selects the closed edge rectangle R1 that has approximately the same aspect ratio as the input rectangle IR from among the closed edge rectangle R1 that is adjacent to the closed edge of a star, the closed edge rectangle R2 that is adjacent to the closed edge of a circle, and the closed edge rectangle R3 that is adjacent to the closed edge of a pentagon.

Also, as shown in FIG. 10, when comparing only the center positions, the most similar rectangle selection unit 12cb selects the closed edge rectangle R2, which has a center position that is approximately the same as that of the input rectangle IR, from among the closed edge rectangles R1, R2, and R3.

Also, as shown in FIG. 11, when comparing only the width/height, the most similar rectangle selection unit 12cb selects the closed edge rectangle R3, which is the one that is most similar in width/height to the input rectangle IR, from among the closed edge rectangles R1, R2, and R3.

As shown in FIG. 12, for example, the input rectangle IR1 contains only a portion of closed edges, and no closed edges are detected within the input rectangle IR1. In such a case, as shown in the figure, the most similar rectangle selection unit 12cb expands the input rectangle IR1 by a certain amount, detects closed edges within the expanded input rectangle IR2, and creates a closed edge rectangle R1.

In this way, by assisting in determining the boundaries of an object from inside and outside the input rectangle IR, it becomes unnecessary to create a precise input rectangle IR, thereby reducing the burden on the user U.

In the explanation using Figure 5, an example of a screen UI (User Interface) that asks the user U whether or not to adopt the rectangle AR after automatic adjustment to the input rectangle IR by selecting "Yes" or "No" was given, but the screen UI is not limited to this.

Various specific examples of screen UIs will be explained below with reference to Figs. 13 to 26. Figs. 13 to 26 are diagrams (1) to (14) showing specific examples of screen UIs.

As shown in FIG. 13, for example, the learning support device 10 may automatically adopt the automatically adjusted rectangle AR as BB for the input rectangle IR input by the user U, without inquiring about the above-mentioned adoption.

As also shown in FIG. 5, as shown in FIG. 14, for example, the learning support device 10 may inquire whether or not to adopt the automatically adjusted rectangle AR for the input rectangle IR input by the user U.

In this case, as shown in FIG. 15, if the user U selects "Yes," the learning support device 10 will adopt the automatically adjusted rectangle AR as BB. On the other hand, as shown in FIG. 16, if the user U selects "No," the learning support device 10 will adopt the input rectangle IR as BB.

As shown in FIG. 17, the input rectangle IR input by the user U includes multiple closed edges. This case also includes the case where the input rectangle IR is expanded. In addition, it includes two closed edges, a pentagon and a circle.

In such a case, as shown in the figure, for example, the learning support device 10 may automatically adopt as BB the rectangle AR that has the greatest similarity to the input rectangle IR input by the user U, without inquiring about whether or not to adopt the rectangle AR.

In a similar case, as shown in FIG. 18, for example, the learning support device 10 may first inquire whether or not to adopt the first candidate rectangle AR1 that has the highest similarity for the input rectangle IR input by the user U.

In this case, as shown in FIG. 19, if the user U selects "Yes," the learning support device 10 will adopt the first candidate rectangle AR1 as BB. On the other hand, as shown in FIG. 20, if the user U selects "No," the learning support device 10 will inquire as to whether or not to adopt the second candidate rectangle AR2, which has the next highest similarity.

In this case, if the user U selects "Yes" as shown in FIG. 21, the learning support device 10 will adopt the second candidate rectangle AR2 as BB. On the other hand, if the user U selects "No" as shown in FIG. 22, the learning support device 10 will adopt the input rectangle IR as BB.

For example, the learning support device 10 may be configured to allow selection with the directional keys on the keyboard and adoption with the Enter key, as shown in FIG. 23, in the flows shown in FIG. 18 to FIG. 22.

Similarly, for example, the learning support device 10 may be configured to allow selection and adoption by clicking with a mouse or tapping with a finger, as shown in FIG. 24.

Next, consider the case where the input rectangle IR includes an already annotated rectangle AR. As shown in FIG. 25, assume that the input rectangle IR includes annotated rectangles AR1 and AR2.

In such a case, as shown in the figure, for example, the learning support device 10 selects and adopts a rectangle AR3 that has not yet been annotated, using one of the methods shown in Figures 13 to 16.

As shown in FIG. 26, the input rectangle IR includes an annotated rectangle AR1, as well as multiple closed edges that have not been annotated.

In such a case, as shown in the figure, the learning support device 10 selects and adopts, for example, rectangle AR2 from among rectangles AR2 and AR3 that have not yet been annotated, using one of the methods shown in Figures 17 to 24.

Next, the processing sequence executed by the learning support device 10 will be described with reference to FIG. 27. FIG. 27 shows the processing sequence executed by the learning support device 10 according to the embodiment. Note that FIG. 27 shows the processing sequence up to the creation of one BB in one target image.

First, when a target image for annotation is selected by the user U via the HMI unit 3 (step S101), the image depiction unit 12a displays the target image on the HMI unit 3 (step S102). The image depiction unit 12a also transmits the target image to the image analysis unit 12b.

The image analysis unit 12b causes the edge extraction unit 12ba to execute edge extraction processing based on the target image and the extraction algorithm information 11b in the storage unit 11 (step S103), and writes the edge information 11c, which is the processing result, to the storage unit 11.

Then, when a rectangle input is received from the user U via the HMI unit 3 (step S104), the image depiction unit 12a draws the input rectangle IR on the HMI unit 3 (step S105). The image depiction unit 12a also transmits input rectangle information related to the input rectangle IR to the rectangle creation unit 12c.

The closed edge rectangle creation unit 12ca of the rectangle creation unit 12c executes a closed edge rectangle creation process based on the input rectangle IR and the edge information 11c of the memory unit 11 (step S106).

Then, the most similar rectangle selection unit 12cb executes the most similar rectangle selection process based on the result of the closed edge rectangle creation process (step S107), and transmits most similar rectangle information about the most similar rectangle, which is the result of the process, to the image depiction unit 12a. The image depiction unit 12a draws the most similar rectangle on the HMI unit 3 based on the received most similar rectangle information (step S108).

Then, when the image depiction unit 12a receives a selection of the rectangular AR to be adopted from the user U via the HMI unit 3 (step S109), it draws the rectangular AR on the HMI unit 3 (step S110) and transmits adopted rectangle information regarding the adopted rectangular AR to the rectangle creation unit 12c (step S111).

Then, the rectangle creation unit 12c writes the received adopted rectangle information to the rectangle information DB 11e of the storage unit 11 (step S112) and ends the process.

As described above, the learning support device 10 according to the embodiment includes an edge extraction unit 12ba (corresponding to an example of an "extraction unit"), a closed edge rectangle creation unit 12ca (corresponding to an example of a "creation unit"), and a most similar rectangle selection unit 12cb (corresponding to an example of a "selection unit"). The edge extraction unit 12ba extracts edges of objects in a target image for annotation. The closed edge rectangle creation unit 12ca creates closed edge rectangles that are tangent to closed edges, which are edges extracted as closed curves from among the edges extracted by the edge extraction unit 12ba. The most similar rectangle selection unit 12cb calculates the similarity between an input rectangle IR input to the target image by a user U and a closed edge rectangle, and selects the closed edge rectangle with the greatest similarity.

Therefore, according to the learning support device 10 of the embodiment, it is possible to ensure the quality of annotations for an AI model for image recognition.

The edge extraction unit 12ba also extracts the edges of the object using an edge extraction algorithm arbitrarily selected by the user U.

Therefore, the learning support device 10 according to the embodiment can realize functions that are not dependent on the characteristics of the teacher data. For example, it is possible to extract the edges of an object with high accuracy by using different algorithms depending on the color or shape of the object, or the brightness or saturation of the image.

The closed edge rectangle creation unit 12ca also creates the closed edge rectangle, with the coordinate positions of the four corners defined by the minimum horizontal position, maximum horizontal position, minimum vertical position, and maximum vertical position of the closed edge being the vertex positions.

Therefore, according to the learning support device 10 of the embodiment, it is possible to automatically and accurately create a closed edge rectangle that is in contact with a closed edge, and the input rectangle IR can be automatically adjusted with high precision.

The most similar rectangle selection unit 12cb also calculates the similarity between the input rectangle IR and the closed edge rectangle using at least the aspect ratio, the center position, or the width and height as comparison elements.

Therefore, according to the learning support device 10 of the embodiment, it is possible to accurately calculate the similarity using at least the aspect ratio, the center position, or the width and height as comparison elements, and to appropriately select the most similar rectangle AR based on the similarity.

In addition, when the input rectangle IR only partially contains the closed edge, the most similar rectangle selection unit 12cb expands the input rectangle IR by a certain amount so that the closed edge can be detected within the input rectangle IR.

Therefore, according to the learning support device 10 of the embodiment, even if the input rectangle IR is not created accurately, it is possible to automatically correct it.

In the above embodiment, the similarity calculation information 11d, which is the basis for calculating the similarity, is selected in advance by the user U, for example, and can be set in advance as static information, but the setting may be dynamically changed. For example, based on the calculated similarity and the history of the user U's adoption results for the closed edge rectangles presented based on the similarity, the tendency of the adoption pattern of the user U may be learned by machine learning, and the learning results may be dynamically reflected in the similarity calculation information 11d. In addition, such dynamic reflection may be applied to the extraction algorithm information 11b.

Further advantages and modifications may readily occur to those skilled in the art. Therefore, the invention in its broader aspects is not limited to the specific details and representative embodiments shown and described above. Accordingly, various modifications may be made without departing from the spirit or scope of the general inventive concept as defined by the appended claims and equivalents thereof.

3 HMI section 3a Input section 3b Output section 10 Learning support device 11 Storage section 11a Target image DB
11b Extraction algorithm information 11c Edge information 11d Similarity calculation information 11e Rectangle information DB
12 Control unit 12a Image depiction unit 12aa Rectangle drawing unit 12b Image analysis unit 12ba Edge extraction unit 12c Rectangle creation unit 12ca Closed edge rectangle creation unit 12cb Most similar rectangle selection unit IR Input rectangle U User

Claims (7)

Extracting edges of objects in the target images for annotation in creating training data ,
A closed edge rectangle is created that is tangent to a closed edge that is an edge extracted as a closed curve from among the extracted edges;
Calculating a similarity between an input rectangle input by a user for the target image and the closed edge rectangle ;
Selecting the closed edge rectangle with the greatest similarity;
adopting the selected closed edge rectangle as a bounding box for the annotation;
A learning support device equipped with a controller . The controller :
Extracting edges of the object using an edge extraction algorithm selected by a user ;
The learning support device according to claim 1 . The controller :
a closed edge rectangle is created , the vertices of which are the coordinate positions of the four corners defined by the minimum horizontal position value, the maximum horizontal position value, the minimum vertical position value, and the maximum vertical position value of the closed edge;
The learning support device according to claim 1 or 2. The controller :
calculating the similarity between the input rectangle and the closed edge rectangle using at least one of an aspect ratio, a center position, or a width and a height as a comparison element ;
4. The learning support device according to claim 1, 2 or 3. The controller :
When the closed edge is only partially included in the input rectangle, the input rectangle is expanded by a certain amount so that the closed edge can be detected within the input rectangle .
The learning support device according to any one of claims 1 to 4. Extracting edges of objects in the target images for annotation in creating training data ,
A closed edge rectangle is created that is tangent to a closed edge that is an edge extracted as a closed curve from among the extracted edges;
Calculating a similarity between an input rectangle input by a user for the target image and the closed edge rectangle ;
Selecting the closed edge rectangle with the greatest similarity;
adopting the selected closed edge rectangle as a bounding box for the annotation;
A learning support method implemented by a controller . Extracting edges of objects in the target images for annotation in creating training data ,
A closed edge rectangle is created that is tangent to a closed edge that is an edge extracted as a closed curve from among the extracted edges;
Calculating a similarity between an input rectangle input by a user for the target image and the closed edge rectangle ;
Selecting the closed edge rectangle with the greatest similarity;
adopting the selected closed edge rectangle as a bounding box for the annotation;
A learning support program executed by the controller .

Images