Motif similarity calculation device and model learning device

Abstract

To enable the evaluation of a similarity level between a design and a motif.SOLUTION: A motif similarity level calculating device obtains, by an obtaining unit, an object image that represents an object. The motif similarity level calculating device inputs, by a motif similarity level calculating unit, the object image obtained by the obtaining unit to an already-learnt model to output a similarity level between the object image and a motif image that represents the motif, thereby obtaining the similarity level between the object image and the motif image.SELECTED DRAWING: Figure 3

Description

[Technical field]

[0001] The present invention relates to a motif similarity calculation device and a model learning device. [Background technology]

[0002] Conventionally, image data classification devices that automatically classify image data are known (for example, Patent Document 1). The image data classification device of Patent Document 1 extracts features of image data from the image data, and evaluates the relevance of the features extracted from multiple pieces of image data. The image data classification device then performs clustering on phenomena represented by the features of the multiple pieces of image data based on the evaluation result of the relevance, and classifies the multiple pieces of image data based on the clustering result.

[0003] Also, an image design support device that supports image design is known (for example, Patent Document 2). For example, the image design support device of Patent Document 2 calculates visual feature quantities such as average brightness, light-dark balance, average chromaticity, cool-warm balance, and average saturation for each image. Furthermore, the image design support device selects two visual features with a large difference in average visual feature quantities between the similar impression image group and the contrast impression image group, and displays each image of the similar impression image group and the contrast impression image group as a two-dimensional graph on the display unit.

[0004] Also, an efficient graphic classification and search technique that is robust against noise and local deformation is known (for example, Patent Document 3). The graphic classification method of Patent Document 3 inputs a graphic, extracts its structural features, applies transformation rules to the extracted structural features to synthesize structural features of a transformed graphic, calculates an index from each of the structural features, and classifies the graphic according to each of the structural features. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] JP 2001-256244 A [Patent Document 2] Japanese Patent Application Publication No. 10-188023 [Patent Document 3] Japanese Patent Application Publication No. 11-203461 Summary of the Invention [Problem to be solved by the invention]

[0006] By the way, motifs are often reflected in industrial and architectural designs, such as trees and flowers.

[0007] However, tree and flower motifs come in a variety of colors and shapes, making it difficult to evaluate the extent to which the motif is reflected in the design and the extent to which the design and the motif are similar.

[0008] In this regard, the techniques of Patent Documents 1 to 3 are methods for deductively evaluating or classifying the similarity of images based on preset feature amounts, etc. However, as described above, motifs of trees and flowers have various colors and shapes, and it is difficult to appropriately evaluate the similarity between a design and a motif using deductive methods such as those of Patent Documents 1 to 3.

[0009] Therefore, the techniques of the above Patent Documents 1 to 3 have a problem in that they are unable to appropriately evaluate the similarity between a design and a motif.

[0010] Furthermore, to incorporate a motif into the design of a building or industrial product, it is necessary to abstract the features of the motif and incorporate them into the shapes, dimensions, and layout of various components, rather than simply tracing the shape of the original motif. There are various methods of abstraction depending on which feature of the motif the designer focuses on, which component the designer chooses to incorporate, etc. For this reason, it has been difficult to evaluate whether the design adequately expresses the features of the motif.

[0011] In consideration of the above facts, the present invention aims to evaluate the similarity between a design and a motif by using a trained model generated inductively from training data. [Means for solving the problem]

[0012] In order to achieve the above object, the motif similarity calculation device of the present invention includes an acquisition unit that acquires an object image representing an object, and a motif similarity calculation unit that acquires the similarity between the object image and the motif image representing a motif by inputting the object image acquired by the acquisition unit into a trained model that outputs the similarity between the object image and the motif image. This makes it possible to appropriately evaluate the degree to which the design of an object, such as an industrial product or a building, is similar to a motif.

[0013] In addition, in the motif similarity calculation device, the object image and the motif image can be monochrome images, which makes it possible to evaluate the degree to which the object resembles the motif without taking into account the color of the object or motif, and as a result, it is possible to evaluate the similarity between the object and the motif while taking into account their shapes.

[0014] In addition, in the motif similarity calculation device, the object may be a building, and an object image in which the building is captured may be used. This allows appropriate evaluation of the degree to which the design of the building is similar to the motif.

[0015] The model learning device of the present invention includes a learning data acquisition unit that acquires learning data in which learning object images, which are images representing learning objects, are associated with learning motif images, which are images representing learning motifs, and a learning unit that generates a trained model for outputting a similarity between an object image representing an object and a motif image representing a motif by machine learning a learning model based on the learning data acquired by the learning data acquisition unit. This makes it possible to obtain a trained model for outputting a similarity between an image representing an object, such as an industrial product or a building, and an image representing the motif. Effect of the Invention

[0016] According to the present invention, the effect of evaluating the similarity between a design and a motif is achieved by utilizing a trained model that is inductively generated from training data. [Brief description of the drawings]

[0017] [Figure 1] A figure showing an example of an image of calculating the similarity between a motif and a building using a trained model trained by the method of an embodiment of the present invention. [Diagram 2] 1 is a block diagram showing a configuration of a model learning device according to an embodiment of the present invention. [Diagram 3] 1 is a block diagram showing a configuration of a motif similarity calculation device according to an embodiment of the present invention. [Figure 4] 4 is a flowchart showing a model learning process in the model learning device according to the embodiment of the present invention. [Diagram 5] 4 is a flowchart showing a motif similarity calculation process in the motif similarity calculation device according to the embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0018] [Embodiments of the present invention] Hereinafter, a model learning device and a motif similarity calculation device according to an embodiment of the present invention will be described with reference to the drawings. Although the model learning device and the motif similarity calculation device are configured as separate devices, they may be configured as an integrated device.

[0019] FIG. 1 is a diagram showing an example of an image of calculation of the similarity between a motif and a building by a trained model trained by the method of the embodiment of the present invention. As shown in FIG. 1, the degree of similarity of each architectural design image of a building, which is an object image, to a tree motif is output as a score of similarity between the motif image and the architectural design. This allows the designer to understand how similar the architectural design being designed is to the motif. Note that the similarity between the motif image may be the similarity between the motif itself (i.e., the entire group of motif images), or a motif image included in the motif (a part of all the motif images) may be specified to be the similarity between the specified motif image and the architectural design.

[0020] Fig. 2 is a block diagram showing the configuration of a model learning device according to an embodiment of the present invention. As shown in Fig. 2, the model learning device 100 functionally comprises a learning data acquisition unit 110, a learning unit 112, and a model storage unit 114. The model learning device 100 is realized by a computer including a CPU (Central Processing Unit), a ROM (Read Only Memory) that stores programs for implementing each processing routine, a RAM (Random Access Memory) that temporarily stores data, a memory as a storage means, a network interface, and the like.

[0021] The learning data acquisition unit 110 acquires learning data used for learning. The learning data is data in which learning object images and learning motif images are associated with each other. The learning object images are images that represent the learning object, and when the object is a building, a group of architectural design images are used. The object is not limited to buildings, and general structures having design are the target. Therefore, in addition to buildings, design images of industrial products (objects, tools, tableware, containers, etc.) can be used as learning object images. The learning motif images are images that represent motifs for learning. For example, if the motif is a tree, a group of tree images is used, and if the motif is a lotus flower, a group of lotus flower images is used.

[0022] Here, the learning object image and the learning motif image in the learning data may be monochrome images with monochrome gradation. By making the images monochrome, the learning data can be obtained from which information such as color scheme, color, and color tone is removed, and therefore a model capable of evaluating similarity focusing on shape properties can be learned. By using such learning data, a trained model capable of evaluating similarity with emphasis on shape patterns and structural patterns between the motif and the object can be obtained. The conversion to monochrome images may be performed by the learning data acquisition unit 110, or the learning object image and the learning motif image may be acquired by converting the learning object image and the learning motif image to monochrome images in advance externally before acquisition.

[0023] The learning unit 112 generates a trained model for outputting a similarity between an object image representing an object and a motif image representing a motif by machine learning the learning model based on the learning data acquired by the learning data acquisition unit 110. An arbitrary neural network is used for the learning model. As a machine learning method, a method of training parameters of a neural network by supervised machine learning and generating a trained neural network may be used.

[0024] In addition, during learning, element information on geometric properties or structural properties to be emphasized (or not to be emphasized) and the designation of the importance of the element information may be accepted to weight the element information of the motif. For example, in the case of a lotus flower motif, the element information is assumed to be the ratio of the petals and the stigma to the whole flower, the arrangement of the petals and the stigma, structural properties such as the number of petals, and geometric properties such as the shape of the petals and the shape of the stigma. The learning unit 112 performs learning by including such designation of element information and the importance of the element information in the learning data. The designation of element information may be accepted from the predetermined user interface provided in the model learning device 100 or an external device that can be operated by the user. The importance of element information to be emphasized may be set high, and that of element information to be not emphasized may be set low. This makes it possible to learn a trained model capable of evaluation reflecting the similarity of specific element information of the motif. In addition, by combining this with a trained model that has been trained without weighting the elements and comparing the similarity results, it is possible to quantitatively evaluate the similarity of the specific elements that the subject wants to represent through the motif.

[0025] Next, a motif similarity calculation device will be described. FIG. 3 is a block diagram showing the configuration of a motif similarity calculation device according to an embodiment of the present invention. As shown in FIG. 3, the motif similarity calculation device 200 functionally comprises a model storage unit 210, an acquisition unit 212, and a motif similarity calculation unit 214. The motif similarity calculation device 200 is realized by a computer including a CPU, a ROM storing programs for implementing each processing routine, a RAM for temporarily storing data, a memory as a storage means, and a network interface. In addition, the motif similarity calculation device 200 is provided with a predetermined user interface that allows the display of the results of the similarity calculation and the user to operate it.

[0026] The model storage unit 210 stores a trained model generated by the model learning device 100. The trained model is trained as a model for outputting a similarity between an object image and a motif image representing a motif.

[0027] The acquisition unit 212 acquires an object image representing an object. The object image is, for example, an architectural design image of a building. The object may also be an object having other designs, as in the case of the model learning device 100. The object image may also be a monochrome image.

[0028] The motif similarity calculation unit 214 inputs the object image acquired by the acquisition unit 212 into the trained model stored in the model memory unit 210, and acquires the similarity between the object image and the motif image based on the output from the trained model.

[0029] In addition, it is also assumed that the designer may incorporate multiple motifs into the architectural design. In this case, a trained model may be trained to output similarities with multiple motifs and output the ratio of similarities between the motifs as a score in the learning and similarity calculation. The model learning device 100 performs machine learning of the learning model using learning motif images of the multiple motifs as learning data. In the learning, the trained model is trained to output the similarities of each motif and a ratio indicating which of the motifs is closer. The motif similarity calculation device 200 inputs an object image to the trained model and outputs the similarities for each of the multiple motifs and the ratio between the motifs. For example, for two motifs, a lotus flower motif and a tree motif, the score is output as a lotus flower score: 0.3, the tree score: 0.1, etc. Then, together with the score, the ratio may be output as a lotus flower motif ratio of 70%, a tree motif ratio of 30%, etc. This allows the designer to grasp the extent to which each motif is distributed in the architectural design he is designing.

[0030] Next, the operation of the embodiment of the present invention will be described. First, the operation of the model learning device 100 will be described. Fig. 4 is a flowchart showing the model learning process in the model learning device 100 according to the embodiment of the present invention. The CPU of the model learning device 100 reads out and executes programs and various data from the ROM, and the CPU executes the processes of each part of the model learning device 100.

[0031] In step S100, the learning data acquisition unit 110 acquires learning data to be used for learning. The learning data is data in which learning object images and learning motif images are associated with each other.

[0032] In step S102, the learning unit 112 generates a trained model for outputting the similarity between an object image representing an object and a motif image representing a motif by machine learning a learning model based on the learning data acquired in step S100.

[0033] In step S104, the learning unit 112 stores the trained model trained in step S102 in the model storage unit 114. This completes the model training process.

[0034] As described above, according to the model learning device 100 of an embodiment of the present invention, it is possible to obtain a trained model for outputting the similarity between an image representing an object such as an industrial product or a building, and an image representing a motif.

[0035] Next, a description will be given of the operation of the motif similarity calculation device 200. Fig. 5 is a flowchart showing the motif similarity calculation process in the motif similarity calculation device 200 according to an embodiment of the present invention. The CPU of the motif similarity calculation device 200 reads out and executes programs and various data from the ROM, and the CPU executes the processes of each part of the motif similarity calculation device 200.

[0036] In step S200, the acquisition unit 212 acquires an object image representing an object.

[0037] In step S202, the motif similarity calculation unit 214 inputs the object image obtained in step S200 into the trained model stored in the model memory unit 210, and obtains the similarity between the object image and the motif image based on the output from the trained model.

[0038] In step S204, the motif similarity calculation unit 214 displays the similarity between the object image and the motif image acquired in step S202 on a predetermined user interface. The above is the motif similarity calculation process.

[0039] As described above, according to the motif similarity calculation device 200 of an embodiment of the present invention, it is possible to evaluate the similarity between a design and a motif by utilizing a trained model that is inductively generated from training data.

[0040] The present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the spirit and scope of the present invention. [Explanation of symbols]

[0041] 100 Model Learning Device 110 Learning data acquisition unit 112 Learning Department 114 Model Memory Section 200 Motif Similarity Calculator 210 Model Memory Unit 212 Acquisition Department 214 Motif Similarity Calculation Unit

Claims

1. an acquisition unit for acquiring an object image representing an object; a motif similarity calculation unit that inputs the object image acquired by the acquisition unit into a trained model for outputting a similarity between the object image and a motif image representing a motif and a ratio indicating which of a plurality of motifs the object image is closer to, the trained model being trained using element information on the shape and structural properties of the motif and the importance of the element information as training data, thereby acquiring a similarity of element information between the object image and the motif image and a ratio indicating which of a plurality of motifs the object image is closer to; A motif similarity calculation device comprising:

2. The object image and the motif image are monochrome images. The motif similarity calculation device according to claim 1 .

3. The object is a building, and an object image in which the building is captured is used. The motif similarity calculation device according to claim 1 or 2.

4. a learning data acquisition unit that acquires learning data including a learning object image, which is an image representing an object for learning, and a learning motif image, which is an image representing a motif for learning, associated with each other, and element information regarding shape properties and structural properties of the motif and the importance degree of the element information; a learning unit that generates a trained model by machine learning a learning model based on the learning data acquired by the learning data acquisition unit, and when an object image representing an object is input as an input, outputs a similarity of element information between the object image and a motif image representing a motif and a ratio indicating which of a plurality of motifs the object image is closer to; A model learning device comprising:

Images