Identification system, learning apparatus, identification apparatus, learning method, identification method, and program

Abstract

To identify the internal state of an object from its appearance.SOLUTION: An identification system comprises: a training image acquisition unit for acquiring a training object image obtained by capturing a training object, with as the object, a matter which may, during a growth process, develop, in its interior, a space or an area having a density different from surroundings, and as an identification target, an internal state which includes whether or not the space or the area having a density different from the surroundings exists in the object; a training unit for generating a trained model that identifies the internal state from an image by learning the correspondence between the object and the internal state using a data set in which the training object image is associated with a label indicating the internal state; an identification image acquisition unit for acquiring an identification object image obtained by capturing the identification object; an identification unit for identifying the internal state of the identification object from the identification object image using the trained model; and an output unit for outputting an identification result obtained by the identification unit.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to an identification system, a learning device, an identification device, a learning method, an identification method, and a program.

Background Art

[0002] There is a method for non-destructively determining the quality of fresh fruits such as apples (for example, Patent Document 1). By irradiating fresh fruits with light rays such as infrared rays and estimating the sugar content of the fresh fruits based on the detection results of the light rays transmitted or reflected by the fresh fruits, the quality can be determined without destroying the fresh fruits.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, with the technology of Patent Document 1, it is difficult to identify the internal state from the appearance. For example, among apples as fruits, there are some in which a space called "hollow crack" or a region having a different density from the periphery is generated inside. It is difficult to visually determine whether there is a "hollow crack" inside from the appearance of an apple, and it becomes known only after cutting the apple in half. For this reason, when a "hollow crack" is discovered at the stage when a consumer tries to eat an apple after shipment, it often becomes a factor that reduces the commercial value even if there is no problem with the taste.

[0005] The present invention has been made in view of such a situation, and an object thereof is to provide an identification system, a learning device, an identification device, a learning method, an identification method, and a program capable of identifying the internal state from the appearance of an object.

Means for Solving the Problems

[0006] The identification system of the present invention targets an object in which a region having a density different from that of the inside or the periphery can occur during the growth process, and identifies an internal state including whether or not there is a region having a density different from that of the inside or the periphery inside the object as an object to be identified. A training image acquisition unit that acquires a training target image obtained by imaging the training target object, and a data set in which a label indicating the internal state is associated with the training target image are used to learn the correspondence between the object and the internal state, thereby generating a trained model that identifies the internal state from an image. A training unit, an identification image acquisition unit that acquires an identification target image obtained by imaging the identification target object, an identification unit that uses the trained model to identify the internal state in the identification target object from the identification target image, and an output unit that outputs an identification result by the identification unit.

[0007] The learning device of the present invention targets an object in which a region having a density different from that of the inside or the periphery can occur during the growth process, and identifies an internal state including whether or not there is a region having a density different from that of the inside or the periphery inside the object as an object to be identified. A training image acquisition unit that acquires a training target image obtained by imaging the training target object, and a data set in which a label indicating the internal state is associated with the training target image are used to learn the correspondence between the object and the internal state, thereby generating a trained model that identifies the internal state from an image. A training unit.

[0008] The identification device of the present invention targets an object in which a region having a density different from that of the interior or the periphery can occur during the growth process, and identifies an internal state including whether there is a region having a density different from that of the interior or the periphery in the object as an object to be identified. The identification device includes an identification image acquisition unit that acquires an identification target image obtained by imaging the object for identification, an identification unit that uses a trained model to identify the internal state of the object for identification from the identification target image, and an output unit that outputs an identification result by the identification unit. The trained model is a model that trains the internal state from an image, which is generated by learning the correspondence between the object and the internal state using a learning dataset in which a label indicating the internal state is associated with a training target image obtained by imaging the object for training.

[0009] The identification method of the present invention is an identification method performed by a computer. An object in which a region having a density different from that of the interior or the periphery can occur during the growth process is targeted, and an internal state including whether there is a density different from that of the interior or the periphery in the object is identified as an object to be identified. A training image acquisition unit acquires a training target image obtained by imaging the object for training, a training unit generates a trained model that identifies the internal state from an image by learning the correspondence between the object and the internal state using a dataset in which a label indicating the internal state is associated with the training target image, an identification image acquisition unit acquires an identification target image obtained by imaging the object for identification, an identification unit uses the trained model to identify the internal state of the object for identification from the identification target image, and an output unit outputs an identification result by the identification unit.

[0010] The learning method of the present invention is a learning method performed by a computer. An object in which a region having a density different from that of the interior or the periphery can occur in the growth process is used as an object. An internal state including whether or not there is a region having a density different from that of the interior or the periphery in the object is used as an object to be identified. A training image acquisition unit acquires a training target image in which the training target object is imaged. A training unit uses a data set in which a label indicating the internal state is associated with the training target image to learn the correspondence between the object and the internal state, thereby generating a trained model for identifying the internal state from an image.

[0011] The identification method of the present invention is an identification method performed by a computer. An object in which a region having a density different from that of the interior or the periphery can occur in the growth process is used as an object. An internal state including whether or not there is a region having a density different from that of the interior or the periphery in the object is used as an object to be identified. An identification image acquisition unit acquires an identification target image in which the identification target object is imaged. An identification unit uses a trained model to identify the internal state in the identification target object from the identification target image. An output unit outputs the identification result by the identification unit. The trained model is a model for training the internal state from an image, which is generated by learning the correspondence between the object and the internal state using a learning data set in which a label indicating the internal state is associated with a training target image in which the training target object is imaged.

[0012] The program of the present invention is a program for operating a computer as the learning device described above, and is a program for causing the computer to function as each part included in the learning device.

[0013] The program of the present invention is a program for operating a computer as the identification device described in claim 9, and is a program for causing the computer to function as each part included in the identification device.

Advantages of the Invention

[0014] According to the present invention, the internal state can be identified from the appearance of the object.

Brief Description of the Drawings

[0015]

Figure 1

Figure 2

Figure 3A

Figure 3B

Figure 3C

Figure 4

Figure 5

Figure 6A

Figure 6B

Figure 7

Figure 8

Figure 9

Figure 10

Mode for Carrying Out the Invention

[0016] Hereinafter, embodiments of the present invention will be described with reference to the drawings. Hereinafter, a case will be exemplified in which an apple as an object is imaged to identify whether there is "cracking" inside the apple as an object to be identified, but the present invention is not limited to this. This embodiment can be applied to any object having a space or a region with a density different from that of the periphery inside during the growth process. For example, it can be applied to fruits of the Rosaceae family such as strawberries, figs, cherries, pears, peaches, citrus fruits such as oranges, fruits of the Cucurbitaceae family such as melons, actual vegetables such as tomatoes, eggplants, and cucumbers, root vegetables such as daikon radishes and burdocks, and tubers. In addition, it can be applied to an object in which a region having a space or a density different from that of the periphery can occur, for example, detection of the state of fish eggs inside a fish in the process of identifying the quality of the fish, whether there are air bubbles in the solder in the process of soldering electronic components etc. on a substrate, and detection of whether the inside of the plastic is in a non-uniform state.

[0017] (Regarding the identification system 1) FIG. 1 is a block diagram showing a configuration example of the identification system 1 in the embodiment. The identification system 1 is a system for identifying whether there is "cracking" inside an apple. As shown in FIG. 1, the identification system 1 includes, for example, an information processing device 10, an imaging device 20, and a display device 30.

[0018] The information processing device 10 and the imaging device 20, and the information processing device 10 and the display device 30 are communicably connected via a communication network such as the Internet, or a short-range wireless communication method such as Bluetooth (registered trademark) or infrared communication, or a wired connection such as a USB cable.

[0019] The imaging device 20 is a camera that images an apple as the object T. The imaging device 20 images the apple and transmits the image information of the captured image to the information processing device 10.

[0020] The information processing apparatus 10 is a computer that identifies whether there is a "vine crack" inside an apple. The information processing apparatus 10 receives the image information of the image of the apple captured by the imaging apparatus 20, and uses the received image information to determine whether there is a "vine crack" inside the apple. The information processing apparatus 10 transmits the identification result to the display apparatus 30.

[0021] The display apparatus 30 is a display that displays images and the like. The display apparatus 30 receives the identification result of whether there is a "vine crack" inside the apple from the information processing apparatus 10, and displays the identification result.

[0022] (Regarding the information processing apparatus 10) As shown in FIG. 1, the information processing apparatus 10 includes, for example, an imaging control unit 11, an image acquisition unit 12, a binarized image generation unit 13, a training unit 14, an identification unit 15, an output unit 16, a dataset storage unit 17, and a trained model storage unit 18.

[0023] The imaging control unit 11 controls the imaging of the apple. FIG. 2 is a diagram showing an example of an imaging environment for imaging an apple as the object T in the embodiment. As shown in FIG. 2, the imaging environment is provided with a light-shielding part SL which is a housing formed of a member that blocks light. The light-shielding part SL is provided with a placement part P and an irradiation part L. An apple as the object T is placed on the placement part P. The irradiation part L irradiates the object T with light, for example, an LED (light-emitting diode). The imaging apparatus 20 images the object T from the direction opposite to the direction where the irradiation position of the light irradiated from the irradiation part L is located, with reference to the installation position of the object T. In the example of this figure, the irradiation part L is installed below the placement part P, and irradiates light upward toward the object T. The imaging apparatus 20 is installed above the object T so that the imaging direction is downward, and images the downward direction where the object T is placed. The imaging control unit 11 controls the irradiation timing of the irradiation part L and the imaging timing of the imaging apparatus 20 so that the object T is imaged in a state where the object T is irradiated with light by the irradiation part L in the imaging environment as shown in FIG. 2. For example, the imaging control unit 11 causes the imaging device to image the object T from the side with the fruit stalk (kakou), that is, the side where the "tsuru" is present, while irradiating light on the side opposite to the side where the fruit stalk is connected to the branch.

[0024] Returning to FIG. 1, the image acquisition unit 12 acquires the image information of the image in which the object T is imaged. When the object T is imaged by the imaging device 20, the image acquisition unit 12 acquires the image information of the image in which the object T is imaged from the imaging device 20. The image information is information in which a pixel value (for example, an RGB value) indicating a color is associated with each pixel. The image acquisition unit 12 outputs the acquired image information of the image to the binarization image generation unit 13.

[0025] The binarization image generation unit 13 generates a binarized image obtained by binarizing the image in which the object T is imaged. The binarization image generation unit 13 converts a pixel value equal to or greater than a threshold value into a specific pixel value (first pixel value) and converts a pixel value less than the threshold value into a specific pixel value (second pixel value) different from the first pixel value based on the image information of the image in which the object T is imaged, thereby generating a binarized image. The binarization image generation unit 13 generates a binarized image using any method used as an existing technique, for example, Otsu's binarization. The binarization image generation unit 13 outputs the generated binarized image to the training unit 14.

[0026] FIG. 3 (FIGS. 3A to 3C) is a diagram showing an example of an image obtained by imaging an object. FIG. 3A shows an example of an image in which an apple as the object T is imaged in a state where no light is irradiated from the irradiation unit L. FIG. 3B shows an example of an image in which an apple as the object T is imaged in a state where light is irradiated from the irradiation unit L.

[0027] Above Figure 3C, as (1), an image of an apple with vine crack (AVC) is shown. On the left, as (a), an example of an image of an apple with vine crack imaged in a state where no light is irradiated from the irradiation unit L is shown. In the center, as (b), a state of a cross-section of an apple with vine crack divided into two is shown. On the right, as (c), an example of a binarized image generated from an image of an apple with vine crack imaged in a state where light is irradiated from the irradiation unit L is shown.

[0028] Below Figure 3C, as (2), an image of a normal apple (NOA) without vine crack is shown. On the left, as (a), an example of an image of a normal apple imaged in a state where no light is irradiated from the irradiation unit L is shown. In the center, as (b), a state of a cross-section of a normal apple divided into two is shown. On the right, as (c), an example of a binarized image generated from an image of a normal apple imaged in a state where light is irradiated from the irradiation unit L is shown.

[0029] Returning to Figure 1, the training unit 14 generates a trained model that identifies the internal state of an apple from an image in which the apple is imaged by training the learning model.

[0030] First, the training unit 14 generates a data set for training the learning model. A method for the training unit 14 to generate a data set will be described with reference to Figure 4. Figure 4 is a flowchart for explaining the processing flow performed by the information processing apparatus 10 of the embodiment.

[0031] The imaging control unit 11 of the information processing apparatus 10 controls the imaging apparatus 20 to image the object T (step S10). The imaging control unit 11 places an apple as the object T on the placement unit P, irradiates the object T with light from the irradiation unit L, and causes the imaging apparatus 20 to image the apple illuminated by the light. The image acquisition unit 12 of the information processing apparatus 10 acquires an image in which the object T illuminated by light is imaged (step S11). The information processing apparatus 10 acquires a determination result of determining the presence or absence of "vine cracking" in an apple as the object T. For example, after the apple as the object T is imaged, the person in charge cuts the apple in half and determines whether there is "vine cracking" inside. The inspector inputs the determination result into the information processing apparatus 10 by operating a keyboard or the like. Thereby, the information processing apparatus 10 acquires the determination result of determining the presence or absence of "vine cracking". Also, the binarization image generation unit 13 of the information processing apparatus 10 generates a binarization image (step S13). The training unit 14 of the information processing apparatus 10 generates a data set in which the determination result of determining the presence or absence of "vine cracking" as a label is attached to the binarization image (step S14). The training unit 14 stores the generated data set in the data set storage unit 17 (step S15). The training unit 14 determines whether to end the generation of the data set (step S16). For example, the training unit 14 determines to end the generation of the data set when both the number of data sets with "vine cracking" and the number of data sets without "vine cracking" reach a predetermined number set in advance. On the other hand, for example, when at least one of the number of data sets with "vine cracking" or the number of data sets without "vine cracking" has not reached the predetermined number set in advance, the training unit 14 determines not to end the generation of the data set. Here, the predetermined number set in advance may be arbitrarily set according to the situation of the object T and the object to be identified. Also, the predetermined number set in advance may be the same number regardless of the presence or absence of "vine cracking", or may be different numbers according to the presence or absence of "vine cracking". When the training unit 14 determines not to end the generation of the data set, it returns to step S10 to continue the generation of the data set. When it determines to end the generation of the data set, the training unit 14 ends the generation of the data set.

[0032] Next, the training unit 14 generates a trained model using the dataset. A method for the training unit 14 to generate a trained model will be described with reference to FIG. 5. FIG. 5 is a flowchart for explaining the processing flow performed by the information processing apparatus 10 according to the embodiment.

[0033] The training unit 14 of the information processing apparatus 10 divides the dataset into a training dataset and a verification dataset (step S20). The training unit 14 sets a predetermined ratio for the dataset, for example, such that the ratio of the training dataset to the entire dataset is equal to or greater than a predetermined value (for example, 0.7). The training unit 14 performs principal component analysis using the binarized images in the training dataset (step S21). The training unit 14 sets the number of principal components to be extracted by the principal component analysis such that the cumulative contribution rate exceeds a predetermined ratio (for example, 0.8). The training unit 14 generates a quantum kernel using the principal components extracted in step S21 (from the first principal component to the tenth principal component) as feature amounts (step S22). As a result, a quantum kernel with the principal components as feature amounts is generated. The training unit 14 embeds the quantum kernel into an SVM (support vector machine) and constructs a learning model using the training data (step S23). The training unit 14 determines whether the trained SVM satisfies the generation conditions as a trained model (step S24). For example, the training unit 14 determines whether the learning model satisfies the reference value based on evaluation indices such as accuracy, recall rate, and F1 score (F value), and determines whether the generation conditions as a trained model are satisfied. For example, when the F1 score is used as the evaluation index, it is determined that the generation conditions are satisfied when the F1 score is generally between 0.85 and 0.9. For example, when the number of data sets trained by the training unit 14 using SVM reaches a predetermined number, the training unit 14 determines whether the trained SVM satisfies the generation conditions as a trained model. The training unit 14 inputs the binarized image in the verification data set to the trained SVM, and causes the trained SVM to identify whether there is "vine cracking" inside the object T (apple) corresponding to the binarized image, and classifies whether the identification result matches the label in the data set. When the evaluation index (accuracy, recall rate, F1 score (F value), etc.) based on the classification is equal to or higher than the threshold value, the training unit 14 determines that the trained SVM satisfies the generation conditions as a trained model. On the other hand, when the evaluation index is less than the threshold value, the training unit 14 determines that the trained SVM does not satisfy the generation conditions as a trained model. When the trained SVM satisfies the generation conditions as a trained model, the training unit 14 stores the trained SVM in the trained model storage unit 18 as a trained model (step S25).

[0034] Note that in step S22, the training unit 14 only needs to be able to map to a high-dimensional feature space (feature space after storage) having more dimensions than the dimension of the feature space (feature space before mapping) corresponding to the feature amount using the kernel method. The training unit 14 may perform the mapping to the second feature space using a classical kernel instead of the quantum kernel. In this way, the training unit 14 generates a trained model by performing the processes shown in steps S21 to S25.

[0035] The identification unit 15 uses the trained model to identify whether there is "vine cracking" inside the object T (apple). The identification unit 15 inputs the binarized image of the verification data set to the trained model. The trained model identifies whether there is "vine cracking" inside the object T (apple) imaged in the input binarized image based on the input binarized image, and outputs the identification result. The identification unit 15 uses the identification result output from the trained model as the identification result of whether there is "vine cracking" inside the object T (apple).

[0036] The output unit 16 outputs the identification result by the identification unit 15. The output unit 16 outputs the identification result to the display device 30. Thereby, the display device 30 displays the identification result. The dataset storage unit 17 stores the dataset generated by the training unit 14. The trained model storage unit 18 stores the trained model generated by the training unit 14.

[0037] The storage unit (including the dataset storage unit 17 and the trained model storage unit 18) provided in the information processing apparatus 10 is composed of a storage medium such as an HDD (Hard Disk Drive), a flash memory, an EEPROM (Electrically Erasable Programmable Read Only Memory), a RAM (Random Access read / write Memory), a ROM (Read Only Memory), or a combination thereof. The storage unit provided in the information processing apparatus 10 stores a program for executing various processes for realizing the functions of the information processing apparatus 10 and temporary data used when performing various processes.

[0038] The functional units (including the imaging control unit 11, the image acquisition unit 12, the binarized image generation unit 13, the training unit 14, the identification unit 15, and the output unit 16) provided in the information processing apparatus 10 are realized by causing the CPU (Central Processing Unit) and / or GPU (Graphics Processing Unit) provided as hardware in the information processing apparatus 10 to execute the program stored in the storage unit provided in the information processing apparatus 10.

[0039] Here, the quantum kernel QK used in the embodiment will be described with reference to FIG. 6 (FIGS. 6A and 6B). FIG. 6 is a diagram showing an example of a quantum circuit for generating a quantum kernel in the embodiment. FIG. 6A shows an example of a basic configuration of a quantum circuit for generating a quantum kernel QK. The quantum circuit is a circuit that takes quantum bits QBT (qubits) as inputs and outputs data to classical bits CBT (bits) after measurement. The initial state of the quantum bits QBT (Qubits) is set to "|0>". In the quantum circuit, the quantum bits set to the initial state are prepared with a state vector encoded by the matrix S(x i )|0> in the data encoding circuit. To create an inner product, the encoding circuit matrix S(x i ) † is prepared, and the inner product S(x i )·S(x i ) † is obtained. Here, the encoding circuit matrix S is a matrix whose elements are the principal components (features) obtained from the binary image. The dagger S † of the encoding circuit matrix is a matrix (dagger matrix) corresponding to the complex conjugate transpose of the matrix S in the unitary matrix. The matrix S(x i ) is a matrix with the principal component (feature) x i obtained from a certain binary image as its elements. The encoding circuit matrix S, and the control rotation gate Ry provided in S † is a rotation gate with the Y-axis as the rotation axis. Here, the computational basis is the Z-axis, and the projection component in the Z-axis direction is measured. The kernel corresponding to each element of the quantum bits QBT is observed in the measurement circuit MS with the computational basis as the Z-axis and the projection component in the Z-axis direction. In the measurement circuit MS, it is converted into classical bits CBT (bits) with a value of 0 or 1 according to the probability.

[0040] FIG. 6B shows the configuration of a quantum circuit for generating each of five types of quantum kernels QK (quantum kernels QK1 to QK5) as an example of the quantum kernel used in the process of the training unit 14 generating a trained model. In the example of this figure, only the configuration of the feature matrix U in the quantum circuit is shown, and the description of the dagger matrix U † corresponding to the Hermitian conjugate is omitted.

[0041] The quantum circuit for generating the quantum kernel QK1 is a circuit in which a controlled rotation gate Ry is provided for each of the qubits QBT. The quantum circuit for generating the quantum kernel QK2 is a circuit in which a controlled rotation gate Ry is provided for the 0 (zero) -th qubit QBT, and a CNOT gate (Controlled NOT gate, XNOR gate) is provided for each of the qubits QBT after the 1 -st qubit. The quantum circuit for generating the quantum kernel QK3 is a circuit in which a controlled rotation gate Ry and a CNOT gate are provided for each of the qubits QBT. The quantum circuit for generating the quantum kernel QK4 is a circuit in which a Hadamard gate H and a controlled rotation gate Ry are provided for each of the qubits QBT. The quantum kernel QK4 is a circuit in which quantum entanglement is formed. The quantum circuit for generating the quantum kernel QK5 is a circuit in which a Hadamard gate H and a controlled rotation gate Rx are provided for each of the qubits QBT. This quantum kernel QK5 is also a circuit in which quantum entanglement is used.

[0042] Here, the effect of discrimination using the quantum kernels used in the embodiment will be described with reference to FIGS. 7 to 10. FIGS. 7 to 10 are diagrams for explaining the effects in the embodiment.

[0043] FIG. 7 shows the relationship between the training size and the evaluation index. The horizontal axis of FIG. 7 represents the training size, and the vertical axis represents the evaluation index. Here, the training size is the number of data sets used to train the trained model. In FIG. 7, for comparison, in addition to the quantum kernel QK1, the relationship between the training size and the evaluation index in each of the two classical kernels CK (classical kernel CK(linear) and classical kernel CK(RBF)) is shown. The classical kernel CK(linear) adopts a linear function as the kernel function used in the kernel method. The quantum kernel QK1 adopts the quantum circuit corresponding to the quantum kernel QK1 in FIG. 6B as the kernel function used in the kernel method. The classical kernel CK(RBF) adopts a radial basis function as the kernel function used in the kernel method.

[0044] In FIG. 7, as the evaluation indexes, two types of indexes, the accuracy rate (Ac, Accuracy), and the F value (F1-score) are shown.

[0045] The accuracy rate Ac is calculated by the following formula (1). Ac = (TP + TN) / (TP + TN + FP + FN) …(1) Here, TP is the true positive, which is the number of object Ts correctly identified as having "vine cracks" for the object T with "vine cracks". TN is the true negative, which is the number of object Ts correctly identified as not having "vine cracks" for the object T without "vine cracks". FP is the false positive, which is the number of object Ts misidentified as not having "vine cracks" for the object T with "vine cracks". FN is the false negative, which is the number of object Ts misidentified as having "vine cracks" for the object T without "vine cracks".

[0046] The F value F1 is calculated by the following formula (2). F1 = 2×(precision×recall) / (precision + recall) …(2) However, Precision is the precision rate, an indicator represented by TP / (TP+FP). Recall is the recall rate, an indicator represented by TP / (TP+FN).

[0047] As shown in FIG. 7, in the case of the quantum kernel QK1, regardless of the training size, both evaluation indicators (accuracy rate Ac and F-value F1) exceed 0.7. This indicates that the trained model using the quantum kernel QK1 can accurately identify with relatively high accuracy regardless of the training size. In contrast, for the classical kernel CK, the evaluation indicator tends to increase as the training size increases. In the classical kernel CK (linear), the evaluation indicator for the case where the training size is 48 is about 0.6. In the classical kernel CK (RBF), in the case where the training size is 30, the evaluation indicator is about 0.68, and in the case where the training size is 48, the evaluation indicator is about 0.75. This suggests that as the training size increases, the kernel trick may function effectively in the classical kernel CK (RBF).

[0048] FIG. 8 shows the evaluation indicators in the trained models generated using each of a plurality of different kernels (classical kernel CK (RBF), quantum kernels QK1 to QK5). The horizontal axis in FIG. 8 represents the kernel, and the vertical axis represents the evaluation indicator. FIG. 8 shows two types of indicators, the accuracy rate Ac and the F-value F1, as the evaluation indicators. In the trained models shown in FIG. 8, the conditions for training each trained model are unified to the same conditions, that is, the training size is 48 and the number of principal components used as features is 10, that is, in principal component analysis, the first principal component to the tenth principal component are extracted, except that the kernels are different. As shown in FIG. 8, when the training size is 48 and the number of principal components is 10, the evaluation indicators (accuracy rate Ac and F-value F1) tend to be higher in the order of quantum kernels QK4, QK3, and QK1. For all of these quantum kernels QK4, QK3, and QK1, both evaluation indicators exceed 0.8.

[0049] FIG. 9 shows the relationship between the number of principal components used as features and the evaluation indicators for different kernels (classical kernel CK(RBF), quantum kernels QK1, and QK4). The first axis (lower side) of the horizontal axis in FIG. 9 represents the number of principal components, the second axis (upper side) of the horizontal axis represents the cumulative contribution rate, and the vertical axis represents the evaluation indicators. FIG. 9 shows two types of evaluation indicators, the accuracy rate Ac and the F-value F1. As shown in FIG. 9, as the number of principal components decreases, the evaluation indicators (accuracy rate Ac and F-value F1) of the classical kernel CK(RBF) and the quantum kernel QK1 tend to decrease. This is considered to suggest that the kernel trick works more effectively as the number of principal components increases. On the other hand, the evaluation index of the quantum kernel QK4 with quantum entanglement shows a tendency to increase as the number of principal components decreases. In the trained model using the quantum kernel QK4, the evaluation indicators when extracting up to the third principal component exceed 0.9. This is considered to be because the quantum entanglement maps the feature space to a clear discrimination boundary based on features corresponding to the cumulative contribution rate of 0.63 up to the third principal component, which are characteristic principal components for identifying the presence or absence of "vine cracking".

[0050] Figure 10 shows the binarized image of the object T and its identification result. In each image, predict as the predicted value and true as the true value are shown in binary (0 (zero) or 1). When the predicted value (predict) t is 0 (zero), it is identified that there is no "vine crack" inside the object T, and when the predicted value (predict) is 1, it indicates that there is a "vine crack" inside the object T. When the true value (true) is 0 (zero), it shows that there is no "vine crack" inside the object T, and when the true value (true) is 1, it shows that there is a "vine crack" inside the object T. As shown in Figure 10, in most of the binarized images, good identification results in which the predicted value (predict) and the true value (true) match are obtained.

[0051] As described above, the identification system 1 of the embodiment includes an image acquisition unit 12, a training unit 14, an identification unit 15, and an output unit 16. The identification system 1 uses an object in which a space can be generated inside during the growth process as the object T. The internal state including whether there is a space inside the object T is the identification target. The image acquisition unit 12 acquires a training target image in which the training object T is imaged. In this case, the image acquisition unit 12 is an example of a "training image acquisition unit". The training unit 14 generates a trained model. The trained model is a model that identifies the internal state of the object T imaged in the image from the image. The training unit 14 uses a dataset in which a label indicating the internal state of the object T imaged in the training target image is associated with the training target image, and generates a trained model by learning the correspondence between the object T and its internal state. The image acquisition unit 12 acquires an identification target image in which the identification object T is imaged. In this case, the image acquisition unit 12 is an example of an "identification image acquisition unit". The identification unit 15 uses the trained model to identify the internal state of the identification object T in the identification target image. The output unit 16 outputs the identification result by the identification unit 15. Thus, in the identification system 1 of the embodiment, since a trained model for identifying the internal state of the object T imaged in the image can be generated from the image, the internal state of the object T can be identified from the identification image showing the appearance of the object T using the trained model.

[0052] Also, in the identification system 1 of the embodiment, the training unit 14 uses the kernel method to map the feature amounts (for example, the first to tenth principal components extracted by principal component analysis) extracted from the training target image to a feature space of a higher dimension than the number of dimensions that the feature amounts have, and generates feature amounts. The training unit 14 generates a trained model by learning an identification boundary for classifying the mapped feature amounts into a plurality of clusters. The training unit 14 embeds a quantum kernel into an SVM and constructs a learning model from the training data. For example, the training unit 14 acquires a Gram matrix from the quantum kernel circuit generated in step S22. The training unit 14 generates a transformation matrix (Gram matrix) that maps to a feature space of a higher dimension than the number of dimensions that the feature amount has by inputting the feature amount (principal component) into the quantum kernel circuit. The training unit 14 generates the mapped feature amount by multiplying the feature amount by the Gram matrix. The mapped feature amount is a feature amount of a higher dimension than the number of dimensions that the feature amount before mapping has. That is, the training unit 14 maps the feature amount to a feature space of a higher dimension than the feature amount before mapping by multiplying the feature amount by the Gram matrix, and generates a mapped feature amount having a larger number of dimensions than the feature amount. The training unit 14 performs training using the SVM with the mapped feature amount. The training unit 14 performs training by causing the SVM to learn an identification boundary for classifying the mapped feature amounts into a plurality of clusters, and constructs a trained learning model. Thereby, in the identification system 1 of the embodiment, the feature amount obtained from the image can be mapped to a feature space of a higher dimension. Therefore, even when the feature amount obtained from the image is complex and it is difficult to set an identification boundary, such a complex feature amount can be classified into classes by mapping the feature amount obtained from the image to a feature space of a higher dimension.

[0053] In addition, in the identification system 1 of the embodiment, the training unit 14 generates feature amounts after mapping using the quantum kernel QK created by a quantum circuit. Thereby, in the identification system 1 of the embodiment, it is possible to suppress the calculation cost. In the kernel method using a classical kernel, as the feature space becomes high-dimensional, the calculation required to estimate the kernel function becomes complicated, and it is known that it becomes difficult to calculate a kernel function that can accurately identify. In such a case, by using a quantum kernel instead of a classical kernel, it becomes possible to calculate a kernel function that can accurately identify the feature amounts mapped to a high-dimensional feature space.

[0054] In addition, in the identification system 1 of the embodiment, the training unit 14 extracts feature amounts by performing principal component analysis on an image. The training unit 14 sets the number of principal components so that the cumulative contribution rate of the extracted principal components exceeds a threshold value. Thereby, in the identification system 1 of the embodiment, it is possible to extract the features of the object T until they reach a predetermined ratio or more with respect to all the feature amounts.

[0055] In addition, the identification system 1 of the embodiment further includes an imaging control unit 11. The imaging control unit 11 controls the imaging of the object T. The imaging control unit 11 causes the object T to be imaged from the direction opposite to the direction in which the light irradiation position is located, with respect to the installation position of the object T, in a state where the object T is irradiated with light. The image acquisition unit 12 acquires an image of the object T imaged according to the control by the imaging control unit 11. Thereby, in the identification system 1 of the embodiment, it is possible to acquire an image with high contrast and make it easier to extract the features of the object T.

[0056] In addition, in the identification system 1 of the embodiment, the object T is an apple. The identification target is whether or not there is a "vine crack" inside the apple. The imaging control unit 11 causes the object T to be imaged from the side with the stalk in a state where light is irradiated on the side opposite to the side where the stalk to which the apple was connected to the branch is located. Thereby, in the identification system 1 of the embodiment, it is possible to image an image from which features indicating the presence or absence of a "vine crack" inside the apple are easily extracted.

[0057] Also, in the identification system 1 of the embodiment, the training unit 14 generates a trained model using a data set in which labels are associated with binarized images (training images obtained by binarizing training target images). The identification unit 15 identifies the internal state of the object T by inputting the binarized image into the trained model. Thereby, in the identification system 1 of the embodiment, training and identification using binarized images can be performed, and image information with a smaller capacity can be used compared to the case of using color images, so that the calculation cost can be suppressed.

[0058] All or part of the identification system 1 and the information processing apparatus 10 in the above-described embodiment may be realized by a computer. In that case, a program for realizing this function may be recorded on a computer-readable recording medium, and the program recorded on this recording medium may be read into a computer system and executed to realize it. Here, the “computer system” is assumed to include hardware such as an OS and peripheral devices. Further, the “computer-readable recording medium” refers to a portable medium such as a flexible disk, a magneto-optical disk, a ROM, a CD-ROM, etc., and a storage device such as a hard disk built in a computer system. Furthermore, the “computer-readable recording medium” also includes something that holds a program dynamically for a short time, such as a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line, and something that holds a program for a certain time, such as a volatile memory inside a computer system that becomes a server or a client in that case. Also, the above program may be for realizing a part of the above-described functions, and may further be realized in combination with a program already recorded in the computer system, or may be realized using a programmable logic device such as an FPGA.

[0059] As described above, the embodiments of the present invention have been described in detail with reference to the drawings. However, the specific configuration is not limited to this embodiment, and designs and the like within the scope not departing from the gist of the present invention are also included.

Explanation of Reference Numerals

[0060] 1…Identification system 10…Information processing apparatus 11…Imaging control unit 12…Image acquisition unit 13…Binarized image generation unit 14…Training unit 15…Identification unit 16…Output unit 17…Dataset storage unit 18…Trained model storage unit 20…Imaging device

Claims

1. An object capable of generating a space inside during the growth process is used as an object, An internal state including whether there is a region having a density different from that of the space or the periphery inside the object is set as an object to be identified, A training image acquisition unit that acquires a training target image obtained by imaging the training object, A training unit that generates a trained model for identifying the internal state from an image by learning the correspondence between the object and the internal state using a dataset in which a label indicating the internal state is associated with the training target image, An identification image acquisition unit that acquires an identification target image obtained by imaging the identification object, An identification unit that identifies the internal state in the identification object from the identification target image using the trained model, An output unit that outputs the identification result by the identification unit, An identification system comprising:

2. The training unit generates a feature amount obtained by mapping a feature amount extracted from the training target image to a feature space of a higher dimension than the number of dimensions of the feature amount using a kernel method, and generates a trained model by learning an identification boundary for classifying the mapped feature amounts into a plurality of clusters. The identification system according to claim 1.

3. The training unit generates the mapped feature amount using a quantum kernel created by a quantum circuit. The identification system according to claim 2.

4. The training unit extracts feature amounts by principal component analysis and sets the number of principal components so that the cumulative contribution rate exceeds a threshold value. The identification system according to claim 2 or claim 3.

5. The identification system further includes an imaging control unit that controls imaging of the object, The imaging control unit causes the object to be imaged from a direction opposite to the direction in which the light irradiation position is located with respect to the installation position of the object while irradiating the object with light. The training image acquisition unit acquires the training target image obtained by imaging the object according to the control by the imaging control unit. The identification image acquisition unit acquires the identification target image obtained by imaging the object according to the control by the imaging control unit. The identification system according to claim 1.

6. The object is an apple, The object to be identified is whether there is a stem crack inside the apple, The imaging control unit causes the object to be imaged from the side with the fruit stalk while irradiating light on the side opposite to the side with the fruit stalk to which the apple was connected to the branch. The identification system according to claim 5.

7. The training unit generates the trained model using a dataset in which the label is associated with a training image obtained by binarizing the training target image, The identification unit identifies the internal state of the object for identification by inputting an identification image obtained by binarizing the identification target image into the trained model, The identification system according to claim 1.

8. An object in which a region having a density different from that of the space or periphery may be generated inside during the growth process is used as the object, An internal state including whether or not there is a region having a density different from that of the space or periphery inside the object is set as the identification target, A training image acquisition unit that acquires a training target image obtained by imaging the training object, A training unit that generates a trained model for identifying the internal state from an image by learning the correspondence between the object and the internal state using a dataset in which the label indicating the internal state is associated with the training target image, A learning device comprising:

9. An object in which a region having a density different from that of the space or periphery may be generated inside during the growth process is used as the object, An internal state including whether or not there is a region having a density different from that of the space or periphery inside the object is set as the identification target, An identification image acquisition unit that acquires an identification target image obtained by imaging the object for identification, An identification unit that identifies the internal state of the object for identification from the identification target image using the trained model, An output unit that outputs the identification result by the identification unit, Comprising: The trained model is a model for training the internal state from an image, which is generated by learning the correspondence between the object and the internal state using a learning dataset in which the label indicating the internal state is associated with a training target image obtained by imaging the training object, An identification device.

10. An identification method performed by a computer, An object in which a region having a density different from that of the space or periphery may be generated inside during the growth process is used as the object, An internal state including whether or not there is a region having a density different from that of the space or periphery inside the object is set as the identification target, A training image acquisition unit acquires a training target image obtained by imaging the training object, The training unit generates a trained model for identifying the internal state from an image by learning the correspondence between the object and the internal state using a dataset in which the training target image is associated with a label indicating the internal state. The identification image acquisition unit acquires an identification target image in which the object for identification is imaged. The identification unit identifies the internal state of the object for identification in the identification target image using the trained model. The output unit outputs the identification result by the identification unit. Identification method.

11. A learning method performed by a computer, using, as an object, an object in which a region having a density different from that of the space or the periphery may be generated inside during the growth process, identifying, as an identification target, an internal state including whether or not there is a region having a density different from that of the space or the periphery inside the object, a training image acquisition unit acquires a training target image in which the training object is imaged, the training unit generates a trained model for identifying the internal state from an image by learning the correspondence between the object and the internal state using a dataset in which the training target image is associated with a label indicating the internal state. Learning method.

12. An identification method performed by a computer, using, as an object, an object in which a region having a density different from that of the space or the periphery may be generated inside during the growth process, identifying, as an identification target, an internal state including whether or not there is a region having a density different from that of the space or the periphery inside the object, an identification image acquisition unit acquires an identification target image in which the object for identification is imaged, the identification unit identifies the internal state of the object for identification in the identification target image using a trained model, the output unit outputs the identification result by the identification unit, the trained model is a model for training the internal state from an image generated by learning the correspondence between the object and the internal state using a learning dataset in which a training target image in which the training object is imaged is associated with a label indicating the internal state. Identification method.

13. A program for operating a computer as the learning device according to claim 8, the program for causing the computer to function as each unit included in the learning device.

14. A program for operating a computer as the identification device according to claim 9, the program for causing the computer to function as each part included in the identification device.

Images