JP7257745B2 - Moisture content estimator - Google Patents

Description

The present invention relates to a moisture content estimation device for estimating the moisture content of sake rice.

In sake brewing, the processing steps of raw rice, such as rice polishing, soaking, and steaming, are important processes that influence the results of subsequent brewing processes. In particular, during the steeping process, daiginjo sake is called limited water absorption. This is important for improving quality. However, it is known that the water content of rice varies greatly depending on the condition of the raw material rice (variety, rice polishing ratio, moisture content, etc.) and the soaking conditions (time, temperature, etc.).

Patent Document 1 discloses optically monitoring rice to be cooked that is placed in a state of being immersed in water, and measuring moisture content-related information related to the moisture content of the rice to be cooked over time. describes a technique for detecting the saturated state of the moisture content of the rice to be cooked estimated from the above, and determining that the rice to be cooked is in a moisture content state suitable for cooking. As the moisture content related information, the expansion rate obtained from the change in the size of the rice to be cooked due to the water absorption of the rice to be cooked, which is obtained by monitoring the light from the rice to be cooked, and the transmitted light from the rice to be cooked. The transmitted light intensity associated with the water absorption of the rice to be cooked, which is obtained by monitoring, is exemplified.

JP 2016-105074 A

The method described in Patent Document 1 can detect a state in which the moisture content of rice is saturated, but in a transitional state in which the moisture content is not saturated, it is possible to determine specifically what the moisture content of rice is. you can't.

The present invention has been made in view of the above problems, and an object thereof is to provide a moisture content estimating apparatus capable of estimating the moisture content of sake rice.

A moisture content estimation apparatus according to the present invention for achieving the above object is characterized by a photographing device for photographing sake rice arranged in a state of being immersed in water, and an image for obtaining image data photographed by the photographing device. a data acquiring unit, a cracking rate calculating unit that refers to the image data acquired by the image data acquiring unit, and calculates a cracking rate, which is the ratio of sake rice with cracks among a plurality of sake rice, Based on the cracking rate calculated by the cracking rate calculation unit and the relationship between the pre-stored cracking rate of sake rice and the moisture content of sake rice placed in a state of being immersed in water , and a moisture content estimating unit for estimating the moisture content of the sake rice arranged in a state of being placed in a state of being filled .

There is a difference between the cracking rate, which is the percentage of sake rice with cracks among the multiple sake rice placed in water, and the moisture content of sake rice placed in water. It has been found that a given relationship holds. In other words, if the crack rate, which is the percentage of sake rice that has cracked due to absorption of water, among the sake rice placed in a state of being immersed in water, is known, it can be placed in a state of being immersed in water. It is possible to estimate the moisture content of multiple sake rice. Therefore, in this characteristic configuration, the crack rate calculation unit refers to the image data acquired by the image data acquisition unit, and calculates the crack rate, which is the ratio of the sake rice with cracks among the plurality of sake rice. , the moisture content estimating unit determines the cracking rate calculated by the cracking rate calculating unit, and the relationship between the pre-stored cracking rate of sake rice and the moisture content of sake rice placed in a state of being immersed in water. Based on this, the moisture content of sake rice placed in a state of being immersed in water is estimated. In other words, even in a transitional state where the water content of sake rice is not saturated, the water content of sake rice placed in water can be estimated in real time from the image data.

Another characteristic configuration of the moisture content estimation device according to the present invention is that the crack rate calculation unit extracts sake rice image data of a region containing only one sake rice from the image data acquired by the image data acquisition unit. An image processing unit that performs image processing including processing to generate processed image data of the one brewer's rice; a morphology determination unit that determines whether or not the one sake rice in the processed image data has cracks; and the morphology determination unit for each of the plurality of sake rice included in the image data acquired by the image data acquisition unit. and a data analysis unit for calculating the crack rate based on the determination result.

According to the above characteristic configuration, the image processing unit serving as the cracking rate calculation unit performs image processing including processing for extracting sake rice image data of an area containing only one sake rice from the image data acquired by the image data acquisition unit. is performed to generate processed image data of the one sake rice, and the morphology determination unit uses the result of executing machine learning on the accumulated training data to determine a piece of the processed image data generated by the image processing unit. The data analysis unit determines whether or not each of the sake rice has cracks, and the data analysis unit determines the crack rate based on the determination result of the morphology determination unit for each of the plurality of sake rice included in the image data acquired by the image data acquisition unit. Calculate That is, the moisture content of the sake rice can be estimated from the image data obtained by the image data obtaining unit, which is obtained by photographing the sake rice arranged in a state of being immersed in water. In particular, in this characteristic configuration, instead of a human being judging whether or not there is a crack in the sake rice, the morphology judgment unit uses the machine learning execution result of the accumulated training data, and the image processing unit generates It is determined whether or not there is a crack in sake rice in the processed image data. As a result, it is possible to obtain a consistent determination result in a short time regarding whether or not the sake rice has cracks. In addition, the image data that the morphology determination unit refers to for determining whether or not there are cracks in sake rice includes only one sake rice in the image data acquired by the image data acquisition unit. This is processed image data after performing image processing including processing for extracting sake rice image data of a region. In other words, the morphology determination unit can determine whether or not there are cracks in sake rice from the image data (processed image data) that has been made suitable for determining whether or not there are cracks in sake rice. The possibility that the determination result is appropriate increases.

Still another characteristic configuration of the moisture content estimation device according to the present invention is that the image processing unit performs the image processing, in the image data acquired by the image data acquisition unit, for a region containing only one sake rice. The point is that it includes a process of extracting image data, a process of rotating the sake rice image so that the one sake rice has a predetermined posture, and a process of adjusting the contrast of the image.

According to the above characteristic configuration, the image data that the morphology determination unit refers to in order to determine whether or not there are cracks in the sake rice is the image data obtained by the image data obtaining unit. Image processing including a process of extracting sake rice image data of an area containing only rice, a process of rotating the sake rice image so that one sake rice has a predetermined posture, and a process of adjusting the contrast of the image is the processed image data after performing In other words, the morphology determination unit can determine whether or not there are cracks in sake rice from the image data (processed image data) that has been made suitable for determining whether or not there are cracks in sake rice. The possibility that the determination result is appropriate increases.

Still another characteristic configuration of the water content estimation device according to the present invention is that the morphology determination unit determines whether the processed image data generated by the image processing unit and sake rice included in the processed image data have cracks. The point is that the machine learning is performed on the training data composed of the combination with the information whether or not.

According to the above characteristic configuration, machine learning is performed using image data (processed image data) in a state suitable for judging whether or not there are cracks in sake rice as training data. Determination of whether or not there are cracks is also performed on the image data (processed image data) in a state suitable for determining whether or not there are cracks in sake rice. As a result, the possibility that the determination result of the morphological determination unit is appropriate increases.

Still another characteristic configuration of the moisture content estimation device according to the present invention is that the crack rate calculation unit determines whether the information on whether or not the sake rice has cracks determined by the morphology determination unit is appropriate. A validity determination unit is provided, and the morphology determination unit receives information on whether or not the sake rice has cracks, which has been determined to be highly appropriate by the validity determination unit, and whether or not the sake rice has cracks. is added to the existing training data to perform the machine learning.
Further, the cracking ratio calculation unit performs image processing including processing for extracting sake rice image data of a region containing only one sake rice in the image data acquired by the image data acquisition unit, Using an image processing unit that generates processed image data of rice and a result of executing machine learning of accumulated training data, the one brewer's rice in the processed image data generated by the image processing unit is cracked. a morphology determination unit that determines whether or not there are cracks in each of the plurality of sake rice included in the image data acquired by the image data acquisition unit; The determination result is stored in a state in which the change history of the morphology of each of the plurality of sake rice after starting to be immersed in water is known. and a data analysis unit that calculates the rate of cracking, and the morphology determination unit determines whether or not the processed image data generated by the image processing unit and the sake rice included in the processed image data have cracks. The machine learning is performed on the training data composed of the combination of the above information, and the breaking rate calculation unit determines that the degree of certainty of the morphology of the sake rice determined by the morphology determination unit is equal to or greater than a predetermined value, Further, when the change history of the morphology of the sake rice after starting to be immersed in water conforms to a predetermined change standard, the validity of the morphology of the sake rice determined by the morphology determination unit is confirmed. The morphological determination unit is provided with information on whether or not the sake rice has cracks, which has been determined to be highly appropriate by the validity determination unit, and whether or not the sake rice has cracks. The machine learning may be performed by adding to the existing training data the combination with the processed image data whose presence or absence has been determined.

According to the above characteristic configuration, the validity determination unit determines whether or not the information on whether or not the sake rice has cracks determined by the morphology determination unit is appropriate. A combination of information on whether or not there are cracks in the sake rice judged to be of high quality and the processed image data in which it is judged whether or not the sake rice has cracks is added to the existing training data, and the machine carry out learning. In other words, the machine learning is performed using image data (processed image data) in a state particularly suitable for judging whether or not there are cracks in sake rice as training data. more likely to be.

It is a figure which shows the structure of the moisture content estimation apparatus of 1st Embodiment. 10 is a flowchart for explaining morphology determination processing; It is a figure explaining the example of image processing. It is a figure which shows the example of the form of sake rice. It is a figure which shows the relationship example between the crack rate of sake rice, and the moisture content of sake rice. 5 is a flowchart for explaining machine learning processing; It is a graph which shows the relationship between the number of times of learning and an accuracy rate. It is a graph which shows the relationship between the number of times of learning and an accuracy rate. It is a graph which shows the relationship between the number of times of learning and an accuracy rate. It is a figure which shows the structure of the moisture content estimation apparatus of 2nd Embodiment. 5 is a flowchart for explaining machine learning processing; 9 is a flowchart for explaining validity determination processing;

<First Embodiment>
A moisture content estimation device 20A (20) according to the first embodiment of the present invention will be described below with reference to the drawings.
FIG. 1 is a diagram illustrating the configuration of the moisture content estimation device 20A. FIG. 2 is a flowchart for explaining the morphology determination process, which is part of the moisture content estimation method. FIG. 3 is a diagram illustrating an example of image processing. A water content estimation device 20A of the present embodiment includes an imaging device 1, an image data acquisition unit 11, a crack rate calculation unit 12, and a water content estimation unit 16. FIG. The moisture content estimation device 20A also includes a storage device 17 that stores information to be handled. In FIG. 1, the water content estimation device 20A includes the photographing device 1, and the sake rice analysis device 10A (10) including the image data acquisition unit 11, the crack ratio calculation unit 12, the water content estimation unit 16, and the storage device 17. It is illustrated in a form composed of

The sake rice analyzer 10A, which constitutes a part of the moisture content estimation device 20A, is realized using one or more computer devices equipped with an information arithmetic processing function, an information input/output function, an information storage function, etc. be done. In that case, the function of the image data acquisition unit 11, the function of the crack rate calculation unit 12 (the function of the image processing unit 13, the function of the morphology determination unit 14, the function of the data analysis unit 15), and the function of the moisture content estimation unit 16 A program (moisture content estimating program) that causes the computer to implement the above may be installed in the computer.

The storage device 17 of the present embodiment includes an image data storage unit 17a that stores image data captured by the imaging device 1, a training data storage unit 17b that stores training data used in machine learning, and a form determination unit 14. and a morphology storage unit 17c for storing the result. Here, the morphology storage unit 17c may store the determination result of the morphology determination unit 14 in a state in which the change history of the morphology of one brewer's rice after starting immersion in water is known.

The output device 2 is, for example, a display device capable of displaying image information, character information, or the like, or a printer capable of printing the image information, character information, or the like on paper or the like.

FIG. 2 is a flowchart for explaining the morphology determination process for determining the morphology of sake rice with reference to image data.
In step #10 of FIG. 2, the image data acquiring unit 11 acquires image data of sake rice arranged in a state of being immersed in water photographed by the photographing device 1 . In other words, an image data acquiring step is executed in which the sake rice placed in water is photographed in the photographing step and the image data thereof is acquired. For example, the photographing device 1 obtains one piece of image data by photographing a plurality of sake rice arranged in a state of being immersed in water. Then, the image data is transmitted to the sake rice analyzer 10A, and the image data acquisition section 11 acquires it. The image data acquired by the image data acquisition section 11 can also be stored in the image data storage section 17 a of the storage device 17 . For example, image data A shown in FIG. 3 is image data acquired by the image data acquisition unit 11 .

Next, the crack rate calculator 12 refers to the image data acquired by the image data acquirer 11 and calculates the crack rate, which is the ratio of the sake rice with cracks among the plurality of sake rice. For example, "cracking" is a state in which the gap between cracks in sake rice is enlarged. In this embodiment, the crack rate calculation unit 12 includes an image processing unit 13 , a morphology determination unit 14 and a data analysis unit 15 . For example, the morphology determination unit 14 can automatically determine whether or not the sake rice is cracked using various image recognition techniques. In the example described later, the morphology determination unit 14 predicts and determines the morphology of sake rice using an algorithm such as a CNN (Convolutional Neural Network). Also, a case will be described in which the morphology determination unit 14 performs machine learning so that information on the morphology of sake rice is output when the processed image data generated by the image processing unit 13 is input. For example, in machine learning, a plurality of weight parameters of CNN, etc., which constitute the algorithm of the morphological determination unit 14 are adjusted.

In step #11 of FIG. 2, the image processing unit 13 performs image processing including processing for extracting sake rice image data of an area containing only one sake rice from the image data acquired by the image data acquisition unit 11, Processed image data of the one brewer's rice is generated. In other words, the image data acquired in the image data acquisition step is subjected to image processing including processing for extracting the sake rice image data of the area containing only one sake rice, and the processed image data of the one sake rice is generated. An image processing step (crack ratio calculating step) is executed. FIG. 3 is a diagram illustrating an example of image processing. In the present embodiment, as the image processing performed by the image processing unit 13, in the image data acquired by the image data acquiring unit 11, a process of extracting sake rice image data of an area containing only one sake rice, It includes a process of rotating the sake rice image so that the sake rice has a predetermined posture, and a process of adjusting the contrast of the image.

For example, image data B shown in FIG. 3 is an image obtained by extracting data including an entire image of one brewer's rice from image data A acquired by the image data acquiring unit 11 . Image data C shown in FIG. 3 is a process of extracting sake rice image data of an area containing only one sake rice by removing images of sake rice other than one sake rice in image data B. This is the image after the trip. The image data D shown in FIG. 3 is an image after the image data C is subjected to the processing of rotating the sake rice image so that one sake rice takes a predetermined posture. In this case, the sake rice image is rotated so that the long axis of the sake rice is along the vertical direction. The image data E shown in FIG. 3 is an image after the image data D is subjected to processing for adjusting the orientation of sake rice (for example, upside-down reversal processing, left-to-right reversal processing, etc.). The image data F shown in FIG. 3 is an image after the image data E is subjected to processing for adjusting the contrast of the image.

In this way, the image data as input data that the morphology determination unit 14 refers to for determining the morphology of sake rice is the image data obtained by the image processing unit 13 from the image data acquired by the image data acquisition unit 11. Image processing including processing to extract sake rice image data of an area containing only sake rice, processing to rotate the sake rice image so that one sake rice is in a predetermined posture, and processing to adjust the contrast of the image It is the processed image data after performing. That is, since the morphology determination unit 14 can determine the morphology of sake rice from the image data (processed image data) in a state suitable for determining the morphology of sake rice, there is a possibility that the determination result is appropriate. increase.

In step #12 of FIG. 2, the morphology determination unit 14 determines whether or not there is a crack. That is, a morphology determination step (crack rate calculation step) is performed to determine whether or not one piece of sake rice in the processed image data generated in the image processing step has cracks. In the present embodiment, the morphology determination unit 14 uses the result of executing machine learning on the accumulated training data to determine one sake rice in the processed image data (image data F in FIG. It is determined whether or not there is a crack in the

In this embodiment, the morphology determination unit 14 determines whether the sake rice is not damaged, whether the sake rice is cracked, or whether the longitudinal section along the longitudinal direction of the sake rice is open. One of four types is determined, that is, the form in which longitudinal cracks occur or the form in which horizontal cracks open in the cross section along the short axis direction of sake rice occurs. Of these, as shown in Table 1, the form in which the sake rice was not damaged and the form in which cracks occurred in the sake rice were judged to be "no cracks." In addition, when the sake rice has horizontal cracks, it is determined that there are cracks.
When there are a plurality of pieces of sake rice immersed in water, it is possible to appropriately set how many of the pieces of sake rice the morphology determination unit 14 determines the morphology of.

The morphology determination unit 14 is configured with an algorithm so that when the processed image data generated by the image processing unit 13 is input, information regarding the morphology of sake rice is output. For example, the CNN executed by the form determination unit 14 performs convolution processing on an input layer to which the processed image data generated by the image processing unit 13 is input, and on the processed image data input to the input layer. A hierarchical network is composed of a convolution layer that obtains a feature map, a pool layer that reduces the feature map output from the convolution layer, a fully connected layer that connects all units, and an output layer. The combination of the convolutional layer and the pool layer is repeatedly provided multiple times, and the fully connected layer is also provided multiple times.

In this embodiment, four types of information (no damage, cracks, vertical cracks, horizontal cracks) are assumed as information on the shape of sake rice that is the determination result of the shape determination unit 14, so the output layer unit of the CNN The number becomes 4. A softmax function can be used as the likelihood function in the output layer. Since the softmax function divides the value of each unit by the accumulated value of all units in the output layer for normalization, the likelihood of each unit in the output layer takes a value between 0 and 1. Then, the morphology determination unit 14 determines the class with the maximum likelihood as the classification class, that is, the morphology of sake rice (determination result). Then, the determination result by the morphology determination unit 14 is stored in the morphology storage unit 17c for each sake rice in association with the identifier assigned to each of the plurality of sake rice. The determination result of the morphology determination unit 14 can be stored in the morphology storage unit 17c in a state in which the change history of the morphology of one piece of sake rice after starting to be immersed in water can be known.

The determination result of the morphological determination unit 14 can also be output from the output device 2 . For example, it is also possible to calculate the ratio of sake rice with cracks (crack rate) among a plurality of sake rice, and output it from the output device 2 in real time. In addition, when the determination result of the morphology determination unit 14 is stored in the morphology storage unit 17c in a state in which the change history of the morphology of one piece of sake rice after the start of immersion in water is known, the crack rate It is also possible to calculate the temporal change of , and output it from the output device 2 .

FIG. 4 is a diagram showing an example of the form of sake rice. Specifically, FIG. 4(a) is a diagram showing a form in which sake rice is not damaged. FIG. 4(b) shows the morphology of cracked sake rice. Figs. 4(c) and 4(d) show the morphology of cracks in which the gap between cracks in the sake rice is enlarged. As for the form of the cracks, there is a form in which a horizontal crack occurs in the transverse cross section along the short axis direction of the sake rice (Fig. 4(c)), and a longitudinal crack in which the longitudinal cross section along the long axis direction of the sake rice opens. There is another form (FIG. 4(d)). Moreover, FIG. 4(b) shows a form in which cracks are generated along the short axis direction in sake rice. In this way, the morphology determination unit 14 determines whether the morphology of the crack is a vertical crack in which the longitudinal section along the long axis direction of the sake rice is open, or a cross section along the short axis direction of the sake rice. It is also determined whether or not it is a form in which an open lateral crack has occurred.

As described above, in the present embodiment, instead of humans determining the morphology of sake rice, the morphology determination unit 14 uses machine learning execution results of accumulated training data, and the image processing unit 13 generates The morphology of one brewer's rice in the processed image data is determined. As a result, it is possible to obtain consistent determination results in a short period of time regarding the morphology of sake rice.

The data analysis unit 15 calculates the crack rate based on the determination result of the morphology determination unit 14 for each of the plurality of sake rice included in the image data acquired by the image data acquisition unit 11 . In other words, a data analysis step (crack ratio calculation step) is performed to calculate the crack ratio based on the determination result of the form determination step. For example, the morphology storage unit 17c stores the determination results of the morphology determination unit 14 for each of a plurality of sake rice in a state in which the change history of the morphology of one sake rice after starting immersion in water is known. It is For example, the morphology storage unit 17c stores one or a plurality of morphologies from the past to the present (latest) as a change history for each sake rice. Then, the data analysis unit 15 reads out the current (latest) morphology of the plurality of sake rice stored in the morphology storage unit 17c, and calculates the proportion of the sake rice with cracks among the plurality of sake rice. Calculate a certain crack rate.

The moisture content estimating unit 16 estimates the moisture content of the sake rice based on the cracking rate calculated by the cracking rate calculating unit 12 and the pre-stored relationship between the cracking rate of the sake rice and the moisture content of the sake rice. to estimate FIG. 5 is a graph showing an example of the relationship between the moisture content of sake rice and the cracking rate of sake rice.
Specifically, the graph shown in FIG. By calculating the average value of the moisture content of a plurality of sake rice measured by the method, the relationship between the crack rate of sake rice and the moisture content is plotted. Then, the approximation line (indicated by the solid line in FIG. 5) for the plurality of plotted points is the relationship between the cracking rate of sake rice and the moisture content of sake rice, which the moisture content estimation unit 16 refers to, as a storage device. 17 is stored. In the present embodiment, the cracking ratio, which is the percentage of sake rice with cracks among the plurality of sake rice, and the moisture content of the plurality of sake rice are defined in a proportional relationship.

As described above, the moisture content estimator 16 calculates the cracking rate of sake rice determined in real time and the relationship between the cracking rate of sake rice and the moisture content of sake rice stored in the storage device 17. , the current moisture content of sake rice placed in water can be estimated. The current moisture content of sake rice estimated by the moisture content estimator 16 can be stored in the storage device 17 and output from the output device 2 . It is also possible to monitor temporal changes in water content and output from the output device 2 when the water content reaches a predetermined value. In this way, the current moisture content of the sake rice placed in water can be estimated in real time from the image data, even in a transient state where the moisture content of the sake rice is not saturated. As a result, for example, the immersion process in sake brewing can be performed appropriately.
In addition, in the relationship between the pre-stored cracking rate of sake rice and the moisture content of sake rice, the range of cracking rate (upper limit, lower limit, etc.) and the range of moisture content (upper limit, lower limit, etc.) and the moisture content may be estimated only within that range.

FIG. 6 is a flow chart illustrating machine learning processing that is part of the moisture content estimation method.
In step #20, the image data acquisition unit 11 acquires image data of sake rice. In other words, an image data acquisition step is executed for acquiring image data obtained by photographing sake rice arranged in a state of being immersed in water. Then, in step #21, the image processing unit 13 performs image processing to generate processed image data. Next, in step #22, the morphology determination unit 14 trains training data composed of a combination of the processed image data generated by the image processing unit 13 and information on the morphology of sake rice included in the processed image data. The training data is stored in the data storage unit 17b, and machine learning is performed on the training data in step #23. For example, one piece of training data is created by labeling one piece of processed image data with the correct data determined by a human regarding the form of sake rice included in the processed image data. be done. A plurality of training data created in this way are stored in the training data storage unit 17b.

In this embodiment, a softmax function is used as the likelihood function in the output layer of the CNN executed by the morphology determination unit 14 . Since the softmax function divides the value of each unit by the accumulated value of all units in the output layer for normalization, the likelihood of each unit in the output layer takes a value between 0 and 1. Then, the morphological determination unit 14 uses error backpropagation to reduce the error between the estimated data (“likelihood”) generated in the output layer and the correct data (“1”). , an algorithm is obtained as a machine learning execution result that corrects a plurality of weight parameters of the CNN to be executed, and outputs information on the form of sake rice when the processed image data generated by the image processing unit 13 is input.

In this way, machine learning is performed using image data (processed image data) in a state suitable for judging the morphology of sake rice as training data. This is performed on the image data (processed image data) in a state suitable for judging the morphology. As a result, the possibility that the determination result of the morphological determination unit 14 is appropriate increases.

Next, effects obtained by performing machine learning using processed image data as training data as in this embodiment will be described.
7 to 9 are graphs showing the relationship between the number of learning times and the accuracy rate. Specifically, FIG. 7 is a graph showing the relationship between the number of times of learning and the accuracy rate when machine learning is performed using training data in which correct data is labeled with image data B in FIG. FIG. 8 is a graph showing the relationship between the number of times of learning and the accuracy rate when machine learning is performed using the image data D of FIG. 3 as training data. FIG. 9 is a graph showing the relationship between the number of times of learning and the accuracy rate when machine learning is performed using the image data F of FIG. 3 as training data.

As shown in FIG. 7, when image data B is used as training data, the accuracy rate is less than 80% even if the number of times of learning reaches 20,000. The correct answer rate of 80% corresponds to, for example, the correct answer rate when a person with some skill determines the form of sake rice from image data. On the other hand, when machine learning is performed using training data in which correct data is labeled with image data D, the number of times of learning is about 12000, and the accuracy rate exceeds 80%. Also, when machine learning is performed using training data in which correct data is labeled with image data F, the number of times of learning is about 5000, and the accuracy rate exceeds 80%. In this way, by using the processed image data after image processing, that is, the image data in a state suitable for judging the morphology of sake rice, as training data, the accuracy rate can be improved even if the number of times of learning is small. had the effect of increasing

As described above, in the moisture content estimating device, by referring to the image data of a plurality of sake rice placed in a state of being immersed in water, the ratio of the sake rice with cracks among the plurality of sake rice A certain crack rate is calculated, and based on the pre-stored relationship between the crack rate of sake rice and the moisture content of sake rice, the moisture content of a plurality of sake rice placed in a state of being immersed in water is estimated. A moisture content estimation method is implemented.

<Second embodiment>
The moisture content estimation device 20B (20) of the second embodiment differs from the above embodiment in the contents of the machine learning process. The moisture content estimation device 20B of the second embodiment will be described below, but the description of the same configuration as in the above embodiment will be omitted.

FIG. 10 is a diagram showing the configuration of a moisture content estimation device 20B of the second embodiment. As illustrated, the moisture content estimation device 20B of the second embodiment further includes a validity determination unit 18 that determines the validity of the form of sake rice determined by the form determination unit 14 . This validity determination unit 18 is provided in the crack ratio calculation unit 12 of the sake rice analyzer 10B (10). Then, the morphology determination unit 14 adds a combination of the information on the morphology of sake rice determined to be highly appropriate by the validity determination unit 18 and the processed image data for which the morphology has been determined to the existing training data. to perform machine learning.
Also in this embodiment, the sake rice analyzer 10B is implemented using one or a plurality of computer devices having an information arithmetic processing function, an information input/output function, an information storage function, and the like. In that case, the function of the image data acquisition unit 11 and the function of the crack rate calculation unit 12 (the function of the image processing unit 13, the function of the morphology determination unit 14, the function of the data analysis unit 15, the function of the validity determination unit 18) , a program (moisture content estimation program) for realizing the functions of the moisture content estimating unit 16 in the computer device.

The morphology storage unit 17c stores the determination result of the morphology determination unit 14 in a state in which the change history of the morphology of one brewer's rice after starting immersion in water can be known. For example, the morphology storage unit 17c stores the determination results of the morphology determination unit 14 from the past to the present for each of a plurality of sake rice whose morphology is to be determined, in a state in which the change history of the morphology is known. are doing. As a result, it is possible to know what kind of change history the morphology of one sake rice showed.

FIG. 11 is a flowchart for explaining machine learning processing according to the second embodiment. This machine learning process is performed in addition to the machine learning process (FIG. 6) described in the first embodiment. Specifically, in step #30, the validity determination section 18 determines the validity of the current form of sake rice determined by the form determination section 14 . In other words, a validity determination step (break rate calculation step) is performed to determine whether the shape of sake rice determined in the shape determination step performed by the shape determination unit 14 is high or low. FIG. 12 is a flowchart for explaining the validity determination processing performed by the validity determination unit 18. As shown in FIG. In step #40, the validity determination unit 18 determines whether or not the degree of certainty of the current form of sake rice determined by the form determination unit 14 is equal to or greater than a predetermined value. For example, when the maximum likelihood calculated by the softmax function used in the output layer of the morphology determination unit 14 is the certainty factor, the maximum likelihood calculated by the softmax function is a predetermined value or more (for example, 0 .8 or higher), the validity determination unit 18 determines that the certainty factor of the current form of sake rice is equal to or greater than a predetermined value, and proceeds to step #41.

Next, in step #41, the validity determination unit 18 determines whether or not the change history of the morphology of sake rice conforms to a predetermined standard. The predetermined criteria are the following six patterns. Then, if the change history of the morphology of brewer's rice conforms to any one of the six patterns, the validity determination unit 18 proceeds to step #42 and determines that "the validity is high."

Specifically, for example, if we look at the morphology of one piece of sake rice continuously, the morphology in which cracks have occurred in the sake rice will not return to the state in which the sake rice is not damaged. Nor does it return to the form in which it arose. In other words, if the morphology of the sake rice after starting to be immersed in water is appropriately determined by the morphology determination unit 14, the change history should meet the predetermined change standard. Therefore, in this characteristic configuration, the validity determination unit 18 determines that the degree of certainty of the shape of the sake rice determined by the shape determination unit 14 is equal to or greater than a predetermined value, and that the shape of the sake rice after starting soaking in water conforms to a predetermined change criterion, the morphology of sake rice determined by the morphology determination unit 14 is determined to be highly appropriate. As a result, accumulation of appropriate training data and machine learning can be performed automatically.

On the other hand, if the validity determination unit 18 determines that the degree of certainty is not equal to or greater than the predetermined value in step #40, and if the change history of the morphology of sake rice does not meet the predetermined criteria in step #41 If so, the process moves to step #43 and it is determined that "the validity is low".

As described above, the validity determination unit 18 determines that the degree of certainty of the shape of the sake rice determined by the shape determination unit 14 is equal to or greater than a predetermined value, and that the shape of the sake rice after starting soaking in water When the change history conforms to the predetermined change standard, the morphology determination unit 14 determines that the morphology of the sake rice is highly appropriate, and the degree of certainty of the morphology of the sake rice determined by the morphology determination unit 14 is less than a predetermined value, or if the change history of the morphology of the sake rice after the start of immersion in water does not meet the predetermined change criteria, the morphology determination unit 14 determines the sake rice The validity of the form is determined to be low.

Next, in step #31, a combination of the information on the morphology of brewer's rice determined to be highly appropriate and the processed image data is used as training data. Then, in step #32, machine learning is performed by adding the training data to the existing training data.

In this way, the validity determination unit 18 determines whether the form of sake rice determined by the form determination unit 14 is highly appropriate. Machine learning is performed by adding a combination of information about the morphology of the sake rice obtained from the model and the processed image data from which the morphology was determined to the existing training data. In other words, accumulation of appropriate training data and machine learning can be performed automatically.

<Another embodiment>
<1>
In the above-described embodiment, the configuration of the water content estimation device 20 has been described with a specific example, but the configuration can be changed as appropriate.
Also, the relationship between the moisture content of sake rice and the cracking rate of sake rice as shown in FIG. 5 is described for the purpose of illustration, and can be changed as appropriate.

<2>
In the above-described embodiment, the morphology determination unit 14 uses a CNN (Convolutional Neural Network) to predict and determine the morphology of sake rice. , support vector machines (SVM), k-nearest neighbor algorithms, discriminant functions, etc.
Further, the morphology determination unit 14 may be configured to determine whether or not the sake rice is cracked using various image recognition techniques different from the above embodiment. Alternatively, the morphological determination unit 14 may be configured to receive the result of determination made by a person as to whether or not cracks have occurred in sake rice, and use the result as its own determination result.

<3>
The configurations disclosed in the above embodiments (including other embodiments) can be applied in combination with configurations disclosed in other embodiments unless there is a contradiction. The described embodiment is an example, and the embodiment of the present invention is not limited to this, and can be modified as appropriate without departing from the object of the present invention.

INDUSTRIAL APPLICABILITY The present invention can be used for a moisture content estimation device capable of estimating the moisture content of sake rice.

1 Photographing device 2 Output device 10 (10A, 10B) Sake rice analyzer 11 Image data acquisition unit 12 Crack rate calculation unit 13 Image processing unit 14 Morphology determination unit 15 Data analysis unit 16 Moisture content estimation unit 17 Storage device 17a Image data storage Unit 17b Training data storage unit 17c Morphology storage unit 18 Validity determination unit 20 (20A, 20B) Moisture content estimation device

Claims (6)

a photographing device for photographing sake rice arranged in a state of being immersed in water;
an image data acquisition unit that acquires image data captured by the imaging device;
a crack rate calculation unit that refers to the image data acquired by the image data acquisition unit and calculates a crack rate, which is a ratio of sake rice with cracks among a plurality of sake rice;
Based on the cracking rate calculated by the cracking rate calculation unit and the relationship between the pre-stored cracking rate of sake rice and the moisture content of sake rice placed in a state of being immersed in water , and a moisture content estimating unit for estimating the moisture content of the sake rice placed in a state where the moisture content is estimating. The crack rate calculation unit
The image data acquired by the image data acquisition unit is subjected to image processing including a process of extracting the image data of a region containing only one sake rice to generate the processed image data of the one sake rice. an image processing unit;
a morphology determination unit that determines whether or not the one piece of sake rice in the processed image data generated by the image processing unit has cracks, using machine learning execution results of accumulated training data;
2. The hydrous according to claim 1, further comprising a data analysis unit that calculates the crack rate based on the determination result of the morphology determination unit for each of the plurality of sake rice contained in the image data acquired by the image data acquisition unit. rate estimator. As the image processing, the image processing unit extracts sake rice image data of a region containing only one sake rice from the image data acquired by the image data acquisition unit, and 3. The moisture content estimating apparatus according to claim 2, comprising a process of rotating the sake rice image so that it assumes the posture of and a process of adjusting the contrast of the image. The morphology determination unit determines whether or not the training data is composed of a combination of the processed image data generated by the image processing unit and information on whether or not there are cracks in the sake rice included in the processed image data. 4. The moisture content estimating device according to claim 2 or 3, which performs machine learning. The crack rate calculation unit includes a validity determination unit that determines the validity of the information on whether or not the sake rice has cracks, determined by the morphology determination unit,
The morphology determination unit determines whether or not the sake rice has cracks, which has been determined to be highly appropriate by the validity determination unit, and the processed rice, which has been determined to have cracks. 5. The moisture content estimation device according to claim 4, wherein said machine learning is performed by adding a combination with image data to said existing training data. The crack rate calculation unit
The image data acquired by the image data acquisition unit is subjected to image processing including a process of extracting the image data of a region containing only one sake rice to generate the processed image data of the one sake rice. an image processing unit;
a morphology determination unit that determines whether or not the one piece of sake rice in the processed image data generated by the image processing unit has cracks, using machine learning execution results of accumulated training data;
For each of the plurality of sake rice contained in the image data acquired by the image data acquisition unit, the morphology determination result of whether or not the rice for sake brewing has cracks is determined by the morphology determination unit after starting immersion in water. a data analysis unit that reads out the latest morphology of sake rice stored in a morphology storage unit that stores the change history of each of the plurality of sake rice in the state in which the history of change in the morphology is known, and calculates the cracking rate. have
The morphology determination unit determines whether or not the training data is composed of a combination of the processed image data generated by the image processing unit and information on whether or not there are cracks in the sake rice included in the processed image data. perform machine learning,
The cracking rate calculation unit determines that the degree of certainty of the morphology of the sake rice determined by the morphology determination unit is equal to or greater than a predetermined value, and that the change history of the morphology of the sake rice after starting immersion in water is a validity determination unit that determines that the form of sake rice determined by the form determination unit is highly appropriate when a predetermined change criterion is met,
The morphology determination unit determines whether or not the sake rice has cracks, which has been determined to be highly appropriate by the validity determination unit, and the processed rice, which has been determined to have cracks. 2. The moisture content estimating device according to claim 1, wherein said machine learning is performed by adding a combination with image data to said existing training data.

Images