TWI708194B - Method for real-time estimation of coffee bean agtron baking level - Google Patents

Abstract

The present invention relates to a method for instantly estimating the roasting degree of coffee beans, which includes the following steps: a step of obtaining green coffee bean information, including the origin, variety, season, and moisture content of the green coffee beans, and preparing a sample roast The training image database is used for real-time estimation model training and testing; for the roasting step, a camera and lighting are used to obtain video or images of roasting coffee beans, and various sensors are used to obtain and record coffee Baking information, which includes room temperature, oven temperature, baking time, sound during baking, gas composition analysis, etc., is used for real-time estimation model training and testing; an algorithm step is to use the video Or compare the image with the image of the sample roasting degree, and use an algorithm to cooperate with the test model to estimate the coffee roasting degree in real time, which can be used to record, display or control the roasting operation of the coffee roaster until it is roasted End.

Description

Instant estimation method of coffee bean roasting degree

The invention relates to a technical category of roasted coffee beans, in particular to a method for real-time automatic identification of roasting degree of coffee beans

The traditional coffee bean roasting method is to put the coffee beans into the roasting device for heating and roasting. During the roasting process, the roaster used the Artisan time-temperature curve to control the roasting temperature. From time to time, he watched the color of the beans through the sampling spoon, smelled the aroma with his nose, and listened to the first or second cracking of the beans with his ears. , Until you see the desired degree of roasting when sampling, finally put the beans and cool them quickly with a fan and stirring. The third step above is to stop roasting, put the beans and cool down. The roaster will blow the coffee beans with a powerful fan and stir them together to achieve a rapid cooling and reach the preset coffee roasting level.

During the roasting process of coffee beans, the moisture of the green coffee beans is slowly released, the weight is reduced, the color is darkened, the volume expands, and the aroma-containing oil is slowly released. The large amount of chlorogenic acid originally contained in raw beans will gradually disappear with the baking process and release good-smelling fruit acid. Its aroma varies with the degree of baking, so it is necessary to meet the needs of different baking flavors from all parties. The mastery of roasting degree is the most important effort in the roasting process of coffee beans.

At present, the mastery of the degree of baking is mainly based on the frequent sampling of the roaster during the baking process, and then based on the roaster's personal experience and the SCAA baking degree color chart to judge the degree of baking and the timing of placing the beans. However, the most precise and authoritative roasting indicator in the coffee industry is Agtron’s NIR The roasting index value 100-0 measured by the line-measured caramel analyzer offline respectively represents the degree of roasting from the shallowest to the deepest, as shown in Figure 26. In the roasting process, refer to the eight kinds of coarse SCAA roasting degree and the cracking sound of roasted beans, and compare them with the precise Agtron roasting index.

The above-mentioned traditional roasting of coffee beans has the following shortcomings:

1. In the traditional coffee bean roasting process, the roaster will take samples from time to time to watch the color of the coffee beans, and compare them with the color chart based on experience to judge the roasting degree of the coffee beans in real time. However, information can only be obtained by sampling frequently from roasted beans, and the mental state of the roaster and the light source spectrum of the roasting room will cause artificial color chart comparison errors.

2. Agtron near-infrared caramel analyzer cannot perform real-time estimation. Although the roasting degree of Agtron near-infrared caramel analyzer is an authoritative indicator in the coffee roasting industry, near-infrared is affected by the high temperature of roasting, so it can only be offline Manual operation is performed to recognize the roasting degree, so that the roasting degree of coffee beans cannot be recognized in real time.

For this reason, the inventor invented a method for instantly estimating the roasting degree of coffee beans. Therefore, the main purpose of the present invention is to provide an instant-identifiable method for estimating the roasting degree of coffee beans; the secondary purpose of the present invention is to: A method for real-time estimation of roasting degree of coffee beans to improve the recognition accuracy of roasting degree.

To achieve the above objective, the present invention employs the following technical means: The present invention relates to a method for real-time estimation of roasting degree of coffee beans, which may include the following steps: a step of obtaining green coffee bean information may include the origin, variety, and season, Moisture content and other information about green coffee beans, and create a sample roasting degree image database for real-time estimation model training and testing; the roasting step is carried out with a camera and lighting to obtain roasted coffee beans Video or image, the same Various sensors are used to obtain and record coffee roasting information, including room temperature, oven temperature, roasting time, roasting sound, gas composition analysis, etc., for real-time estimation model training and testing ; An algorithm step is to compare the video or image with the sample roasting degree image, and use an algorithm to cooperate with the test model to estimate the coffee roasting degree in real time, which can be used to record, display or control coffee roasting The baking operation of the cultivator until the baking ends.

The method for real-time estimation of the roasting degree of coffee beans is based on the deep learning technology of a convolutional neural network "time sequential binary classification model", which includes the following steps: a coffee bean roasting degree setting step is applied A chronological binary classification model of coffee beans roasting classification uses a roasting degree classification; a video recording step is to record coffee beans during the roasting process; an image capturing step is to capture ROI Perform real-time capture of the recorded video into an image; and an estimation step, which is to input the captured image into the chronological binary classification model to perform the baking degree estimation calculation process; through the above steps, the system obtains real-time estimation and notification The current roasting degree classification of the baker (man or machine), if it reaches the set roasting degree classification, the baker (man or machine) can stop roasting and place the beans to cool.

The above-mentioned time sequence binary classification model performs baking degree estimation calculation steps: Step (1): Set the initial parameter setting of the time sequence binary classification model: binary model counter k=1, threshold value of consecutive times T= 4. The continuous counter t=0, the predetermined baking degree classification I; Step (2): Switch to the P _k binary model to identify the kth and k+1 categories; Step (3): Input the next screen to P _{k to} do Classification; Step (4): Is the result of classification the k+1 class? If yes, then t=t+1, if not, then t=0, and go back to the above step (3); Step (5): Is t=T? , If yes, go to the next step (6), if not, go to the above step (3); step (6): whether k+1=I? If it is, it declares "has reached the first category of the predetermined baking degree" and ends the algorithm. If not, then i=k+1, it is declared that it has entered the i-th type of baking degree, and the above step (2) is returned.

The input of the above-mentioned time sequence binary classification model is the k-th image training set BA, and the output is the P _k binary model, and the training steps of the P _k binary model are as follows: Step (1): Take RGB training The image set BA belongs to the k-th type and the k+1-th type training image set; step (2): extract the images into training and verification; step (3): start training.

The above-mentioned method for making the k-th image training set BA further includes: Step (1): Make training images B TIB (Training Images (Baking)); Step (2): Increase the number of training set images to form training images TIBA (Training Images(Baking)(augmentation)), that is, based on the type K baking degree training image set BA obtained in the previous step (1), using data augmentation technology to increase the number of training image sets, and use data to enhance ( data augmentation) technology can be upgraded to 3600 (50x72) sheets.

The method for real-time estimation of the roasting degree of coffee beans is based on the "convolutional neural network method", which includes the following steps: a coffee bean roasting degree setting step is based on a convolutional neural network suitable for coffee bean roasting classification The road model uses a roasting degree classification; a video recording step is to record the coffee beans in the roasting process; an image capturing step is to capture the video into an image in real time by ROI capture; and An estimation step is to estimate both the captured image and the roasting degree classification selected by the convolutional neural network model; through the above steps, if the coffee beans reach the set roasting degree classification during the roasting process, Then stop roasting and put the beans to cool.

The convolutional neural network model used above further includes the following: step (1) placing the roasted sample coffee beans; step (2) roasting the K-th sample coffee beans without heating; step ( 3) Record the roasting video of type K sample coffee beans; step (4) make RGB image set; step (5) convolutional neural network training and parameter adjustment.

The sample coffee beans used above are set to 10 types of sample coffee beans corresponding to the Agtron value.

The convolutional neural network training and parameter adjustment system used above further includes the following: step (1) refer to the CNN model; step (2) modify the FEP parameters of the CNN model; step (3) input the RGB image set to start training; step (4) Test and determine the best parameters.

The above-mentioned FEP parameter system includes the size of the convolution kernel of the convolutional layer, the excitation function, the size of the filter of the pooling layer and its steps, and the color space of the coffee roasting color.

Through the above technical means, the present invention can achieve the following effects:

1. The present invention can provide a time-baking degree curve that more directly reflects the baking degree than the traditional and widely used time-temperature curve. Although the standard coffee Agtron roast degree is obtained by measuring the caramel index by infrared, it is also highly correlated with color and shape information. Therefore, the present invention divides the roasting degree into ten categories (levels), uses a general color camera to take the image, and uses a deep learning convolutional neural network for coffee roasting degree training to obtain a coffee bean roasting time and roasting degree The graph is for real-time identification of roasted coffee beans.

。因每次只分兩類，其精準度預期將可大幅提升，且不會產生烘焙度順序錯誤現象。 2. When the coffee beans are actually roasted, the Agtron value of the roasting degree should be Monotonically Decreasing. If the current roasting degree category is the kth category, the next category must be the k+1th category, so the present invention proposes "Temporal Sequential Binary Classifier (TSBC)", if the total number of categories is K, this model is composed of K-1 binary classification models. The P _k- th binary classification model only classifies the _k- th category and the k+1-th category,

. Since there are only two categories at a time, its accuracy is expected to be greatly improved, and there will be no wrong baking order.

A: Instant estimation method of coffee beans roasting degree based on chronological binary classification

B: Real-time estimation method of coffee beans roasting degree based on convolutional neural network method

Figure 1: is a flow chart of the steps of the instant estimation method of coffee beans roasting degree of the present invention.

Fig. 2 is a flow chart of the steps of the instant estimation method of coffee beans roasting degree based on the chronological binary classification method of the present invention.

Figure 3: is a flowchart of the time sequence binary classification algorithm of the present invention.

Figure 4: is a training flowchart of the time sequence binary classification model of the present invention.

Figure 5: is a flow chart of the steps of making the RGB image training set BA of the present invention.

Fig. 6 is a schematic diagram of the data augmentation method of the training image set BA TIBA of the present invention, in which (a) rotation (b) horizontal flipping (c) cropping.

Fig. 7 is a flow chart of the steps of the instant coffee bean roasting degree estimation method based on the convolutional neural network method of the present invention.

Fig. 8 is a flow chart of the steps of making an RGB image training set in the present invention.

Fig. 9 is a color map of 10 types of coffee beans with respect to Agtron value of the present invention.

Fig. 10 is a flow chart of the steps for making 10 types of sample coffee beans corresponding to the Agtron value in the present invention.

Fig. 11 is a flow chart of the steps of making a convolutional neural network model suitable for coffee bean roasting degree recognition according to the present invention.

Figure 12: This is the reference convolutional neural network model of the present invention (refer to the CNN model).

Figure 13: This is a CNN model suitable for the identification of coffee beans roasting degree according to the present invention.

Figure 14 is a comparison diagram of smoke (a) a small amount of smoke (b) a large amount of smoke in the image of the present invention.

Figure 15: Line graph of 10 types of coffee beans roasting video test using TSBC test.

Figure 16: A graph showing the roasting time and roasting temperature of Artisan coffee beans.

First of all, please refer to Figure 1. The present invention relates to a method for real-time roasting degree estimation of coffee beans, which uses a camera (visible light or invisible light) and appropriate lighting (lighting) (visible light or invisible light) to obtain instant roasting The image of coffee beans can be obtained by appropriate sensors to obtain relevant information during coffee roasting. Based on one or more of the above information as input, the appropriate algorithm can estimate the roasting of coffee beans in real time degree. Step 1: Obtain information about green coffee beans, such as the season of origin, washing or sun treatment, and moisture content of green coffee beans, and create a sample roasting degree image database; Step 2: When roasting , Use appropriate (visible light and or invisible light) cameras with appropriate lighting to obtain video or images of roasting coffee beans. At the same time, use other appropriate sensors to obtain and record coffee roasting related information, such as room temperature, oven temperature, roasting time, sound during roasting, gas composition analysis, etc.; and Step 3: This video or The image is compared with the image of the roasting degree of the sample, and an algorithm is used to estimate the roasting degree of coffee in real time, which can be used to record, display or control the roasting operation of the coffee roaster until the roasting is completed.

Please refer to Figure 2. The present invention relates to a coffee bean roasting degree instant estimation method. As shown in Figure 2, it is a coffee bean roasting degree instant estimation method A based on a chronological binary classification method, which includes the following The steps described: a step of setting the roasting degree of coffee beans, which is based on a Temporal Sequential Binary Classifier (TSBC) suitable for coffee roasting classification, selecting a roasting degree classification; The coffee beans in the training process are captured and recorded; an image capture step is to capture the recorded video into an image in real time by ROI capture; and an estimation step is to input the captured image in the time sequence binary classification The model is used to perform the roasting degree estimation calculation program; through the above steps, if the coffee beans reach the set roasting degree classification during the roasting process, the roasting is stopped and the beans are cooled.

The above-mentioned method for real-time roasting degree estimation of coffee beans is as shown in Fig. 3, wherein the chronological binary classification model (TSBC) performs roasting degree estimation calculation steps as follows:

Step (1): Set the initial parameter settings of the chronological binary classification model: binary model counter k=1, continuous number threshold T=4, continuous number counter t=0, predetermined baking degree classification I.

Step (2): Switch to the P _k binary model to identify the k and k +1 types.

Step (3): Input the next picture to P _k for classification.

Step (4): Is the result of classification the k+1 class? If yes, then t = t +1, if not, then t =0, and return to the above step (3).

Step (5): Is t=T? , If yes, go to the next step (6), if not, go to the above step (3).

Step (6): Is k+1=I? If it is, it declares "has reached the first category of the predetermined baking degree" and ends the algorithm. If not, then i=k+1, it is declared that it has entered the i-th type of baking degree, and the above step (2) is returned.

，如果類別總數為K，則由K-1個二元分類模型依序運作而成。P _k 二元模型訓練流程的輸入為第k類影像訓練集BA，而輸出則為P _k 二元模型，其訓練步驟如下述： Further, the above-mentioned training process of the temporal sequential binary classification model (TSBC) is shown in FIG. 4. Because TSBC is composed of multiple binary classifiers

, If the total number of categories is K, then K-1 binary classification models are operated in sequence. The input of the P _k binary model training process is the k-th image training set BA, and the output is the P _k binary model. The training steps are as follows:

Step (1): Take a training image set belonging to the kth category and the k+1th category in the RGB training image set BA. Here, only two consecutive types of training image sets are taken each time, and the initial value of k is 1, so at the beginning, the first and second types will be selected.

Step (2): The extracted image will be divided into training and verification. There are 3,600 images in each category, randomly divided into: 90% training and 10% verification. Because the baked video test will be used, there is no test image here.

，就能得到TSBC模型

。 Step (3): Start training. In this step, the training image set will be used for training, and the verification image set will be used to verify the accuracy after each epoch, and the weight with the highest accuracy will be retained to obtain P _k . Repeat the above steps until the

, You can get the TSBC model

.

Further, the method of making the RGB image training set BA, as shown in Figure 5, is to directly input the k-th sample coffee bean roasting video for making sample coffee beans, and take the last 50 frames to make the RGB image training set B TIB (Training Images(Baking)), and the number of training images for each type is set to 50 frames, which can better reflect the actual coffee bean color and smoke conditions during heating and roasting. Finally, the number of images is increased to obtain the RGB image training set A TIBA (Training Images (Baking) (A ugmentation)).

Step (1): Make training images B TIB (Training Images (Baking)). The production method of the k-th training image set TIB_k is to input the K-th sample coffee bean roasting video, and only take the frame of the last T=2 seconds of the video. Because in our experimental framework, the image capture screen contains other objects that are not coffee beans, we capture a 50x50 square block containing only coffee beans. Since the frame rate during playback is set to 25f/s, the number of TIB_k is 25x2=50 frames. TIB={TIB_k, k=1,2,..K} The total number of images is 50x10=500

Step (2): Increase the number of images in the training set to form a training image TIBA (Training Images (Baking) (augmentation)). If the amount of training set is too small, it will cause insufficient training of the model; if there is enough training amount, although T can be increased to increase the number of Tx25 training image sets, the coffee roasting degree becomes faster and faster with time. Increasing T will allow the training set to include Too many training images with a lower final baking degree than the K-th category. This is the dilemma of time-sequential sampling, which is worthy of further discussion.

The present invention is based on the K-th baking degree training image set BA obtained in the aforementioned step (1), and adopts data augmentation technology to increase the number of training image sets: first, rotate each 100x100 training image, and then add Rotate up 90 degrees 3 times (90 degrees, 180 degrees, 270 degrees) to expand into four 100x100 images, as shown in Figure 6(a). Each image is then cropped into 9 sub-images, as shown in Figure 6(c), the next step is to flip horizontally, flip the input image horizontally (left and right), as shown in Figure 6(b), and repeat the above rotation and cropping Cut steps, so it can finally be expanded into 72 (4x9x2) 50x50 sub-images. Therefore, the 50 100x100 images obtained in this step (1) can be upgraded to 3600 (50x72) images by data augmentation technology, which is called the training image set B.

The present invention relates to another method for real-time estimation of roasting degree of coffee beans, namely method B for real-time estimation of roasting degree of coffee beans based on the convolutional neural network method. As shown in FIG. 7, it includes the following steps: The setting step is to select a roasting degree classification in a Convolution Neural Network (CNN) model suitable for coffee roasting classification; a video recording step is to record coffee beans in the roasting process An image capture step is to capture the video into an image in real time by ROI capture method. Because the frame screen is not all coffee beans, it is necessary to take out a 50x50 square block containing only coffee beans; and an estimation The step is to estimate both the captured image and the baking degree classification selected by the convolutional neural network model. When the CNN classification output category (2-10 categories) reaches the baking degree category set by the user, the system will "Target xx roasting degree Agtron xx degree has been reached" is issued to remind the user to immediately stop roasting and cooling the beans; through the above steps, if the coffee beans reach the set roasting degree classification during the roasting process, stop roasting and Place the beans to cool.

In the above-mentioned method for real-time estimation of roasting degree of coffee beans, as shown in FIG. 8, the convolutional neural network model used therein further includes the following steps: Step (1) Enter the roasted sample coffee beans , And there are K sample coffees; step (2) roasting the K sample coffee beans without heating; step (3) recording the roasting video of the K sample coffee beans; step (4) making RGB image set; step ( 5) Convolutional neural network training and parameter adjustment.

In particular, the sample coffee beans used are set to 10 types of sample coffee beans corresponding to the Agtron value (including the whole-life bean samples), and the color of the 10 types of sample coffee beans with the Agtron value is shown in Figure 9; The manufacturing method of Agtron’s 10 sample coffee beans, as shown in Figure 20, includes the following steps: step (1) preparation and setting; step (2) bean insertion and video recording; step (3) stop video recording and at the same time (Or immediately) remove heat and cool down the beans; step (4) measure the Agtron roasting degree. Through the above-mentioned sample coffee bean manufacturing method, such as In this way, you can obtain the K-th sample coffee beans and the K-th sample coffee roasting video based on the Agtron value, where K=2, 3, 4...10.

In the above-mentioned real-time estimation method of coffee beans roasting degree, as shown in Figure 21, the used convolutional neural network training and parameter adjustment system further includes the following steps: Step (1) Refer to the CNN model, as shown in Figure 22; Step (2) Modify the FEP (Features Extraction Parameters, FEP) parameters of the CNN model; Step (3) Input the RGB image set to start training; Step (4) Test and determine the best parameters. The purpose of the present invention by adjusting the FEP parameters is to improve the coffee bean roasting degree recognition ability, so as to obtain a CNN model suitable for coffee bean roasting degree recognition, as shown in FIG. 23.

In the above-mentioned method for real-time estimation of roasting degree of coffee beans, the FEP parameter report includes the color space, the size of the convolution kernel of the convolutional layer, the excitation function, and the filter size of the pooling layer. The optimal FEP parameter values are shown in Table 1 Shown.

The present invention uses a line graph to evaluate the "real-time estimation method of coffee beans roasting degree based on convolutional neural network method" to test the results of coffee roasting video. The vertical axis is the Agtron value and the horizontal axis is the roasting time (seconds). The convolutional neural network recognizes the classification once per second, and finally connects the results to obtain the sample coffee bean roasting video test line chart. The present invention will use this line chart to check whether the classification result is consistent with roasted coffee beans Features to ensure that there will be no classification errors during actual baking.

The present invention then uses the estimated difference to evaluate the "real-time coffee bean roasting degree estimation method based on the convolutional neural network method", which is the result of testing the coffee bean roasting video and will be used to view the convolutional neural network The difference between the Agtron value of the estimated category and the actual tested category Agtron value in the last second of the baking video is shown in the following table.

It can be seen from the above table that the "real-time estimation method of coffee beans roasting degree based on the convolutional neural network method" is completely correct for the second, third, sixth, and seventh types. Class 8, 9 Class and Class 10 can only be classified into Class 7 and cannot be classified deeper. This may be due to smoke, as shown in the figure, or the relationship between the training image and the actual roasting image of coffee beans. Shown in Figure 24.

Regarding the evaluation of the "real-time estimation method of roasting degree of coffee beans based on the chronological binary classification method", the present invention uses two evaluation methods (1) using a line graph to evaluate whether it meets the characteristics of roasted coffee beans (2) calculating the absolute difference The value is used to evaluate the difference between the value measured by the Agtron near-infrared caramel analyzer, as follows: Use a line graph to evaluate: Please refer to Figure 25, in the use of a line graph to evaluate whether it meets the characteristics of roasted coffee beans, as shown below It can be found that after using TSBC, there is no violation of the characteristics of roasted coffee beans, because TSBC has added chronological information, but it is different from the convolutional neural network model suitable for coffee bean roasting degree recognition, because in TSBC In the definition, it is necessary for the P _k- th model to be classified into the k +1-th category for T consecutive times. At this time, it can be regarded as the k +1-th category. Therefore, the dashed box in the fifth category of the line chart, although the It is jumping in the 5th and 6th categories, but the number of the 6th category has not reached the continuous number threshold, so it is still in the 5th category. Compared with the convolutional neural network suitable for coffee bean roasting degree recognition, it has successfully solved the violation of the characteristics of roasted coffee beans, and correctly identified and classified the 8, 9, and 10 types of smoke in the video.

Use TSBC to test the results of coffee roasting video, because the present invention defines that in TSBC, the P _k- th model must be classified to the k+1-th consecutive T times, and then it is considered to be the k+1-th category, so finally The estimated classification Agtron value is TSBC's final identification classification category and then converted to the corresponding Agtron value, which is different from the last image of the identification classification video in Chapter 4. In terms of using TSBC to estimate the difference evaluation, because all the identification and classification are correct, the Agtron difference is 0.

Based on the above, the following conclusions can be drawn: The present invention uses the visible light color of a general color camera and a convolutional neural network to estimate in real time the most authoritative indicator Agtron roasting degree in coffee roasting, because Agtron roasting degree is highly correlated with color information. Therefore, the general convolutional neural network structure and parameters are adjusted to develop a "convolutional neural network suitable for coffee bean roasting degree recognition", but the recognition ability is insufficient due to other factors (such as smoke), so according to roasted coffee beans The characteristics of the chronological binary classification model (TSBC) was developed, and finally a satisfactory result was obtained, and it was also in line with our expectations. It can provide instant delivery of roasting degrees to roasters to reduce the workload of professional roasters and make general Amateur roasters can also roast professional coffee beans.

Claims (6)

A method for real-time coffee bean roasting degree estimation includes the following steps: a step of obtaining green coffee bean information, which includes the origin, variety, season, and water content of the green coffee beans, and forming a sample roasting degree image data The library is used for real-time estimation model training and testing; for the roasting step, a camera and lighting are used to obtain video or images of roasting coffee beans. At the same time, various sensors are used to obtain and record coffee roasting information. It should include room temperature, furnace temperature, baking time, sound during baking, gas composition analysis, etc., for real-time estimation model training and testing; an algorithm step is to combine the video or image with the sample The roasting degree images are compared, and the coffee roasting degree is estimated in real time with an algorithm and real-time estimation model, which can be used to record, display or control the roasting operation of the coffee roaster until the roasting is completed; The algorithm steps are based on the chronological binary classification method, which includes the following steps: a coffee bean roasting degree setting step is to select a roasting degree classification in a time sequential binary classification model suitable for coffee bean roasting classification; The video step is to take a video of the coffee beans during the roasting process; an image capture step is to capture the video into an image in real time by way of ROI capture, in which a 50x50 block containing only coffee beans is captured And an estimation step, which is to input the captured images in the chronological binary classification model to perform the roasting degree estimation calculation procedure; through the above steps, if the coffee beans reach the set roasting degree classification during the roasting process, then The roasting is stopped and the beans are cooled down; wherein the time sequence binary classification model performs roasting degree estimation calculation steps: Step (1): Set the initial parameter setting of the time sequence binary classification model: binary model counter k=1 , The threshold value of consecutive times T=4, the consecutive times counter t=0, the predetermined baking degree classification I; Step (2): switch to the P _k binary model to identify the k and k +1 categories; Step (3): Enter the next screen to P _k for classification; Step (4): Is the classification result of the k+1 category? If yes, then t = t +1, if not, then t =0, and go back to the above step (3); Step (5): Is t=T? , If yes, go to the next step (6), if not, go to the above step (3); and step (6): whether k+1=I? If it is, it declares "has reached the first category of the predetermined baking degree" and ends the algorithm. If not, then i=k+1, it is declared that it has entered the i-th type of baking degree, and the above step (2) is returned. The method for real-time coffee bean roasting degree estimation according to claim 1, wherein the input of the chronological binary classification model is the k-th image training set BA, and the output is the P _k binary model, and the P _k binary classification model The training steps of the model are as follows: Step (1): Take the k-th and k+1-th training image sets from the RGB training image set BA; Step (2): Train and verify the taken images; Step (3) : Start training. The method for real-time coffee bean roasting degree estimation according to claim 2, wherein the method for making the k-th image training set BA further includes: Step (1): making training image B TIB (Training Images (Baking)); Step (2): Increase the number of images in the training set to form training images TIBA (Training Images (Baking) (augmentation)), that is, the K-th type of baking training image obtained in the previous step (1) Based on the BA, data augmentation technology is used to increase the number of training image sets. The method for real-time coffee roasting degree estimation as described in claim 1, in which, in addition to the camera, other sensors can obtain the coffee roasting information as input to obtain and record the coffee roasting information, including room temperature, oven The internal temperature, baking time, sound during baking, and gas composition analysis data provide the baker (man or machine) to fully understand the baking conditions in real time and obtain the best baking quality required. A method for real-time coffee bean roasting degree estimation includes the following steps: a step of obtaining green coffee bean information, which includes the origin, variety, season, and water content of the green coffee beans, and forming a sample roasting degree image data The library is used for real-time estimation model training and testing; for the roasting step, a camera and lighting are used to obtain video or images of roasting coffee beans. At the same time, various sensors are used to obtain and record coffee roasting information. It should include room temperature, furnace temperature, baking time, sound during baking, gas composition analysis, etc., for real-time estimation model training and testing; An algorithm step is to compare the video or image with the sample roasting degree image, and use an algorithm to cooperate with the real-time estimation model to estimate the coffee roasting degree in real time, which can be used to record, display or control coffee roasting The roasting operation of the roasting machine until the roasting ends; the algorithm step is based on the convolutional neural network method, which includes the following steps: a coffee bean roasting degree setting step is applied to a roll of coffee beans roasting classification The product neural network model uses a roasting degree classification; a video recording step is to record the coffee beans in the roasting process; an image capturing step is to capture the video in real time by ROI capture An image in which a 50x50 block containing only coffee beans is extracted; and an estimation step is to estimate both the captured image and the roasting degree selected by the convolutional neural network model; by the above steps If the coffee beans reach the set roasting degree classification during the roasting process, the roasting will be stopped and the beans will be cooled down; the convolutional neural network model used further includes the following: step (1) the roasting is completed Sample coffee beans; step (2) roasting type K sample coffee beans without heating; step (3) recording the roasting video of type K sample coffee beans; step (4) making RGB image collection; step (5) volume Convolutional neural network training and parameter adjustment; the used convolutional neural network training and parameter adjustment system further includes the following: step (1) refer to the CNN model; step (2) modify the FEP parameters of the CNN model; step (3) Input the RGB image set to start training; step (4) test and determine the best parameters; and the FEP parameter The number system report includes the size of the convolution kernel of the convolution layer, the excitation function, the filter size of the pooling layer and its steps, and the color space of the coffee roasting degree; the aforementioned sample coffee beans are set to 10 types of samples corresponding to the Agtron value Coffee beans. As described in claim 5, the coffee bean roasting degree real-time estimation method, in which, in addition to the camera, other sensors can obtain the coffee roasting information as input to obtain and record the coffee roasting information, including room temperature, oven The internal temperature, baking time, sound during baking, and gas composition analysis data provide the baker (man or machine) to fully understand the baking conditions in real time and obtain the best baking quality required.

Images