A Deep Learning Tobacco Leaf Grading Method Based on Expert Experience Guidance

Abstract

The present invention discloses a deep learning tobacco leaf grading method guided by expert experience. A deep network feature vector of a tobacco leaf image is obtained by using a preset deep feature extraction network. A manual feature vector of the tobacco leaf image is obtained by using traditional computer vision methods. The deep network feature vector and the manual feature vector are spliced into a joint feature vector. The joint feature vector is input into a preset grading judgment network to output a tobacco leaf grading result. The expert re-grades the tobacco leaf images with incorrect grading to obtain updated classification labels, and learns the grading judgment network according to the updated classification labels to obtain an optimized grading judgment network. The present invention has objectivity and consistency in tobacco leaf grading and can achieve a high grading accuracy. In view of the occasional incorrect tobacco leaf grading situation, the domain knowledge of experts is used to assist in improving the grading accuracy.

Description

Technical Field

[0001] The present invention relates to a deep learning tobacco leaf grading method guided by expert experience, belonging to the technical fields of computer vision and tobacco leaf automatic grading. Background Art

[0002] Tobacco leaves are an important crop. When purchasing tobacco leaves, it is necessary to grade tobacco leaves according to different tobacco leaf conditions to give a reasonable purchase price for tobacco leaves. However, tobacco leaves are complex and diverse, the grading standards are difficult to quantify, and they are easily affected by subjective factors. Often, experienced experts are required to grade them, which leads to problems of low efficiency and low accuracy in the tobacco leaf grading process and ultimately affects the quality of finished cigarettes.

[0003] Computer vision and image processing use the extracted features to perform related tasks. Feature engineering relies on professional knowledge to select effective features, but manual features limit the selection flexibility and it is difficult to significantly improve the system performance. With the development of deep learning, neural networks can automatically extract effective features from images, but a large amount of data is required for model parameter update. The combination of the advantages of artificial features and deep features can effectively improve the system grading accuracy and reduce the data requirements.

[0004] The training of existing neural networks requires a large amount of data. When the data is not sufficient, the prior knowledge of people can be used to guide the model. Human-in-the-loop is a human-computer interaction technology that makes full use of the relevant domain knowledge of experts, enables humans and machines to interact with each other, and continuously improves the model performance.

[0005] A heat map is a machine learning visualization technology used to show the key attention areas of the current model. The model can be artificially guided to update by analyzing the model heat map, thereby improving the model performance.

[0006] Neural network architecture search is an automatic machine learning technology. It designs the model structure for relevant data and can obtain the model structure most suitable for the data set, and the effect is significantly better than the network model designed manually. Thanks to the improvement of computing power, neural network architecture search can often automatically search within the predefined parameter space, greatly reducing the workload.

[0007] Therefore, those skilled in the art are eager to integrate the above technologies to solve the problems that the existing tobacco leaf grading methods rely too much on manual grading, resulting in low efficiency and low accuracy. Summary of the Invention

[0008] Objective: In order to overcome the deficiencies in the prior art, the present invention provides a deep learning tobacco leaf grading method guided by expert experience.

[0009] Technical solution: To solve the above technical problems, the technical solution adopted by the present invention is as follows:

[0010] A deep learning tobacco leaf grading method based on expert experience guidance, comprising the following steps:

[0011] Obtain the deep network feature vector of the tobacco leaf image by using a preset deep feature extraction network.

[0012] Obtain the handcrafted feature vector of the tobacco leaf image by using traditional computer vision methods.

[0013] Concatenate the deep network feature vector and the handcrafted feature vector into a joint feature vector.

[0014] Input the joint feature vector into a preset grading judgment network to output the tobacco leaf grading result.

[0015] The expert regrades the tobacco leaf images with incorrect grading to obtain updated classification labels, and learns the grading judgment network according to the updated classification labels to obtain an optimized grading judgment network.

[0016] As a preferred solution, it further includes: the deep feature extraction network outputs an attention map, the expert corrects the attention area of the attention map corresponding to the tobacco leaf image with incorrect grading to obtain an attention guidance map, learns the deep feature extraction network according to the attention guidance map to obtain an optimized deep feature extraction network, obtains an optimized deep network feature vector according to the optimized deep feature extraction network, obtains an optimized joint feature vector according to the optimized deep network feature vector, and learns the grading judgment network according to the optimized joint feature vector and the updated classification labels to obtain an optimized grading judgment network.

[0017] As a preferred solution, the method for obtaining a preset deep feature extraction network includes the following steps:

[0018] Determine the neural network basic parameters of the neural network architecture search process, and the neural network basic parameters include: image resolution, network depth.

[0019] Define the search space of the neural network architecture search, and the search space includes: basic convolution module, convolution kernel size, dilation coefficient, and number of channels.

[0020] Use the random sampling method to sample the network parameters in the search space and calculate the search reward of the model corresponding to the network parameters.

[0021] Repeat the network parameter sampling and calculate the search reward of the model corresponding to the network parameters until the model search reward meets the requirements or the number of sampling times reaches the limit. Select the network parameters with the highest search reward to construct the corresponding model as the deep feature extraction network.

[0022] As a preferred solution, the manual feature vector includes at least one of the length, width, length-width ratio, perimeter, area, circularity, breakage rate, average values of the RGB red, green, and blue channels, or average values of the HSV hue, saturation, and value channels of the tobacco leaves.

[0023] As a preferred solution, the search space of the basic convolution module is {MBConv, Fused-MBConv}, the search space of the convolution kernel size is {3x3, 5x5}, the search space of the dilation coefficient is {1, 2, 4, 6}, and the search space of the number of channels is {32, 64, 128, 256, 512}.

[0024] As a preferred solution, the grading judgment network includes an input layer, a hidden layer, and an output layer. Among them, the number of neurons in the input layer is equal to the length of the joint feature vector output by the joint feature extraction module, and the number of neurons in the output layer is equal to the number of tobacco leaf categories.

[0025] As a preferred solution, the specific method of inputting the joint feature vector into a preset grading judgment network and outputting the tobacco leaf grading result includes the following steps:

[0026] Input the joint feature vector into the grading judgment network to output a predicted probability vector.

[0027] Calculate the probability entropy using the predicted probability vector.

[0028] Judge the relationship between the calculated probability entropy and the set threshold.

[0029] When the probability entropy is less than the set threshold, the grading ends and the tobacco leaf grading result is output.

[0030] When the probability entropy is greater than the set threshold, expert intervention is required to check the tobacco leaf grading result.

[0031] Beneficial effects: A deep learning tobacco leaf grading method based on expert experience guidance provided by the present invention. In view of the strong subjectivity of the existing tobacco leaf grading, this method has objectivity and consistency in tobacco leaf grading and can achieve a high grading accuracy rate. For the occasional incorrect tobacco leaf grading situation, the present invention adopts an expert guidance method, utilizes the domain knowledge of experts, and assists in improving the grading accuracy rate. The method of the present invention can be not only used for tobacco leaf grading, but also applied to other classification fields, such as agricultural products, etc. Description of the Drawings

[0032] Figure 1 It is a schematic structural diagram of the grading device according to the embodiment of the present invention.

[0033] Figure 2 It is a flow chart of the joint feature extraction of the present invention.

[0034] Figure 3 This is a flowchart for constructing a deep feature extraction network using neural network architecture search in the present invention.

[0035] Figure 4 This is a flowchart of the method for the hierarchical judgment network in the present invention.

[0036] Figure 5 This is a flowchart of the method with expert guidance in the present invention. Detailed implementation manners

[0037] The present invention will be further described below in conjunction with specific embodiments.

[0038] As Figure 1 shown, a deep learning tobacco leaf grading device based on expert experience guidance. The proposed device includes three modules, namely a joint feature extraction module, a grading judgment module, and an expert guidance module. Among them, using tobacco leaf grading experience, several effective manual features of tobacco leaves are extracted from tobacco leaf images. At the same time, deep network features of tobacco leaves are extracted using a deep neural network. Aiming at the disadvantage of low flexibility of the existing method model, this method uses neural network architecture search to obtain the deep neural network model most suitable for tobacco leaf data. Finally, the manual features and deep features are fused to obtain joint features for grading. The joint features are input into the neural network of the grading judgment module to predict the final grading result, and the grading misjudgment probability is calculated to prompt the expert to re-judge the uncertain tobacco leaves. The expert guidance module uses expert knowledge to continuously improve the grading accuracy of the entire model. When the grading judgment module points out the tobacco leaves with a relatively high grading error probability, or the expert actively discovers the misgraded tobacco leaves, the expert can point out the correct category of the tobacco leaves to the system and provide an attention guidance map as a judgment basis to assist the model in improving the grading accuracy.

[0039] A deep learning tobacco leaf grading method based on expert experience guidance. The method proposed by the present invention includes the following steps:

[0040] S0: Obtain a deep feature extraction network using neural network architecture search and perform the following processing:

[0041] S01: Determine the basic neural network parameters in the neural network architecture search process. The basic neural network parameters include: image resolution, network depth.

[0042] S02: Define the search space for neural network architecture search. The search space includes: basic convolution module, convolution kernel size, dilation coefficient, and number of channels.

[0043] S03: Use the random sampling method to sample the network parameters in the search space and calculate the search reward of the model corresponding to the network parameters.

[0044] S04: Repeat the sampling of network parameters and calculate the search reward of the model corresponding to the network parameters until the model search reward meets the requirements or the number of samplings reaches the limit. Select the network parameters with the highest search reward to construct the corresponding model as the deep feature extraction network.

[0045] S1: Use traditional computer vision methods to extract handcrafted feature vectors, use the deep feature extraction network to extract deep network feature vectors, and obtain joint feature vectors based on the handcrafted feature vectors and deep network feature vectors, and perform the following processing:

[0046] S11: Use traditional computer vision methods to segment the tobacco leaf image, and obtain a 13-dimensional handcrafted feature vector f m , the handcrafted feature vector f m includes the length, width, aspect ratio, perimeter, area, circularity, breakage rate of the tobacco leaf, and the means of six color channels (red, green, blue, hue, saturation, and lightness);

[0047] S12: Use the deep feature extraction network to extract features from the tobacco leaf image and obtain the deep network feature vector f d ;

[0048] S13: Concatenate the handcrafted feature vector f m and the deep network feature vector f d to obtain the joint feature vector f t .

[0049] S2: Use the joint feature vector to grade the tobacco leaves through the hierarchical judgment network, and perform the following processing:

[0050] S21: Input the joint feature vector into the hierarchical judgment network and output the predicted probability vector.

[0051] S22: Calculate the probability entropy using the predicted probability vector.

[0052] S23: Judge the relationship between the calculated probability entropy and the set threshold.

[0053] S24: When the probability entropy is less than the set threshold, the grading ends.

[0054] S25: When the probability entropy is greater than the set threshold, expert intervention is required to check the grading results.

[0055] S3: The expert re-grades the tobacco leaf images with incorrect grading to obtain updated classification labels, and learns the hierarchical judgment network based on the updated classification labels to obtain an optimized hierarchical judgment network.

[0056] S4: The deep feature extraction network outputs an attention map. The expert focuses on the attention map corresponding to the tobacco leaf image with incorrect grading, corrects the region of interest to obtain an attention guidance map, learns the deep feature extraction network according to the attention guidance map to obtain an optimized deep feature extraction network, obtains an optimized deep network feature vector according to the optimized deep feature extraction network, obtains an optimized combined feature vector according to the optimized deep network feature vector, and learns the grading judgment network according to the optimized combined feature vector and the updated classification label to obtain an optimized grading judgment network.

[0057] Example 1:

[0058] A deep learning tobacco leaf grading method based on expert experience guidance includes the following steps:

[0059] As Figure 2 shown, first, use traditional computer vision methods to extract manual features. Collect tobacco leaf images on the device, and use existing image segmentation techniques to obtain the mask of the tobacco leaf area.

[0060] Find the tobacco leaf contour in the binary mask image. Based on the pixel neighborhood, the tobacco leaf edge can be calculated, and the transition position of the pixel value in the mask image can be regarded as the tobacco leaf contour. To calculate the length and width of the tobacco leaf, the minimum bounding rectangle of the tobacco leaf needs to be obtained. Using the obtained tobacco leaf contour, the minimum convex hull enclosing the tobacco leaf can be obtained. By enumerating the circumscribed rectangles of the convex hull and comparing the areas of the circumscribed rectangles, the minimum bounding rectangle of the tobacco leaf can be obtained. The length of the circumscribed rectangle can represent the length of the tobacco leaf, the width of the circumscribed rectangle can represent the width of the tobacco leaf, and the ratio of the length to the width of the circumscribed rectangle can represent the aspect ratio of the tobacco leaf. Since the tobacco leaf contour has been obtained, the perimeter and area of the tobacco leaf can be directly calculated using the number of pixels in the tobacco leaf contour.

[0061] Using the obtained perimeter and area of the tobacco leaf, the circularity of the tobacco leaf is expressed as:

[0062]

[0063] In the formula, E represents the circularity of the tobacco leaf, A represents the area of the tobacco leaf, and P represents the perimeter of the tobacco leaf.

[0064] Usually, the calculated total area is slightly lower than the area inside the contour. By calculating the ratio of the area of the damaged area inside the tobacco leaf contour to the total area inside the tobacco leaf contour, the damage rate of the tobacco leaf can be obtained, which is expressed as:

[0065]

[0066] In the formula, R represents the damage rate of the tobacco leaf, and S represents the area of the damaged area.

[0067] Under normal circumstances, the images obtained by visible light cameras are divided into three RGB channels. The three channels can be directly separated, and the mean value of each channel image is calculated to obtain the color mean feature of each channel. In addition to directly calculating the RGB three-channel color means, the tobacco leaf images obtained in the RGB color model are converted into images represented by the HSV color model, and the mean value of each channel is calculated to obtain the color mean features of the HSV three channels.

[0068] The features extracted at this stage include the length, width, aspect ratio, perimeter, area, circularity, breakage rate of the tobacco leaves, and the means of six color channels (red, green, blue, hue, saturation, and lightness), a total of 13 kinds. All the manually extracted feature values are combined into a 13-dimensional vector f. m 。

[0069] As Figure 3 shown, the method uses neural network architecture search to determine the best model. The classification accuracy of traditional neural networks on the tobacco leaf dataset is limited, so neural network architecture search is used to determine the best model on a small amount of tobacco leaf data.

[0070] During the neural network architecture search process, the basic parameters of the neural network, namely the input image resolution and network depth, are first determined.

[0071] The search space for the basic convolutional module is {MBConv, Fused-MBConv}, the search space for the convolutional kernel size is {3x3, 5x5}, the search space for the dilation coefficient is {1, 2, 4, 6}, and the search space for the number of channels is {32, 64, 128, 256, 512}.

[0072] During the search process, in order to compare the performance of different structures and select the most suitable model, it is necessary to define a search reward function according to the design requirements. Under normal circumstances, the design goals are: while improving the model accuracy, reducing the model calculation amount, and at the same time reducing the model parameter amount.

[0073] The search reward function can be defined as:

[0074] R = Acc·S w ·P v

[0075] where Acc represents the accuracy, S represents the model calculation time, P represents the model parameter amount, and w, v represent the hyperparameters for weighing the calculation time and parameter amount.

[0076] By finding the model with the largest search reward, the optimal model structure is obtained. During the search process, it is necessary to record the accuracy, calculation time, and model parameter amount, and use the search reward function to calculate the search reward of each model, and select the model with the highest search reward as the deep feature extraction network.

[0077] The neural network architecture search process uses random sampling, that is, randomly generating model parameters and calculating the search reward of the model on the corresponding data set. Every time a model is obtained, its search reward is recorded and the next model parameters are obtained by random sampling.

[0078] Each time a model search is performed, the model parameters are randomly obtained and the search reward is calculated until the search reward meets the requirement or the number of model samples reaches the limit.

[0079] The tobacco leaf image is input into the deep feature extraction network obtained by the neural network architecture search, the deep features in the tobacco leaf image are extracted, and the deep network feature vector f is output d .

[0080] The manual feature vector f m and the deep network feature vector f d Concatenate to get the joint feature vector f t The features calculated by the joint feature extraction module are input into the grading judgment network as the basis for tobacco leaf grading.

[0081] like Figure 4 As shown, the hierarchical judgment network includes an input layer, a hidden layer, and an output layer, wherein the number of neurons in the input layer is equal to the length of the joint feature vector output by the joint feature extraction module, the number of neurons in the output layer is equal to the number of tobacco leaf categories, and the number of neurons in the hidden layer is set based on experience.

[0082] The calculation process of the hierarchical judgment network can be expressed as:

[0083] P = softmax(b (2) +w (2) ·Swish(b (1) +w (1) ·f t ))

[0084] Among them, P represents the predicted probability vector output by the classification judgment module, b (1) , b (2) represents the bias parameter of the neural network, w (1) , w (2) represents the neural network weight parameter, softmax(·) represents the softmax function, and Swish(·) represents the Swish activation function, that is:

[0085]

[0086] The predicted probability vector P output by the neural network is [p1, p2, ..., p N ] T , where p irepresents the probability that the current tobacco leaf is of the i-th type, and N represents the number of tobacco leaf grading categories.

[0087] The predicted category of the tobacco leaf can be obtained and expressed as:

[0088]

[0089] According to the predicted probability vector P, the probability distribution is statistically analyzed, and the probability entropy of the current tobacco leaf prediction is calculated:

[0090]

[0091] Using the calculated probability entropy, the probability of tobacco leaf grading error is judged, and a prompt is made according to the preset probability entropy threshold.

[0092] Let η represent the threshold parameter set by the system. During the grading process, the predicted probability entropy is compared with the set threshold. When E < η, it indicates that the confidence in the grading result is relatively high and expert inspection is not required for the time being. When E ≥ η, the confidence in the grading result is relatively low and experts need to intervene to check the grading result.

[0093] In actual situations, the threshold can be flexibly selected according to the situation. When the threshold η is relatively low, the number of tobacco leaves that need to be inspected for the grading result is relatively large. When the threshold η is relatively high, the number of tobacco leaves that need to be inspected for the grading result is relatively small.

[0094] The expert guidance module of the described method makes full use of the interactivity between humans and machines. Through expert manual intervention, the misjudged and uncertain key attention parts in the attention map are manually marked. Through the updated attention guidance map, the depth feature extraction network's attention to this part is increased to further improve the feature extraction effect.

[0095] For misclassified tobacco leaves, experts can check them and provide the correct grading result to guide the grading judgment network to learn and improve the grading effect.

[0096] Such as Figure 5 As shown, when the attention area of the depth feature extraction network is incorrect, experts can provide an attention guidance map to guide the depth feature extraction network to focus on important areas to improve the feature extraction accuracy. After the experts correct the grading result and features, the parameters of the depth feature extraction network and the grading judgment network are updated using the corrected grading result and features.

[0097] In order to display the attention area of the depth feature extraction network on the basis of grading and rely on experts to correct the attention area, an attention branch needs to be added on the basis of the depth feature extraction network obtained by using the neural network architecture search.

[0098] The calculation process of the depth feature extraction network is expressed as:

[0099] map1 = F(I)

[0100] Where F(·) is a convolutional feature extractor, map1 represents the feature map output by the convolutional feature extractor, and I represents the input tobacco leaf image data.

[0101] The attention branch added to the deep feature extraction network is expressed as:

[0102] M net (I) = sigmoid(conv1(map1))

[0103] map2 = conv2(map1)

[0104] v1 = GAP(map2)

[0105] Where M net (I) represents the output single-channel attention map, sigmoid(·) represents the sigmoid activation function, conv1(·) represents a convolutional layer with the same number of input channels as the feature map and 1 output channel, conv2(·) represents a convolutional layer with the same number of input channels as the feature map and the number of output channels equal to the length of the deep feature vector, and GAP(·) is the global average pooling operation. v1 represents the feature vector output after the global average pooling of the feature map map2.

[0106] The attention map M net (I) will be used to calculate the new feature v2, expressed as:

[0107] map3 = conv3(map1 + M net (I)·map1)

[0108] v2 = GAP(map3)

[0109] Where conv3(·) represents a convolutional layer with the same number of input channels as the feature map and the number of output channels equal to the length of the deep feature vector, map3 represents the feature map output by the convolutional layer conv3(·), and v2 represents the feature vector output after the global average pooling of the feature map map3.

[0110] At this time, the deep feature vector can be expressed as:

[0111] f d = v1 + v2

[0112] Using the attention map output by the attention branch, the reason for the model to make a prediction result can be analyzed.

[0113] Under normal circumstances, only a small area in the tobacco leaf image often plays a decisive role in tobacco leaf grading. Therefore, by checking the attention map, experts can determine whether the attention area of the model is correct. If the model focuses its attention on the wrong area, experts need to correct the attention area of the attention map.

[0114] When asking experts to check the results of the attention area of the attention map, the module requires experts to mark the basis area for this grading case in the form of a pop-up window, forming an attention guidance map M expert (I). Calculate the error between the attention guidance map provided by the expert and the attention map to obtain the attention loss, expressed as:

[0115] L att =α|M expert (I)-M net (I)|

[0116] where α represents the attention guidance weight.

[0117] When experts do not provide an attention guidance map and only use the tobacco leaf classification label for model training, the training process calculates the classification loss using the output of the grading judgment network and the correct grading label of the tobacco leaf, expressed as:

[0118] L c =CE(P,label e )

[0119] where label e represents the correct grading label provided by the expert, and CE(·) represents the cross-entropy loss function.

[0120] When experts provide an attention guidance map, the model training loss function is expressed as:

[0121] L = L att +L c

[0122] Using the correct grading results and attention guidance maps provided by experts, perform online incremental learning on the deep feature extraction network and the grading judgment network respectively.

[0123] During the operation of this method, relying on the guidance of experts, correct the grading results and attention areas, and continuously improve the accuracy of tobacco leaf grading.

[0124] The above is only the preferred embodiment of the present invention. It should be noted that: for those of ordinary skill in the art of this technology, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A deep learning-based tobacco leaf grading method guided by expert experience, characterized in that: It includes the following steps: Obtain the deep network feature vector of the tobacco leaf image by using a pre-set deep feature extraction network; Obtain the manual feature vector of the tobacco leaf image by using traditional computer vision methods; Concatenate the deep network feature vector and the manual feature vector into a joint feature vector; Input the joint feature vector into a pre-set grading judgment network to output the tobacco leaf grading result; Experts re-grade the tobacco leaf images with incorrect grading to obtain updated classification labels, and learn the grading judgment network according to the updated classification labels to obtain an optimized grading judgment network; The deep feature extraction network outputs an attention map. Experts correct the attention area of the attention map corresponding to the tobacco leaf images with incorrect grading to obtain an attention guidance map, learn the deep feature extraction network according to the attention guidance map to obtain an optimized deep feature extraction network, obtain an optimized deep network feature vector according to the optimized deep feature extraction network, obtain an optimized joint feature vector according to the optimized deep network feature vector, and learn the grading judgment network according to the optimized joint feature vector and the updated classification labels to obtain an optimized grading judgment network.

2. The deep learning tobacco leaf grading method based on expert experience guidance according to claim 1, wherein: The method for obtaining a pre-set deep feature extraction network includes the following steps: Determine the neural network basic parameters in the neural network architecture search process. The neural network basic parameters include: image resolution, network depth; Define the search space for neural network architecture search. The search space includes: basic convolution module, convolution kernel size, dilation coefficient, and number of channels; Use the random sampling method to sample the network parameters in the search space and calculate the search reward of the model corresponding to the network parameters; Repeat the network parameter sampling and calculate the search reward of the model corresponding to the network parameters until the model search reward meets the requirements or the number of sampling times reaches the limit; select the network parameters with the highest search reward to construct the corresponding model as the deep feature extraction network.

3. The deep learning tobacco leaf grading method based on expert experience guidance according to claim 1, characterized in that: The manual feature vector includes at least one of the length, width, aspect ratio, perimeter, area, circularity, breakage rate, mean values of the RGB red, green, and blue channels, or mean values of the HSV hue, saturation, and value channels of the tobacco leaf.

4. A deep learning tobacco leaf grading method based on expert experience guidance according to claim 2, characterized in that: The search space of the basic convolution module is {MBConv, Fused-MBConv}, the search space of the convolution kernel size is {3x3, 5x5}, the search space of the dilation coefficient is {1, 2, 4, 6}, and the search space of the number of channels is {32, 64, 128, 256, 512}.

5. A deep learning tobacco leaf grading method based on expert experience guidance according to claim 1, characterized in that: The grading judgment network includes an input layer, a hidden layer, and an output layer. Among them, the number of neurons in the input layer is equal to the length of the joint feature vector output by the joint feature extraction module, and the number of neurons in the output layer is equal to the number of tobacco leaf categories.

6. The deep learning tobacco leaf grading method based on expert experience guidance according to claim 1, wherein: The specific method for inputting the joint feature vector into a pre-set grading judgment network to output the tobacco leaf grading result includes the following steps: Input the joint feature vector into the grading judgment network to output a predicted probability vector; Calculate the probability entropy by using the predicted probability vector; Judge the relationship between the calculated probability entropy and the set threshold; When the probability entropy is less than the set threshold, the grading ends and the tobacco leaf grading result is output; When the probability entropy is greater than the set threshold, expert intervention is required to check the tobacco leaf grading results.

Images