A method, medium and system for intelligent identification of tobacco growth period

Abstract

The invention provides a method, medium and system for intelligently identifying tobacco growth periods, belonging to the technical field of tobacco cultivation, and comprising: obtaining a tobacco leaf color image; performing computer language processing on the tobacco growth period corresponding to the tobacco leaf image to obtain a tobacco growth period label; designing a convolution randomization network framework, comprising an input layer, a hidden layer and an output layer, wherein the hidden layer uses convolution with a step size and an expansion rate to extract small receptive field and large receptive field features; randomly initializing hidden layer weight parameters of the convolution randomization network; using the tobacco growth period labels processed by computer language to train the convolution randomization network to obtain hidden layer random features; substituting the hidden layer random features into a loss function to solve the output layer weight parameters; loading the hidden layer weight parameters and the output layer weight parameters into the convolution randomization network framework; inputting a new tobacco leaf image, and the convolution randomization network outputs the corresponding tobacco growth period prediction result, and outputs it.

Description

Technical Field

[0001] The present invention belongs to the technical field of tobacco cultivation, and in particular, relates to a method, medium and system for intelligently identifying tobacco growth periods. Background Art

[0002] Tobacco is an important economic crop, and its planting and management plays an important role in agricultural production. The management focus and requirements of tobacco at different growth stages are different. In the seedbed seedling stage, it is necessary to pay attention to key indicators such as seed germination and seedling growth; in the field growth stage, it is necessary to focus on processes such as leaf elongation, flowering and fruiting. Traditionally, tobacco planting managers usually rely on manual observation and experience judgment to identify the growth period of tobacco to guide the next step of planting management measures. However, this method that relies on manual experience has many problems such as low efficiency and high cost.

[0003] In order to improve the automation and intelligence level of tobacco planting management, researchers in related fields have begun to try to use computer vision technology to identify the growth period of tobacco in recent years. Specifically, by analyzing the visual features of tobacco leaves such as color and shape, the current growth stage of tobacco can be automatically determined. This tobacco growth period identification method based on machine vision is more efficient and objective than manual observation.

[0004] However, when using traditional methods such as support vector machine (SVM) and BP neural network to identify the tobacco growth period, they are extremely sensitive to changes in the external environment. Often, slight changes in the tobacco growth environment have a great impact on the identification of the tobacco growth period. When using extreme learning machine (ELM) and breadth learning (BLS) to identify the tobacco growth period, due to their limited hidden layer feature capabilities, the recognition effect is often poor. Summary of the invention

[0005] In view of this, the present invention provides a method, medium and system for intelligently identifying the growth period of tobacco, which can solve the technical problem that when the existing technology identifies the growth period of tobacco leaves, slight changes in the tobacco growth environment have a great impact on the identification of the tobacco growth period, resulting in poor identification effect.

[0006] The present invention is achieved in that:

[0007] A first aspect of the present invention provides a method for intelligently identifying tobacco growth stages, comprising the following steps:

[0008] S10, obtaining the collected tobacco leaf color image;

[0009] S20, performing computer language processing on the tobacco growth period corresponding to the tobacco leaf image to obtain a tobacco growth period label for computer recognition;

[0010] S30. Design a convolutional randomized network framework, including input layer, hidden layer and output layer. The hidden layer uses convolution with step size and dilation rate to extract small receptive field and large receptive field features.

[0011] S40, randomly initialize the hidden layer weight parameters of the convolution randomization network;

[0012] S50, using the tobacco growth period labels processed by computer language to train the convolutional randomization network to obtain hidden layer random features;

[0013] S60, substituting the hidden layer random features into the loss function to solve the output layer weight parameters;

[0014] S70, loading the hidden layer weight parameters and the output layer weight parameters into the convolution randomization network framework;

[0015] S80. Input a new tobacco leaf image, and the convolutional randomization network outputs the corresponding tobacco growth period prediction result and outputs it.

[0016] Among them, the color image of tobacco leaves is collected using a CCD sensor.

[0017] The step S10 specifically comprises: using a CCD imaging sensor to obtain a color image of the tobacco leaf. The CCD sensor uses its photoelectric conversion function to convert the reflected light signal on the surface of the tobacco leaf into a digital electrical signal, forming color image data containing visual information such as the color and shape of the tobacco leaf, providing input for subsequent identification of the tobacco growth period.

[0018] Among them, the step S20 specifically includes: performing computer language processing on the tobacco growth period corresponding to the acquired tobacco leaf color image to obtain a tobacco growth period label for computer recognition. Specifically, the tobacco growth process is divided into a seedbed stage and a field stage, represented by 0 and 1 respectively; each stage is further subdivided into 4 periods, and the encoding methods of 0-1, 0-2, 0-3, 0-4 and 1-1, 1-2, 1-3, 1-4 are used to convert it into an 8-dimensional binary vector.

[0019] Wherein, the step S30 specifically includes: designing a convolutional randomized network framework, including an input layer, a hidden layer and an output layer, wherein the hidden layer uses convolution kernels with different step sizes and expansion rates to extract small receptive field and large receptive field features of the tobacco leaf image; the hidden layer of the network framework is composed of two parallel branches, the upper branch uses a 3×3 convolution kernel with a step size of 1 to extract local detail features, and the lower branch uses a 3×3 convolution kernel with a step size of 2 and a expansion rate of 2 to extract global semantic features, and the two are cascaded to form multi-scale hidden layer features.

[0020] The step S40 specifically includes: randomly initializing the hidden layer weight parameters of the convolutional randomized network. Since the convolutional randomized network is a network structure with fast parameter learning, its hidden layer weight parameters do not need to be obtained through a complex training process, but are directly randomly initialized. This random initialization method not only simplifies the training process, but also can enhance the generalization ability of the network and improve its adaptability to environmental changes.

[0021] Among them, the step S50 specifically includes: using the tobacco growth period labels after computer language processing in step S20 to train the convolutional randomized network to obtain hidden layer random features; the training process includes: inputting the tobacco leaf image into the network, calculating the hidden layer features according to the randomly initialized hidden layer weights; using the corresponding tobacco growth period labels as supervision signals, and learning to obtain the random feature representation of the hidden layer by minimizing the variance loss function.

[0022] Among them, the step S60 is specifically: substituting the hidden layer random features into the loss function to solve the output layer weight parameters, wherein the variance is used as the loss function and the least squares method is used to calculate the optimal weight parameters of the output layer.

[0023] This method of quickly calculating the output layer parameters avoids the complex gradient descent training process of traditional network structures.

[0024] Among them, the step S70 is specifically: loading the hidden layer weight parameters randomly initialized in step S40 and the output layer weight parameters calculated in step S60 into the hidden layer and output layer of the convolutional randomized network respectively, to complete the determination of the entire network parameters.

[0025] Among them, the step S80 specifically includes: inputting the new tobacco leaf color image into the trained convolutional randomized network, using the network parameters loaded in step S70, and outputting the corresponding tobacco growth period prediction result after feature extraction and prediction.

[0026] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the above-mentioned method for intelligent identification of tobacco growth stages.

[0027] A third aspect of the present invention provides a tobacco growth period intelligent identification system, which includes the above-mentioned computer-readable storage medium.

[0028] Compared with the prior art, the beneficial effects of the tobacco growth period intelligent identification method, medium and system provided by the present invention are:

[0029] 1. Design of a convolutional randomized network structure: The present invention adopts a convolutional randomized network (CRN) framework to automatically extract multi-scale feature information from tobacco leaf images. The network structure contains two parallel convolution branches, one with a small receptive field and the other with a large receptive field, which can fully describe the apparent characteristics of tobacco at different growth stages from two dimensions: local details and global semantics. Compared with the traditional fully connected randomized network, CRN has the advantages of fewer parameters and lower computational complexity.

[0030] 2. Enhanced the adaptability of the network to environmental changes: The hidden layer weight parameters of CRN are randomly initialized, without the need for complex gradient descent training. This simple and fast parameter learning method greatly improves the generalization ability of the network, making it more robust to changes in the tobacco growth environment.

[0031] 3. Improved the ability to express tobacco growth period characteristics: The parallel convolution structure of CRN can extract feature information of tobacco leaf images from multiple scales and dimensions. This expressive feature representation enhances the network's ability to recognize different growth periods of tobacco, laying the foundation for subsequent accurate prediction.

[0032] 4. Realize fast and efficient parameter learning: The present invention uses the least squares method to calculate the output layer parameters of CRN, which greatly simplifies the network training process. Compared with the complex gradient descent optimization, this fast parameter learning method not only speeds up the model convergence speed, but also has stronger advantages in practical application deployment.

[0033] In summary, the CRN-based intelligent identification method for tobacco growth periods proposed in the present invention has significant improvements over the existing technology in terms of environmental adaptability, feature expression ability and parameter learning efficiency, and solves the technical problem that when the existing technology is used to identify the growth period of tobacco leaves, slight changes in the tobacco growth environment have a great impact on the identification of the tobacco growth period, resulting in poor identification effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0035] Figure 1 This is a computer language diagram of the tobacco growth period;

[0036] Figure 2 This is a schematic diagram of the convolutional randomized network framework;

[0037] Figure 3 A flow chart of the method provided by the present invention. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0039] like Figure 3 FIG. 1 is a flow chart of a method for intelligently identifying tobacco growth periods provided by the present invention. The method comprises the following steps:

[0040] S10, obtaining a color image of tobacco leaves collected by a CCD sensor;

[0041] S20, performing computer language processing on the tobacco growth period corresponding to the tobacco leaf image to obtain a tobacco growth period label that can be recognized by a computer;

[0042] S30. Design a convolutional randomized network framework, including input layer, hidden layer and output layer. The hidden layer uses convolution with step size and dilation rate to extract small receptive field and large receptive field features.

[0043] S40, randomly initialize the hidden layer weight parameters of the convolution randomization network;

[0044] S50, using the tobacco growth period labels processed by computer language to train the convolutional randomization network to obtain hidden layer random features;

[0045] S60, substituting the hidden layer random features into the loss function to solve the output layer weight parameters;

[0046] S70, loading the hidden layer weight parameters and the output layer weight parameters into the convolution randomization network framework;

[0047] S80. Input a new tobacco leaf image, and the convolutional randomization network outputs the corresponding tobacco growth period prediction result and outputs it.

[0048] The specific implementation methods of the above steps are described in detail below:

[0049] The specific implementation method of step S10 is: using a CCD imaging sensor to obtain a color image of tobacco leaves. A CCD (Charge Coupled Device) imaging sensor is a photoelectric device that converts light signals into electrical signals and is widely used in the fields of digital cameras, video surveillance, etc. In the present invention, a CCD sensor is used to collect color images of tobacco leaves. Specifically, tobacco leaves are placed in an imaging area, and the CCD sensor uses its photoelectric conversion function to convert the reflected light signals of the tobacco leaves into electrical signals to form a digital color image of the tobacco leaves. The color image contains rich visual information such as the color and shape of the tobacco leaves, and provides important input data for the subsequent identification of the tobacco growth period.

[0050] The specific implementation of step S20 is: the tobacco growth period corresponding to the obtained tobacco leaf color image is processed by computer language to obtain a tobacco growth period label that can be recognized by the computer. The growth process of tobacco mainly includes the seedbed stage and the field stage, and each stage can be subdivided into 4 periods. In order to enable the computer to recognize and process the tobacco growth period information, it needs to be converted into a language form that can be understood by the computer. Specifically, the seedbed stage is represented by 0, and the field stage is represented by 1; the four periods of the seedbed stage are represented by 0-1, 0-2, 0-3, and 0-4 respectively; the four periods of the field stage are represented by 1-1, 1-2, 1-3, and 1-4 respectively. In this way, each tobacco growth period can be represented by an 8-dimensional binary vector, and the position with a value of 1 in the vector represents the category of the period. For example, the vector [0,0,0,0,1,0,0,0] represents the seedling slowing period of the field stage, and the vector [0,0,1,0,0,0,0,0] represents the rooting period of the seedbed stage. In this way, the tobacco growth period is converted into a computer-recognizable label form, providing basic data for subsequent intelligent identification.

[0051] The specific implementation of step S30 is: designing a convolutional randomized network framework, including an input layer, a hidden layer and an output layer, wherein the hidden layer uses convolution with a step size and a dilation rate to extract small receptive field and large receptive field features. The convolutional randomized network is a network structure with fast parameter learning, which can effectively extract local features of input data. In the present invention, the network framework mainly consists of three parts:

[0052] 1) Input layer: receiving the tobacco leaf color image data from step S10.

[0053] 2) Hidden layer: It consists of two parallel branches. The upper branch uses a 3×3 convolution kernel with a step size of 1 to extract the two-dimensional small receptive field features of the tobacco leaf image, which contains more underlying detail information. The lower branch uses a 3×3 convolution kernel with a step size of 2 and a dilation rate of 2 to extract the two-dimensional large receptive field features of the tobacco leaf image, which contains more high-level semantic information. The convolution features of the two branches are flattened and cascaded to form hidden layer features with multi-scale information. This parallel structure can enhance the network's ability to express the characteristics of the tobacco growth period.

[0054] 3) Output layer: receives the multi-scale features extracted by the hidden layer and outputs the corresponding tobacco growth period prediction results.

[0055] The convolutional randomized network framework designed in this way can effectively extract the features of tobacco leaf images at different receptive field scales, providing rich input information for subsequent growth period identification.

[0056] The specific implementation of step S40 is: randomly initialize the hidden layer weight parameters of the convolutional randomized network. The convolutional randomized network is a network structure with fast parameter learning, and its hidden layer weight parameters are randomly generated without the need for training through a complex gradient descent algorithm. Specifically, the hidden layer weight parameters ω of the convolutional randomized network are initialized randomly. j,k Random initialization, where j represents the index of the hidden layer node and k represents the index of the input feature. This random initialization method not only simplifies the training process, but also enhances the generalization ability of the network and improves its adaptability to environmental changes.

[0057] The specific implementation of step S50 is: using the tobacco growth period labels processed by computer language in step S20 to train the convolutional randomization network to obtain hidden layer random features. The training process is as follows:

[0058] 1) The tobacco growth period label t obtained in step S20 i As a supervision signal for the network.

[0059] 2) Transform the tobacco leaf image x i,j Input to the input layer of the convolutional randomization network.

[0060] 3) According to the hidden layer weight parameter ω randomly initialized in step S40 j,k , using formula (1) to calculate the random feature h of the hidden layer i,k :

[0061] h i,k =g(x i,j ·ω j,k +b i,k )

[0062] Among them, g(·) is the activation function, b i,k is the bias term.

[0063] 4) The hidden layer random feature h i,k After further processing, the final hidden layer output H is obtained, as shown in equations (2)-(5).

[0064] Through this training process, the network can learn the hidden layer random characteristics of the tobacco growth period, laying the foundation for the subsequent output layer parameter calculation.

[0065] The specific implementation of step S60 is: substitute the hidden layer random features into the loss function to solve the output layer weight parameter. Since the hidden layer weights of the convolutional randomized network are randomly initialized, it is only necessary to calculate the weight parameter β of the output layer. Specifically, the variance is used as the loss function, as shown in formula (6):

[0066]

[0067] Where H is the hidden layer random feature calculated in step S50, and T is the real tobacco growth period label. By minimizing the loss function, the optimal output layer weight parameter β can be obtained, as shown in formula (7):

[0068] β=(H T ·H) -1 ·H T T

[0069] In this way, by solving the output layer parameter β, the parameter learning process of the convolutional randomized network is completed.

[0070] The specific implementation of step S70 is: loading the hidden layer weight parameters and the output layer weight parameters into the convolution randomization network framework. Specifically, the hidden layer weight parameters ω randomly initialized in step S40 are j,k The output layer weight parameter β calculated in step S60 is loaded into the hidden layer and output layer of the convolutional randomization network respectively. In this way, the parameters of the entire convolutional randomization network have been determined and can be used for subsequent tobacco growth period prediction.

[0071] The specific implementation of step S80 is: input a new tobacco leaf image, and the convolution randomization network outputs the corresponding tobacco growth period prediction result. Specifically:

[0072] 1) Input the new tobacco leaf color image into the input layer of the convolutional randomization network.

[0073] 2) Based on the network framework designed in step S30, using the hidden layer weight parameter ω loaded in step S70 j,kand the output layer weight parameter β, after hidden layer feature extraction and output layer prediction, the final tobacco growth period prediction result is calculated, as shown in formula (8):

[0074] Hβ=H·H T ·T

[0075] Among them, H T H T Represents the transpose operation of the hidden layer feature information matrix H.

[0076] 3) Output the prediction result, that is, the current growth period of tobacco.

[0077] Through the above steps, the newly input tobacco leaf images can be intelligently recognized and their growth period can be predicted. The entire recognition process relies on the designed convolutional randomized network framework, which has strong robustness and generalization ability.

[0078] In general, the tobacco growth period intelligent identification method proposed in the present invention mainly includes three key steps: 1) Computer languageization of tobacco growth period data set, converting growth period information into a computer-recognizable form; 2) Designing a convolution randomized network framework, using convolutions with different receptive fields to extract multi-scale features; 3) Using the least squares method to calculate the output layer parameters of the network, to achieve fast parameter learning. This method overcomes the problems of traditional algorithms being sensitive to environmental changes and having limited feature expression capabilities, and shows good performance in the task of tobacco growth period intelligent identification.

[0079] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the above-mentioned method for intelligent identification of tobacco growth stages.

[0080] A third aspect of the present invention provides a tobacco growth period intelligent identification system, which includes the above-mentioned computer-readable storage medium.

[0081] Specifically, the principle of the present invention is: a convolutional randomized network (CRN) framework is designed, and a fast and efficient parameter learning algorithm is used to achieve accurate prediction of tobacco growth period. The reason why this method can solve the problems in the prior art is mainly based on the following technical principles:

[0082] 1. Multi-scale feature extraction: The hidden layer of CRN consists of two parallel convolution branches, which use different convolution kernel parameters to extract small receptive field and large receptive field features of tobacco leaf images. The small receptive field convolution kernel can capture the detailed information of the image, while the large receptive field convolution kernel can extract global semantic features. By fusing these two different scale features, CRN can more comprehensively and deeply describe the apparent characteristics of tobacco at different growth stages. This multi-scale feature extraction method overcomes the problem of limited feature expression ability in the existing technology.

[0083] 2. Randomly initialize the hidden layer: Unlike traditional deep learning networks based on gradient descent, the hidden layer weight parameters of CRN are randomly initialized, without the need for a complex training process. This simple and fast parameter setting method gives CRN strong generalization ability and environmental adaptability. Even in the face of changes in the tobacco growth environment, CRN can maintain a relatively stable recognition performance, overcoming the defect of existing technologies that are sensitive to environmental changes.

[0084] 3. Least squares parameter learning: For CRN, since the hidden layer weights are randomly generated, only the weight parameters of the output layer need to be calculated. The present invention uses the least squares method to solve the output layer parameters, which greatly simplifies the network training process. Compared with the traditional training method based on gradient descent, the least squares method has the characteristics of high computational efficiency and fast convergence. This ability of fast parameter learning makes CRN have more significant advantages in practical applications.

[0085] In general, the intelligent identification method of tobacco growth period proposed in the present invention makes full use of the technical advantages of computer vision and machine learning. A CRN framework that can adapt to changes in the tobacco planting environment is designed, and a fast and efficient parameter learning algorithm is used to achieve accurate prediction of tobacco growth period. This method not only overcomes the limitations of existing technologies, but also has better performance in practical applications.

[0086] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

[0087] A specific embodiment of the present invention is provided below: This embodiment collects a data set containing tobacco leaf images at different growth stages, and compares the performance of the CRN method with several existing representative algorithms in the tobacco growth stage recognition task.

[0088] 1. Experimental Dataset

[0089] The experimental dataset includes 1200 color images of tobacco leaves from different planting areas. These images cover the entire growth process of tobacco, including the emergence, cross, rooting and seedling stages in the seedbed stage, and the seedling stage, root extension stage, vigorous growth stage and maturity stage in the field stage, a total of 8 growth stages. There are 150 images for each growth stage. The resolution of all images is 640×480 pixels, with 3 color channels.

[0090] According to the method described in step S20, the growth period label corresponding to each tobacco leaf image is converted into an 8-dimensional binary vector representation. Specifically, the four periods of the seedbed stage are represented by [1,0,0,0,0,0,0,0], [0,1,0,0,0,0,0,0], [0,0,1,0,0,0,0,0] and [0,0,0,1,0,0,0,0] respectively; the four periods of the field stage are represented by [0,0,0,0,1,0,0,0], [0,0,0,0,0,1,0,0], [0,0,0,0,0,0,1,0] and [0,0,0,0,0,0,0,1] respectively.

[0091] The entire dataset is randomly divided into a training set and a test set, where the training set contains 900 images and the test set contains 300 images.

[0092] 2. Experimental Setup

[0093] The CRN method proposed in the present invention is compared with the following representative tobacco growth period recognition algorithms:

[0094] (1) Support Vector Machine (SVM): Use SVM with radial basis kernel function for classification.

[0095] (2) BP neural network: a standard BP neural network with one hidden layer.

[0096] (3) Extreme Learning Machine (ELM): ELM algorithm using random hidden layer parameters.

[0097] (4) Breadth Learning (BLS): Use the width learning network structure for feature learning and classification.

[0098] For the above four comparison algorithms, the same tobacco leaf image data was used for training and testing. In order to make a fair comparison, the hyperparameter settings of each algorithm were optimized as much as possible to obtain the best recognition performance.

[0099] For the CRN method proposed in this invention, the network structure parameters are set as follows: small receptive field convolution kernel C s Contains 100 3×3 filters, L s =100; large receptive field convolution kernel C lContains 200 3×3 filters, L l =200. Hidden layer weight parameter C s and C l And the bias parameter B s and B l All are randomly initialized to obey The value of .

[0100] The performance of various algorithms in the tobacco growth period recognition task was evaluated by a 5-fold cross-validation method. Specifically, the training set was randomly divided into 5 parts, 4 of which were used as training data each time, and the remaining 1 part was used as validation data, and then the recognition accuracy of the validation was recorded. The final recognition accuracy was the average of the 5 validation results.

[0101] 3. Experimental results and analysis

[0102] (1) Comparison of recognition accuracy

[0103] Table 1 shows the average recognition accuracy results of various algorithms on the tobacco growth period recognition task. It can be seen that the CRN method proposed in the present invention has achieved the highest recognition accuracy, reaching 91.3%, which has obvious performance advantages over other algorithms.

[0104] Table 1 Average recognition accuracy of different algorithms in tobacco growth period recognition task

[0105] algorithm Average recognition accuracy SVM 82.7％ BP Neural Network 85.2％ ELM 86.5％ BLS 88.1％ CRN of the present invention 91.3％

[0106] This result shows that compared with traditional algorithms based on SVM and neural networks, the CRN method proposed in this invention has stronger performance in the task of identifying tobacco growth periods. The main reasons are:

[0107] 1) CRN can effectively extract the feature information of tobacco leaf images at different scales, including detail features and semantic features, so as to better describe the apparent characteristics of tobacco at different growth stages.

[0108] 2) The hidden layer weights of CRN are randomly initialized, and no complicated training process is required. This simple and fast parameter learning method enhances the generalization ability of the network and improves its adaptability to environmental changes.

[0109] 3) The least squares method is used to calculate the output layer parameters, which greatly simplifies the training process of CRN and gives it a stronger advantage in practical application deployment.

[0110] In contrast, traditional algorithms such as SVM and BP neural network are sensitive to environmental changes and are difficult to fully extract the complex characteristics of tobacco growth period, resulting in weak recognition performance. Although ELM and BLS also use the idea of ​​random hidden layer, they cannot achieve the excellent performance of CRN due to their limited feature extraction capabilities.

[0111] (2) Analysis of recognition results

[0112] The recognition effect of CRN on different reproductive periods was further analyzed.

[0113] As can be seen from the figure, the recognition accuracy of CRN in the four growth stages (seedling stage, root extension stage, vigorous growth stage and maturity stage) in the field stage is above 90%, which is an excellent performance. In the seedbed stage, the recognition accuracy of CRN is relatively lower, but also above 85%.

[0114] The reason why CRN has better recognition effect in the field stage is mainly because the tobacco leaf features at this stage are relatively richer and more stable, which helps the network extract more discriminative features. In the seedbed stage, due to the large changes in tobacco leaf morphology and relatively complex features, the learning and generalization ability of the network is affected to a certain extent. However, CRN can still achieve high recognition accuracy with its advantages of multi-scale feature extraction and fast parameter learning.

[0115] In general, through the comparative analysis of the above experimental results, it can be seen that the CRN-based intelligent recognition method for tobacco growth periods proposed in the present invention is significantly superior to several existing representative algorithms in terms of recognition accuracy, environmental adaptability and parameter learning efficiency.

Claims

1. A method for intelligently identifying tobacco growth periods, characterized in that: The following steps are involved: S10, obtaining the collected tobacco leaf color image; S20, performing computer language processing on the tobacco growth period corresponding to the tobacco leaf image to obtain a tobacco growth period label for computer recognition; S30. Design a convolutional randomized network framework, including input layer, hidden layer and output layer. The hidden layer uses convolution with step size and dilation rate to extract small receptive field and large receptive field features. S40, randomly initialize the hidden layer weight parameters of the convolution randomization network; S50, using the tobacco growth period labels processed by computer language to train the convolutional randomization network to obtain hidden layer random features; S60, substituting the hidden layer random features into the loss function to solve the output layer weight parameters; S70, loading the hidden layer weight parameters and the output layer weight parameters into the convolution randomization network framework; S80, input a new tobacco leaf image, and the convolution randomization network outputs the corresponding tobacco growth period prediction result and outputs it; Wherein, the step S30 specifically includes: designing a convolution randomized network framework, including an input layer, a hidden layer and an output layer, wherein the hidden layer uses convolution kernels with different step sizes and dilation rates to extract small receptive field and large receptive field features of the tobacco leaf image; the hidden layer of the network framework is composed of two parallel branches, the upper branch uses a 3×3 convolution kernel with a step size of 1 to extract local detail features, and the lower branch uses a 3×3 convolution kernel with a step size of 2 and a dilation rate of 2 to extract global semantic features, and the two are cascaded to form multi-scale hidden layer features; The step S50 specifically includes: using the tobacco growth period labels processed by computer language in step S20 to train the convolutional randomized network to obtain hidden layer random features; the training process includes: inputting the tobacco leaf image into the network, calculating the hidden layer features according to the randomly initialized hidden layer weights; using the corresponding tobacco growth period labels as supervision signals, and learning to obtain the random feature representation of the hidden layer by minimizing the variance loss function; The step S70 specifically includes: loading the hidden layer weight parameters randomly initialized in step S40 and the output layer weight parameters calculated in step S60 into the hidden layer and output layer of the convolutional randomized network respectively, thereby completing the determination of the parameters of the entire network; The step S80 specifically includes: inputting the new tobacco leaf color image into the trained convolutional randomized network, using the network parameters loaded in step S70, extracting and predicting features, and outputting the corresponding tobacco growth period prediction result.

2. The method for intelligently identifying tobacco growth periods according to claim 1, characterized in that: The color images of tobacco leaves are collected using a CCD sensor.

3. The method for intelligently identifying tobacco growth periods according to claim 1, characterized in that: The step S20 specifically includes: performing computer language processing on the tobacco growth period corresponding to the acquired tobacco leaf color image to obtain a tobacco growth period label for computer recognition. Specifically, the tobacco growth process is divided into a seedbed stage and a field stage, which are represented by 0 and 1 respectively; each stage is further subdivided into 4 periods, and the encoding methods of 0-1, 0-2, 0-3, 0-4 and 1-1, 1-2, 1-3, 1-4 are used to convert it into an 8-dimensional binary vector.

4. The method for intelligently identifying tobacco growth stages according to claim 1, characterized in that: The step S60 specifically comprises: substituting the hidden layer random features into the loss function to solve the output layer weight parameters, wherein the variance is used as the loss function and the least squares method is used to calculate the optimal weight parameters of the output layer.

5. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the tobacco growth stage intelligent identification method according to any one of claims 1 to 4.

6. An intelligent identification system for tobacco growth period, characterized in that: Contains the computer-readable storage medium of claim 5.

Images