An improved MobileNetV3-Large-based method for intelligently identifying field maturity of flue-cured tobacco leaves

Abstract

The application discloses a kind of based on improved MobileNetV3-Large's tobacco leaf field maturity intelligent discrimination method for flue-cured tobacco, applied to intelligent terminal, and the image of flue-cured tobacco leaf to be discriminated field maturity is collected by photographing function, and improved MobileNetV3-Large discrimination model is used to carry out flue-cured tobacco leaf field maturity discrimination processing to the image of flue-cured tobacco leaf;Wherein, the improved MobileNetV3-Large discrimination model takes MobileNetV3-Large as base model, and introduces dynamic convolution module, multilayer perception machine, jump connection branch and SimAM module in base model.The application can enhance image deep feature extraction and channel and position information perception, so that flue-cured tobacco leaf field maturity recognition result is not limited to production area and position, and the discrimination accuracy and speed of flue-cured tobacco leaf field maturity are higher.

Description

Technical Field

[0001] This invention relates to a method for determining the field maturity of flue-cured tobacco leaves, specifically, to an intelligent method for determining the field maturity of flue-cured tobacco leaves based on an improved MobileNetV3-Large. Background Technology

[0002] Flue-cured tobacco is an important economic crop in my country. Rapid and accurate determination of the field maturity of flue-cured tobacco leaves is a prerequisite and foundation for scientific harvesting, and is of great significance for improving the quality and increasing the yield of flue-cured tobacco leaves. Currently, the determination of field maturity of flue-cured tobacco leaves is mainly done manually, which has the advantages of being fast and accurate, but also suffers from drawbacks such as strong subjectivity and reliance on experience.

[0003] Patent CN103245625A discloses a non-destructive testing method for the maturity of fresh flue-cured tobacco leaves. It uses a chlorophyll meter to measure the relative chlorophyll content (SPAD value) of the first 1-3 leaves to be harvested from bottom to top, and a protractor to measure the angle between the stem and petiole. The maturity of the tobacco leaves is then determined based on the measured values. This method has the advantages of being objective, non-destructive, and easy to operate. Patent CN103278458A discloses a rapid non-destructive testing method for the harvest maturity of flue-cured tobacco. It uses a portable colorimeter to measure the L*a*b* or / and L*C*h* color values ​​of the first 1-3 leaves (from bottom to top) to quantify the range of color value variation for leaves at different maturity levels. This method also has the advantages of being objective, non-destructive, and easy to operate. Patent CN109540894A discloses a non-destructive and rapid method for detecting the maturity of flue-cured tobacco leaves. It acquires color images of the tobacco leaves using a digital camera, processes the images using a computer to obtain the R-value of the detected portion, and substitutes this R-value into the maturity calculation formula: Maturity Value = 18*R / 255 to obtain the maturity value of the detected portion of the tobacco leaf. This method has the advantages of being objective, non-destructive, and rapid. Patent CN108198176A discloses a method for determining tobacco maturity based on image analysis. It acquires images of tobacco leaves at close range using a digital camera or other high-resolution image acquisition device, preprocesses the images using a computer, and extracts the R, G, and B values. The PEN correlation analysis method is used to select the maturity parameter as: MT = (BR) / (BG). The maturity is determined by the MT value of the image, which also has the advantages of being objective, non-destructive, and rapid. However, all four patented methods mentioned above require chlorophyll meters, color difference meters, computers, or digital cameras, which are relatively expensive and inconvenient to operate in the field.

[0004] Patent CN105004722A discloses a rapid method for detecting the maturity of tobacco leaves. It utilizes the photography and image processing functions of widely used smartphones to take photos of tobacco leaves on-site. After image preprocessing using the smartphone's image processing function, the frequency values ​​h4-h6 of the tobacco leaf image's hue in the H4-H6 hue region are extracted. The maturity is then determined based on a linear discriminant analysis (LDA) model based on three feature parameters (h4, h5, and h6) of the image hue frequency, which is pre-sampled, regressed, and stored in the smartphone. The method provides a maturity value for the photographed tobacco leaves and has the advantages of being objective, non-destructive, low-cost, and convenient for field operation. Patent CN114609134A discloses a mobile phone-based intelligent method for judging the field maturity of flue-cured tobacco leaves based on linear discriminant analysis. The method involves obtaining images of flue-cured tobacco leaves with the desired field maturity by taking photos with a mobile phone. First, the images undergo background removal preprocessing using a maximum inter-class variance algorithm with a threshold of 0.20–0.40 on the mobile phone. Next, the image feature extraction function on the mobile phone is used to obtain ten feature parameters: red (R), green (G), blue (B), hue (H), saturation (S), brightness (V), energy (ASM), entropy (ENT), moment of inertia (INE), and correlation (CORRL). Then, the ten feature parameters are used as input to a linear discriminant model to judge the field maturity of the flue-cured tobacco leaves. Finally, the results are displayed on the mobile phone. This method has advantages such as being objective and non-destructive, low-cost, and convenient for field operation. Patent CN114609135A discloses a mobile phone-based intelligent method for judging the field maturity of flue-cured tobacco leaves based on a BP neural network. It obtains images of flue-cured tobacco leaves to be judged for field maturity through mobile phone photography and image processing, along with ten feature parameters: red (R), green (G), blue (B), hue (H), saturation (S), brightness (V), energy (ASM), entropy (ENT), moment of inertia (INE), and correlation (CORRL). The method uses a 5-7 BP neural network model mounted on the mobile phone to judge the field maturity of the flue-cured tobacco leaves. Finally, the results are displayed on the mobile phone. This method has advantages such as being objective and non-destructive, low-cost, and convenient for field operation. However, all three patents mentioned above use traditional learning algorithms based on manual feature extraction. Manual extraction is time-consuming and requires certain professional knowledge and experience, which limits the accuracy, precision, and speed of the system.

[0005] With the maturity and popularization of deep learning technology, and driven by massive amounts of data and rich application scenarios, deep convolutional networks, represented by convolutional neural networks, have gradually replaced traditional algorithms based on manual feature extraction in the machine learning era. Among them, MobileNetV3-Large, a lightweight network model proposed by the Google team in 2019, adopts new technologies including the Squeeze-and-Excitation module for channel attention and NAS search method. Compared with MobileNetV2, MobileNetV3-Large improves the accuracy of ImageNet subclassing by 3.2% while reducing latency by 20%, and has achieved excellent performance in mobile image classification, object detection, semantic segmentation and other tasks.

[0006] However, for flue-cured tobacco leaves, color is the most direct visual indicator and a straightforward reflection of their internal quality. Studies have shown that tobacco leaves darken with increasing maturity, thus maturity can be classified based on color. However, flue-cured tobacco leaves from different production areas, different parts of the plant, and at different stages of maturity may exhibit similar color characteristics. Therefore, developing a fast, accurate, and region- and part-specific field maturity determination method for flue-cured tobacco leaves based on MobileNetV3-arge is a pressing issue.

[0007] In order to solve the above problems, people have been seeking an ideal technological solution. Summary of the Invention

[0008] The purpose of this invention is to address the shortcomings of existing technologies by providing a method for determining the field maturity of flue-cured tobacco leaves that is fast, accurate, objective, non-destructive, low-cost, easy to operate in the field, and not limited by production area or part of the plant.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] The first aspect provides an intelligent method for determining the field maturity of flue-cured tobacco leaves based on an improved MobileNetV3-Large, such as... Figure 1 As shown, the intelligent method for determining the field maturity of flue-cured tobacco leaves, applied to smart terminals, includes the following steps:

[0011] Acquire images of flue-cured tobacco leaves whose field maturity needs to be determined;

[0012] An improved MobileNetV3-Large discriminant model was used to perform field maturity discrimination processing on flue-cured tobacco leaf images.

[0013] The improved MobileNetV3-Large discriminative model is based on the MobileNetV3-Large model. The main framework of the MobileNetV3-Large model includes 15 bottleneck structures. The first three bottleneck structures and bottleneck structures 7-12 use 3x3 depthwise convolutions, while bottleneck structures 4-6 and the last three use 5x5 depthwise convolutions. All other settings within the bottleneck structures are the same. Specifically, bottleneck structures 4-6 and 11-15 introduce SE modules, allowing the network to automatically learn the weights for each channel to optimize the feature map data, thus achieving a significant performance improvement with relatively low computational overhead.

[0014] Among them, the convolution method and attention mechanism are the core factors affecting the performance of the MobileNetV3-Large model.

[0015] Regarding the first core factor, MobileNetV3-Large uses the Depthwise Separable Convolution (DSC) module to reduce the model's parameters to achieve lightweighting, but this also causes the loss of target features to some extent, leading to a decrease in model accuracy.

[0016] To improve the model's ability to extract subtle differences in field maturity images of flue-cured tobacco leaves, this invention introduces a dynamic convolution module into the original depthwise separable dynamic convolution (DSC) module of the MobileNetV3-Large model, namely the Depthwise Separable Dynamic Convolution (DSDC) module (as shown in Figure 2). This dynamic convolution module dynamically adjusts the model's convolution kernel parameters based on the input image.

[0017] Specifically, the dynamic convolution module uses depthwise convolution to perform convolution operations on each channel of the input feature map individually, using different convolution kernels to extract different features on each channel; then, the feature map obtained by depthwise convolution is dynamically and adaptively generated by the attention module in dynamic convolution to generate weight coefficients for each convolution kernel. After multiplying each weight coefficient with the corresponding convolution and summing the results to obtain the final weight coefficient, it is convolved with the feature map obtained by depthwise convolution. After batch normalization and ReLU activation function, dynamic adaptive convolution is completed.

[0018] To verify the effectiveness of the dynamic convolution module, this embodiment uses tobacco leaf samples from tobacco fields in Shewan, Pangu, Guoji, and Wangdian towns in Biyang County, Zhumadian. After the tobacco plants entered the maturity stage, tobacco leaves from the lower, middle, and upper parts, whose field maturity was manually determined by experts, were selected. Using the "image capture" function of a mobile phone or other smart terminal, a total of 6,562 images of tobacco leaves at different field maturity levels were collected. Among them, there were 2,207 images of immature tobacco leaves, 2,167 images of mature tobacco leaves, and 2,188 images of overripe tobacco leaves. The immature, mature, and overripe tobacco leaf images were labeled as "1", "2", and "3", respectively.

[0019] The 6562 flue-cured tobacco leaf images were divided into a training set and a validation set in a 7:3 ratio. The training set consisted of 4599 tobacco leaf images, including 1551 images of immature tobacco leaves, 1526 images of mature tobacco leaves, and 1522 images of overripe tobacco leaves. The validation set consisted of 1963 tobacco leaf images, including 656 images of immature tobacco leaves, 641 images of mature tobacco leaves, and 666 images of overripe tobacco leaves.

[0020] The selected deep learning framework was PyTorch 1.10, the programming language was Python, the environment was Python 3.8, and the integrated development environment was PyCharm 2020. The computer configuration for running the program was an Intel® Core® Gold 5320 CPU @ 2.20GHz, 32GB of RAM, an NVIDIA RTX A4000 GPU with 16GB of VRAM, and a 64-bit Windows 10 operating system. The model was trained for 200 epochs, with a uniform batch size of 16. The loss function was classification cross-entropy, and the model was trained using stochastic gradient descent. The training parameters—learning rate, weight decay, and momentum—were set to 0.001, 0.00001, and 0.9, respectively. A learning rate decay strategy was implemented, reducing the learning rate to 80% every 20 epochs. Figure 3 shows the accuracy and speed of the MobileNetV3-Large-DSC model without dynamic convolution and the MobileNetV3-Large-DSDC model with dynamic convolution for determining the field maturity of flue-cured tobacco leaves on the validation set.

[0021] from Figure 3It can be seen that the MobileNetV3-Large-DSDC model, after introducing dynamic convolution into depthwise separable convolution, can dynamically generate the model's convolution kernel parameters according to different input images to extract subtle differences in flue-cured tobacco leaf images. Although the discrimination speed is slightly reduced, the accuracy of field maturity discrimination of tobacco leaves is improved.

[0022] Regarding the second core factor, the MobileNetV3-Large model adds a Squeeze-and-Excitation (SE) module to the bottleneck structure to improve model performance. However, the SE module only considers the dependencies between channel information in the feature maps inside the model, while ignoring the extremely important positional information in the visual space. This results in the model only being able to capture local feature information, and has problems such as scattered regions of interest and limited performance.

[0023] To improve the model's ability to represent the field maturity features of flue-cured tobacco leaves, this invention replaces the SE module with a parameterless attention module (Similarity Attention Module, SimAM). Specifically, the SE modules in the 4th to 6th bottleneck structures and the 11th to 15th bottleneck structures of the basic model are all replaced with the SimAM module.

[0024] Specifically, the SimAM module first compresses the feature map (number of channels C × height H × width W) along the spatial direction, and calculates the mean x of each channel's height H × width W. Next, it calculates the squared error X' of the mean error of each position on X relative to the spatial position within the same channel. Then, it calculates X' / n and t within each channel as channel information. Finally, it calculates the energy of each pixel. The feature map inference 3D attention weight module calculates its own energy using statistical laws, considering both the spatial position and channel of the feature map. It adaptively adjusts the weight of each position in the feature map, thereby focusing on effective features and suppressing ineffective features, improving the model's information perception and feature representation capabilities.

[0025] To verify the effectiveness of the SimAM module in extracting and representing model image features, this application introduces an Efficient Channel (EC) attention module, a Convolutional Block (CB) attention module, and a Coordinate (C) and SimAM module into the original squeezed excitation channel (SE) attention module of MobileNetV3-Large, respectively, to assess the accuracy and speed of discriminating the field maturity of flue-cured tobacco leaves on the validation set. Figure 4 As shown.

[0026] from Figure 4 It can be seen that the MobileNetV3-Large-EC model after introducing effective channel attention in the squeeze excitation channel attention, the MobileNetV3-Large-CB model after introducing convolutional block attention, the MobileNetV3-Large-C model after introducing coordinate attention, and the MobileNetV3-Large-SimAM model after introducing feature map inference of 3D attention weights all improve the accuracy of tobacco leaf field maturity discrimination. Except for the MobileNetV3-Large-CB model after introducing convolutional block attention, which has a slightly slower discrimination speed, the MobileNetV3-Large-C model after introducing effective channel attention... The EC model, the MobileNetV3-Large-C model with coordinate attention, and the MobileNetV3-Large-SimAM model with parameterless attention all improve the speed of tobacco field maturity discrimination. The MobileNetV3-Large-SimAM model has the highest accuracy and speed in tobacco field maturity discrimination, indicating that the MobileNetV3-Large-SimAM model with parameterless attention has the strongest three-dimensional information perception and feature expression ability of flue-cured tobacco field maturity images, and is the best improved MobileNetV3-Large discrimination model in terms of tobacco field maturity discrimination effect.

[0027] Furthermore, since each layer of a neural network extracts different features, the deeper the network, the more feature information is extracted. Therefore, the MobileNetV3-Large model introduces skip connections when the input and output feature map sizes and number of channels are the same. Specifically, skip connections are implemented in the bottleneck structures Bneck3, Bneck5, Bneck6, Bneck8, Bneck9, Bneck10, Bneck12, Bneck14, and Bneck15 of the basic model to ensure that effective features can still be extracted in deep networks, thereby improving the model's accuracy.

[0028] To further improve the utilization of prior features by the bottleneck structures, this application introduces multilayer perceptrons in the 5th bottleneck structure Bneck5, the 6th bottleneck structure Bneck6, the 8th bottleneck structure Bneck8, the 9th bottleneck structure Bneck9, the 10th bottleneck structure Bneck10, the 14th bottleneck structure, and the 15th bottleneck structure, and introduces skip connection branches between the 4th and 6th bottleneck structures, between the 7th and 10th bottleneck structures, and between the 13th and 15th bottleneck structures in the basic model.

[0029] The jump connection branch configuration makes the output of the 6th, 10th, and 15th bottleneck structures determined by their outputs along with the outputs of the previous 2 or 3 bottleneck structures, thus increasing the depth of these bottleneck structures.

[0030] Furthermore, the structure of the Multilayer Perceptron (MLP) is an embedded small network perceptual classifier built by adding a GLU activation function and a 1×1 convolutional fully connected layer after the original bottleneck structure of 5×5 convolution. By adding hidden layers and activation functions to the MLP, the linear model is transformed into a nonlinear model, introducing nonlinear fitting ability, thereby realizing the reuse of image features and enhancing perceptual fitting ability.

[0031] Furthermore, the GLU activation function is chosen, which helps the network selectively focus on certain parts of the input vector and suppress irrelevant or noisy information. This allows the MLP in this application to reuse only the feature vectors of interest. These improvements all enhance the model's feature utilization, making the model unaffected by production area merging and improving detection accuracy.

[0032] Furthermore, the classifier of the original MobileNetV3-Large model was designed for 1000 categories on the ImageNet dataset. It is necessary to first use PW convolution to increase the dimension of the feature map from 96 to 576 dimensions, then perform global average pooling to obtain a vector of length 576, and then pass two fully connected layers to obtain the output result.

[0033] The target categories in the flue-cured tobacco leaf dataset are only nine: upper mature, upper overripe, and upper underripe; middle mature, middle overripe, and middle underripe; lower mature, lower overripe, and lower underripe. This is significantly different from the original network. To better classify the extracted features, this application optimizes the classifier in the model structure based on the dimension of feature extraction. Redundant and time-consuming multiple dimensionality-up operations are removed, specifically the two conv2d convolutional layers before and after AVGpool are deleted. Only a single fully connected layer is used to directly map the pooled feature map to the number of target categories.

[0034] Furthermore, the MobileNetV3-Large model structure formed after the above three improvements is as follows: Figure 5 As shown.

[0035] The second aspect provides an intelligent terminal, including a camera, a discrimination module, and a result display module. The camera is used to capture images of flue-cured tobacco leaves whose field maturity needs to be determined. When the discrimination module is working, it executes the aforementioned intelligent discrimination method for field maturity of flue-cured tobacco leaves based on the improved MobileNetV3-Large. The result display module is used to display the field maturity of the flue-cured tobacco leaves to be determined.

[0036] A third aspect provides a computer device including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described above.

[0037] A fourth aspect provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0038] The fifth aspect provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method.

[0039] This invention has outstanding substantive features and significant progress compared to the prior art. Specifically,

[0040] 1. This invention introduces a dynamic convolution module into the original depthwise separable convolution of the basic model, replacing the SE module in the 4th to 6th bottleneck structures and the 11th to 15th bottleneck structures of the basic model with a parameterless attention module; introducing a multilayer perceptron in the 5th, 6th, 9th, 10th, 14th and 15th bottleneck structures, and introducing skip connection branches between the 4th and 6th bottleneck structures, between the 7th and 10th bottleneck structures, and between the 13th and 15th bottleneck structures of the basic model, thereby enhancing the ability to extract deep image features and perceive channel and location information, and improving the model's feature utilization rate by reusing the feature vectors of interest in a targeted manner, so that the detection results are not limited to the production area and location, thereby improving the detection accuracy;

[0041] 2. By using smart terminals to replace computers and digital cameras, field operations are convenient, and the results of judging the maturity of flue-cured tobacco leaves in the field are displayed on-site, demonstrating high intelligence;

[0042] 3. It is highly objective, avoiding the problems of high subjectivity and reliance on experience that arise from manually extracting features to determine the maturity of flue-cured tobacco leaves in the field;

[0043] 4. Convenient, practical, low-cost, and easy to popularize and apply. Attached Figure Description

[0044] Figure 1 The steps of the intelligent field maturity determination method for flue-cured tobacco leaves based on the improved MobileNetV3-Large described in Example 1;

[0045] Figure 2A schematic diagram of the structure of the dynamic convolutional DSDC module of the MobileNetV3-Large discriminant model for improving the field maturity of flue-cured tobacco leaves;

[0046] Figure 3 This is a schematic diagram illustrating the accuracy and speed of the MobileNetV3-Large-DSC and MobileNetV3-Large-DSDC models in determining the field maturity of flue-cured tobacco leaves on the validation set.

[0047] Figure 4 This is a schematic diagram illustrating the accuracy and speed of the MobileNetV3-Large-SE and MobileNetV3-Large-SimAM models in determining the field maturity of flue-cured tobacco leaves on the validation set.

[0048] Figure 5 A schematic diagram of the overall structure of the MobileNetV3-Large discriminant model for improving the field maturity of flue-cured tobacco leaves;

[0049] Figure 6 The steps of Example 1 are as follows: Intelligent determination method for field maturity of flue-cured tobacco leaves based on improved MobileNetV3-Large;

[0050] Figure 7 The diagram shows the effect of the present invention on the field maturity of flue-cured tobacco leaves in the implementation application examples of patents CN105004722A, CN114609134A, and CN114609135A. Detailed Implementation

[0051] The technical solution of the present invention will be further described in detail below through specific embodiments.

[0052] Example 1

[0053] This embodiment provides an intelligent method for determining the field maturity of flue-cured tobacco leaves based on an improved MobileNetV3-Large, which is applied to smart terminals, such as... Figure 6 As shown, the steps include:

[0054] (1) Obtain images of flue-cured tobacco leaves whose field maturity is to be determined. Preferably, the number of pixels in the flue-cured tobacco leaf images is greater than 224 pixels × 224 pixels.

[0055] (2) The improved MobileNetV3-Large discriminant model was used to perform field maturity discrimination processing on the flue-cured tobacco leaf images.

[0056] Among them, the improved MobileNetV3-Large discriminant model is based on the MobileNetV3-Large as the base model. A dynamic convolution module is introduced into the original depthwise separable convolution of the base model, and the dynamic convolution module dynamically adjusts the convolution kernel parameters of the model according to the input image; the SE modules in the 4th to 6th bottleneck structures and the 11th to 15th bottleneck structures of the base model are replaced with SimAM modules; multi-layer perceptrons are introduced into the 5th bottleneck structure, the 6th bottleneck structure, the 8th bottleneck structure, the 9th bottleneck structure, the 10th bottleneck structure, the 14th bottleneck structure and the 15th bottleneck structure, and skip connection branches are introduced between the 4th bottleneck structure and the 6th bottleneck structure, between the 7th bottleneck structure and the 10th bottleneck structure, and between the 13th bottleneck structure and the 15th bottleneck structure of the base model; meanwhile, the two conv2d convolutional layers before and after the AVGpool are deleted.

[0057] In one embodiment, the dynamic convolution module uses depthwise convolution to perform convolution operations on each channel of the input feature map separately, and different convolution kernels are used on each channel to extract different features; then the feature map obtained by depthwise convolution is passed through the attention module in the dynamic convolution to dynamically and adaptively generate the weight coefficient πk (0 < k ≤ K, K is the number of convolution kernels) of each convolution kernel, and after multiplying each weight coefficient πk by the corresponding convolution and summing to obtain the final weight coefficient, it is convolved with the feature map obtained by depthwise convolution, and dynamic adaptive convolution is completed through batch normalization and the ReLU activation function.

[0058] In one embodiment, the multi-layer perceptron is an embedded small network perception classifier constructed by adding a GLU activation function and a 1×1 convolutional fully connected layer after the 5×5 convolution in the original bottleneck structure.

[0059] Furthermore, in the improved MobileNetV3-Large discriminant model, the two conv2d convolutional layers before and after the AVGpool are deleted.

[0060] The discriminant effects of the implementation application examples of flue-cured tobacco leaf field maturity using the present invention, the patent of CN105004722A, the patent of CN114609134A and the patent of CN114609135A are shown in Figure 7. Among them, the tobacco leaf samples are from the tobacco fields in Pan Gu, Guo Ji and Wang Dian towns of Biyang County, Zhumadian. After the tobacco plants enter the mature period, 657 flue-cured tobacco leaves to be discriminated for field maturity are respectively selected, and the images of the flue-cured tobacco leaves to be discriminated for field maturity are taken by intelligent terminals such as mobile phones, and 2491 flue-cured tobacco leaf images with pixels greater than 224pixel × 224pixel are obtained.

[0061] It can be seen that the accuracy rate of this invention in determining the field maturity of flue-cured tobacco leaves in the application example reaches 97.16%, and the speed reaches 12.51 leaves / min, which is 23.36% and 5.82 leaves / min higher than that of patent CN105004722A, respectively; 13.35% and 7.43 leaves / min higher than that of patent CN114609134A, respectively; and 6.29% and 8.09 leaves / min higher than that of patent CN114609134A, respectively (see [link to patent CN105004722A]). Figure 7 It avoids the disadvantages of patents CN103245625A, CN109540894A, and CN108198176A, which require chlorophyll meters, color difference meters, computers, or digital cameras, resulting in higher costs and inconvenience in field operation. It also avoids the problems of subjective factors and high experience-based nature in manually judging the maturity of flue-cured tobacco leaves in the field. It has the advantages of higher judgment accuracy, objectivity and non-destructive nature, convenient field operation, and low cost.

[0062] Implementation Example 1

[0063] On July 9, 2023, in a tobacco field in Guoji Town, Biyang County, 216 lower tobacco leaves with field maturity to be determined were selected. First, images of the tobacco leaves with field maturity to be determined were obtained by taking pictures using a mobile phone or other smart terminal. Second, the field maturity of the tobacco leaves was determined using a MobileNetV3-Large model with dynamic convolution and coordinate attention modules on the mobile phone or other smart terminal. Finally, the field maturity of the tobacco leaves was displayed using the result display function on the mobile phone or other smart terminal. The accuracy rate of this invention in determining the field maturity of the lower tobacco leaves was 97.13%, and the speed was 12.46 leaves / minute.

[0064] Implementation Example 2

[0065] On August 6, 2023, in a tobacco field in Wangdian Town, Biyang County, 218 tobacco leaves from the middle section of the flue-cured tobacco plant, whose field maturity was to be determined, were selected. First, images of the tobacco leaves were taken using a smartphone or other smart device. Then, the field maturity of the tobacco leaves was determined using a MobileNetV3-Large model equipped with a dynamic convolution module and a coordinate attention module on the smartphone or other smart device. Finally, the field maturity of the tobacco leaves was displayed using the result display function on the smartphone or other smart device. The accuracy rate of this invention in determining the field maturity of the middle section of the flue-cured tobacco leaves was 97.15%, and the speed was 12.49 leaves / minute.

[0066] Implementation Example 3

[0067] On September 2, 2023, in a tobacco field in Pangu Town, Biyang County, 223 upper tobacco leaves of flue-cured tobacco with field maturity to be determined were selected. First, images of the tobacco leaves with field maturity to be determined were obtained by taking pictures using a mobile phone or other smart terminal. Second, the field maturity of the flue-cured tobacco leaves was determined using a MobileNetV3-Large model with a dynamic convolution module and a coordinate attention module, which is equipped on the mobile phone or other smart terminal. Finally, the field maturity of the flue-cured tobacco leaves to be determined was displayed using the result display function of the mobile phone or other smart terminal. The accuracy rate of the present invention in determining the field maturity of the upper tobacco leaves of flue-cured tobacco was 97.21%, and the speed was 12.57 leaves / minute.

[0068] Example 2

[0069] This embodiment provides an intelligent terminal, including a camera, a discrimination module, and a result display module. The camera is used to capture images of flue-cured tobacco leaves whose field maturity needs to be determined. When the discrimination module is working, it executes the intelligent discrimination method for field maturity of flue-cured tobacco leaves based on the improved MobileNetV3-Large described in Embodiment 1. The result display module is used to display the field maturity of the flue-cured tobacco leaves to be determined.

[0070] Example 3

[0071] This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in Embodiment 1.

[0072] Example 4

[0073] This embodiment provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in Embodiment 1.

[0074] Example 5

[0075] This embodiment provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in Embodiment 1.

[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A method for intelligently determining the field maturity of flue-cured tobacco leaves based on an improved MobileNetV3-Large, characterized in that, Includes the following steps: Acquire images of flue-cured tobacco leaves whose field maturity needs to be determined; An improved MobileNetV3-Large discriminant model was used to process images of flue-cured tobacco leaves to determine their field maturity. The improved MobileNetV3-Large discriminative model is based on MobileNetV3-Large. It introduces a dynamic convolution module into the original depthwise separable convolution of the base model. This dynamic convolution module dynamically adjusts the convolution kernel parameters based on the input image. The SE module in the fourth to sixth bottleneck structures and the eleventh to fifteenth bottleneck structures of the base model is replaced with a parameterless attention module. Multilayer perceptrons are introduced in the fifth, sixth, eighth, ninth, tenth, fourteenth, and fifteenth bottleneck structures. Skip connections are introduced between the fourth and sixth bottleneck structures, between the seventh and tenth bottleneck structures, and between the thirteenth and fifteenth bottleneck structures of the base model. In the improved MobileNetV3-Large discriminant model, the two conv2d convolutional layers before and after AVGpool are removed.

2. The intelligent method for determining the field maturity of flue-cured tobacco leaves based on the improved MobileNetV3-Large according to claim 1, characterized in that: The multilayer perceptron is an embedded small-scale network perceptron classifier constructed by adding a GLU activation function and a 1×1 convolutional fully connected layer after the 5×5 convolution in the original bottleneck structure.

3. The intelligent method for determining the field maturity of flue-cured tobacco leaves based on the improved MobileNetV3-Large according to claim 1, characterized in that: The captured images of flue-cured tobacco leaves have a resolution greater than 224 pixels × 224 pixels.

4. A smart terminal, characterized in that: It includes a camera, a discrimination module, and a result display module. The camera is used to capture images of flue-cured tobacco leaves whose field maturity is to be determined. When the discrimination module is working, it executes the intelligent discrimination method for field maturity of flue-cured tobacco leaves based on the improved MobileNetV3-Large as described in any one of claims 1-3. The result display module is used to show the field maturity of the flue-cured tobacco leaves to be identified.

5. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 3.

7. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 3.

Images