Corn leaf stomatal measurement method based on anchor-free frame target detection and stomatal conductance calculation

Abstract

This invention discloses a method for measuring stomatal conductance in maize leaves based on anchor-free target detection and stomatal conductance calculation, comprising the following steps: acquiring stomatal images of maize leaves; preprocessing, data augmentation, and stomatal labeling of the maize leaf stomatal images; constructing an optimal stomatal conductance model; constructing a deep learning model for stomatal identification and measurement, and training and testing the deep learning model. This invention improves upon the CenterNet target detection model, using DLA34 as the backbone network, replacing upsampled 2D convolution with deformable convolution, introducing an attention mechanism, and adding an angle prediction branch to automatically identify and measure stomata with tilt angles in maize leaf microscopic images. Simultaneously, a stomatal conductance loss function is introduced, and the optimal stomatal conductance model is used to automatically measure the stomatal conductance of maize leaves.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and in particular to a corn leaf stomata measurement method based on anchor-free target detection and stomatal conductance calculation. BACKGROUND

[0002] At present, stomata is an important structure for gas and water exchange between terrestrial plant leaves and air. Plants gradually evolved this special water and gas exchange structure in the process of evolving from aquatic environment to terrestrial environment. Figure 1 The micrograph shows the stomata of a plant leaf. Stomata play an important role in the process of photosynthesis in plants, and can regulate the water content and carbon dioxide level in the leaves. Stomatal conductance is an important indicator of the balance and circulation of water and carbon dioxide between plants and the atmosphere. Rapid, efficient and automatic detection of stomata can help researchers conduct large-scale research on stomatal morphology, structure and stomatal conductance model.

[0003] Traditional stomata measurement methods usually use optical microscopes for manual observation and measurement. These methods usually require researchers to manually mark the boundaries and length and width of stomata, which is not efficient, and human errors are likely to occur due to manual operation. In order to automatically detect stomata, researchers have proposed many automatic measurement methods.

[0004] With the development of deep learning technology, it is possible to efficiently and accurately identify and measure stomata. Some machine learning and deep learning-based stomata research has been proposed. Most deep learning-based methods directly use existing models without considering the characteristics of plant stomata images or the problems of accelerating convergence speed and model generalization ability during model training. In addition, many deep learning-based methods can automatically identify and count stomata, but cannot simultaneously achieve automatic measurement of stomatal parameters, and the measurement work still needs to be completed manually in the later stage, and deep learning methods and stomatal conductance are not combined. Moreover, the current method is based on horizontal anchor frame, and when collecting stomata images on plant leaves, it is often not possible to ensure that all stomata are in a horizontal state in the microscope, so most stomata are inclined, and the automatic detection methods proposed cannot identify the rotated stomata well, and an additional step of rotating and correcting the stomata image is usually required, which affects the efficiency of stomata detection. SUMMARY

[0005] The purpose of the present application is to provide a corn leaf stomata measurement method based on anchor-free target detection and stomatal conductance calculation, which can efficiently and quickly identify and detect stomata.

[0006] The technical solution adopted by the present application is:

[0007] Corn leaf stomata measurement method based on anchor-free frame target detection and stomata conductance calculation

[0008] The method comprises the following steps:

[0009] Step 1: Obtain a corn leaf stomata image;

[0010] Step 2: Preprocess, data enhance and mark the corn leaf stomata image;

[0011] Step 3: Construct an optimal stomata conductance model; the optimal stomata conductance model is as follows:

[0012]

[0013] In the formula, d is the diffusion rate of water in air m 2 ·s -1 ; 24.9*10 -6 m 2 ·s -1 at 25 DEG C; SD is the stomata density unit, which is mm -2 ; alpha max is the estimated maximum stomata aperture area, which is pi*SL 2 / 4, wherein SL is the stomata length; l is the stomata cavity depth, which is assumed to be equal to the width of the guard cell; v is the air molar volume (m 3 ·mol -1 ), which is 22.4*10 -3 m 3 ·mol -1 at 25 DEG C and 101.3 kPa;

[0014] Step 4: Construct a deep learning model for stomata identification and measurement; the specific steps are as follows:

[0015] Step 4.1: Construct an iterative deep aggregation structure;

[0016] Step 4.2: Construct a hierarchical deep aggregation structure;

[0017] Step 4.3: Construct a direct learning structure: combine the above iterative deep aggregation structure and hierarchical deep aggregation structure to form a direct learning structure;

[0018] Step 4.4: Construct an attention mechanism module to obtain the features of the target; the attention mechanism module is combined by a channel attention module and a spatial attention module, and the channel attention and spatial attention are evaluated and scored at the same time;

[0019] Step 4.5: Replace 2D convolution with deformable convolution in the upsampling process: replace 2D convolution with deformable convolution in the upsampling process to obtain a larger receptive field after using a small step downsampling rate to extract target features between different levels;

[0020] Step 4.6: Add angle prediction branch; an angle detection head is added to detect angle inclined pores;

[0021] Step 4.7: Introduce angle loss and pore conductance loss in the loss function;

[0022] Step 4.8: Pore segmentation and measurement; the length and width of the rotating bounding box are used as the length and width of the pore, and the pore position is directly located by the rotating bounding box;

[0023] Step 4.9: Calculate the corn leaf stomatal conductance: according to the length and width parameters of each pore in the leaf, the corn leaf stomatal conductance is automatically calculated according to formula (1);

[0024] Step 5: Training and testing of deep learning model; the specific steps are as follows:

[0025] Step 5.1: Divide the pore image into training set, validation set and test set;

[0026] Step 5.2: Train the model on the training set;

[0027] Step 5.3: Evaluate the model performance using accuracy, precision, recall and F1 score indicators on the test set.

[0028] Said step 1 adopts nail polish imprint method to obtain corn leaf pore image; the specific steps are as follows:

[0029] Step 1.1: Apply a layer of transparent nail polish on the corn seedling leaf;

[0030] Step 1.2: Collect nail polish imprint sheet;

[0031] Step 1.3: Make temporary nail polish imprint sheet slide;

[0032] Step 1.4: Observe the temporary slide under a general optical microscope;

[0033] Step 1.5: Capture clear field of view as pore image using ImageView software;

[0034] Step 1.6: Number the pore image and add it to the pore image data set.

[0035] Said step 2 specifically includes the following steps:

[0036] Step 2.1: convert the format and resolution of the maize leaf stomata image;

[0037] Step 2.2: pre-process the maize leaf stomata image using a geometric distortion correction algorithm and a gray scale stretching algorithm;

[0038] Step 2.2: data augmentation of the pre-processed maize leaf stomata image using the CLoDSA method;

[0039] Step 2.3: label the maize leaf stomata image obtained in step 2.2 using labelimg2 software;

[0040] Step 2.4: convert the label file format to a json file that can be recognized by the CenterNet deep learning model.

[0041] The iterative deep aggregation structure IDA process calculation formula is as follows:

[0042]

[0043] Where I represents the entire IDA module, N represents the aggregation node; such as N(X1, X2) represents an aggregation node with inputs X1 and X2.

[0044] The deep aggregation structure HAD process in step 4.2 is calculated as follows:

[0045]

[0046] Where: N is the aggregation node, and R and L are defined as follows:

[0047]

[0048] Where: B represents a convolution block.

[0049] The input of the aggregation node in the IDA structure in step 4.2 is 2, and the input of the aggregation node in the HDA structure is 2 or more. The aggregation node fuses multiple input features to form a single feature and outputs it. In order to reduce the amount of calculation, the aggregation node is constructed using the structure of "convolution-BN-activation function", and a residual connection is added therein, and the expression is as follows:

[0050]

[0051] Where: W i represents the weight, and b represents the bias.

[0052] The step 4.7 specifically comprises the following steps:

[0053] Focal loss is used to calculate the difference between the predicted heatmap information and the actual heatmap information; the heatmap loss function is as follows:

[0054]

[0055] Where p and q represent the predicted heatmap value and the actual heatmap value, respectively; i is the index of the pixel position in the feature map, and n is the number of pixels; α and β are two hyperparameters used to control the contribution of each point.

[0056] To minimize the difference between the predicted and actual values, the remaining loss functions are all optimized using the RegL1 loss function; the center point offset loss is as follows:

[0057]

[0058] in and o k These represent the offset information of the predicted and actual k-th instances, respectively;

[0059] The length and width losses are as follows:

[0060]

[0061] in and s k These represent the width and height information of the predicted and actual k-th instances, respectively;

[0062] The angle loss is as follows:

[0063]

[0064] in and θ k Let represent the predicted and actual angle information of the k-th instance, respectively; the porosity loss is as follows:

[0065]

[0066] in and G k Let represent the predicted and actual stomatal conductance information for the k-th instance, respectively; the stomatal conductance is calculated based on the optimal stomatal conductance model; the total loss function is the sum of the above loss functions, expressed as follows:

[0067]

[0068] This invention improves upon the CenterNet object detection model by using DLA34 as the backbone network, replacing upsampled 2D convolution with deformable convolution, introducing an attention mechanism, and adding an angle prediction branch to automatically identify and measure stomata with tilt angles in microscopic images of maize leaves. It also introduces a stomatal conductance loss function and uses the optimal stomatal conductance model to automatically measure the stomatal conductance of maize leaves. Attached Figure Description

[0069] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0070] Figure 1 This is a flowchart of the present invention;

[0071] Figure 2 This is a schematic diagram of the pore identification and measurement process of the method of the present invention;

[0072] Figure 3 Microscopic images of plant leaf stomata according to the present invention;

[0073] Figure 4 This is a schematic diagram of an iterative deep aggregation structure;

[0074] Figure 5 (Schematic diagram of a layered deep aggregation structure);

[0075] Figure 6 A schematic diagram for constructing a direct learning structure;

[0076] Figure 7 This is a schematic diagram of the attention mechanism module.

[0077] Figure 8 The experimental results are for pore identification;

[0078] Figure 9 A regression comparison chart of pore conductance values;

[0079] Figure 10 For parameter changes during the training process of deep learning models;

[0080] Figure 11 This is an image of the stomata on a maize leaf obtained by the method of the present invention. Detailed Implementation

[0081] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0082] like Figure 1 As shown, the present invention includes the following steps:

[0083] Step 1: Obtain images of stomata on corn leaves using the nail polish impression method; the specific steps are as follows:

[0084] Step 1.1: Apply a layer of clear nail polish to the corn seedling leaves; specifically:

[0085] The test samples were collected at 25℃. The second leaf of the corn seedling was taken as the sampling leaf, and a layer of transparent nail polish was evenly applied to the middle of the leaf.

[0086] Step 1.2: Collect nail polish imprints; specifically:

[0087] After drying for 20 minutes, remove the nail polish layer with tweezers and store the collected nail polish imprint in a 1.5ml test tube.

[0088] Step 1.3: Prepare a temporary slide for the nail polish imprint; specifically:

[0089] Carefully place the nail polish mark onto a glass slide using tweezers, and add a small amount of water to increase the viscosity of the mark and the slide. Add a coverslip to make a temporary slide.

[0090] Step 1.4: Observe the temporary slide using a regular optical microscope; specifically:

[0091] The temporary slides were observed using a regular optical microscope at magnifications of 10×10 and 10×20, respectively.

[0092] Step 1.5: Use ImageView software to capture a clear view as a stomatal image; such as Figure 3 As shown,

[0093] Step 1.6: Number the stomatal images and add them to the stomatal image dataset;

[0094] Step 2: Stomatal image preprocessing, data augmentation, and stomatal labeling; the specific steps are as follows:

[0095] Step 2.1: Convert image format and resolution; specifically:

[0096] Convert the stomatal images from TIFF format to JPG format, and uniformly modify the image resolution to 1000×667 pixels (approximately 6μm / pixel).

[0097] Step 2.2: Preprocess the stomatal images using geometric distortion correction and grayscale stretching algorithms;

[0098] Step 2.2: Perform data augmentation using the CloDSA method; specifically:

[0099] Data augmentation was performed using the CloDSA method. Through a series of operations such as random angle rotation, random occlusion, and the addition of Gaussian noise, the stomata image dataset was expanded to meet the needs of deep learning model training.

[0100] Step 2.3: Use labelimg2 software to label the stomatal images; specifically:

[0101] Use labelimg2 software to label the stomata images, such as Figure 11 As shown. During the labeling process, the smallest possible bounding rectangle is used to fit the pores as closely as possible. After labeling all the pores in the image, the labelimg2 software normalizes the labeling information and automatically generates a label file containing pore coordinate information.

[0102] Step 2.4: Convert the label file format to a JSON file that the CenterNet deep learning model can recognize;

[0103] Step 3: Construct the optimal porosity conductance model; specifically:

[0104] The optimal porosity model is as follows:

[0105]

[0106] In the formula, d is the diffusivity of water in air, m 2 ·s -1 At 25℃, it is 24.9×10. -6 m 2 ·s -1 SD is the pore density (pores per mm²). -2 );α max To estimate the maximum pore area (valued as π×SL) 2 / 4, where SL is the stomatal length; l is the stomatal cavity depth, assumed to be equal to the width of the guard cell; v is the air molar volume (m³). 3 ·mol -1 At 25℃ and 101.3 kPa, the value is 22.4 × 10⁻⁶. -3 m3 ·mol -1 .

[0107] Step 4: Construct a deep learning model for stomatal identification and measurement;

[0108] This invention improves upon the CenterNet object detection model. CenterNet is an end-to-end model that is faster and more accurate than anchor-box-based models. The CenterNet object detection model first scales the input image to a preset size to generate a keypoint heatmap. The keypoint locations in the original image are downsampled and then distributed onto the heatmap using a Gaussian distribution.

[0109] A schematic diagram of the pore identification and measurement process of the method of the present invention is shown below. Figure 2 As shown, the specific steps are as follows:

[0110] Step 4.1: Construct the Iterative Deep Aggregation (IDA) structure; specifically:

[0111] A schematic diagram of Iterative Deep Aggregation (IDA) is shown below. Figure 4 As shown.

[0112] Aggregation nodes aggregate features as they propagate from shallow to deep. This involves breaking down the convolutional neural network (CNN) structure; a CNN consists of multiple stages with consistent resolution, each stage is composed of multiple blocks, and each block contains multiple layers.

[0113] The IDA process calculation formula is as follows:

[0114]

[0115] Where I represents the entire IDA module, and N represents the aggregation node. For example, N(X1,X2) represents an aggregation node with inputs X1 and X2.

[0116] Step 4.2: Construct a Hierarchical Deep Aggregation (HAD) structure; specifically:

[0117] A schematic diagram of Hierarchical Deep Aggregation (HAD) is shown below. Figure 5 As shown.

[0118] IDA can effectively fuse features from multiple stages; however, features from multiple blocks within a stage are not fused. HDA (High-Definition Aspect-Based ...s) can be combined, allowing information to be combined across different layers, resulting in richer features learned by the model.

[0119] The HAD process is calculated as follows:

[0120]

[0121] Where: N is the aggregation node, and R and L are defined as follows:

[0122]

[0123] Where: B represents a convolutional block.

[0124] In the IDA structure, the aggregation node receives two inputs, while in the HDA structure, it receives two or more. The aggregation node fuses multiple input features to form a single feature and outputs it. To reduce computation, this invention employs a "convolution-BN-activation function" structure to construct the aggregation node, incorporating residual connections, as shown in the following expression:

[0125] N(x1,x2,…,x n ,)

[0126] =σ(BatchNorm(∑ i W i x i +b)+x n (5)

[0127] Among them: W i 'b' represents the weight, and 'b' represents the bias.

[0128] Step 4.3: Construct the Direct Learning Architecture (DLA); specifically:

[0129] Combining the aforementioned IDA and HDA forms a Direct Learning Architecture (DLA), such as... Figure 6 As shown.

[0130] Step 4.4: Construct the Convolutional Block Attention Module (CBAM); specifically:

[0131] A Convolutional Block Attention Module (CBAM) is introduced to selectively acquire target features. CBAM combines the Channel Attention Module and the Spatial Attention Module, simultaneously evaluating and scoring both channel and spatial attention. A schematic diagram of the CBAM attention mechanism module is shown below. Figure 7 As shown.

[0132] Step 4.5: During the upsampling process, deformable convolutional networks (DCNs) are used instead of 2D convolutions; specifically:

[0133] First, a small step downsampling rate is used to extract target features between different levels. Then, deformable convolution is used instead of 2D convolution during the upsampling process to obtain a larger receptive field.

[0134] Step 4.6: Add an angle prediction branch; specifically:

[0135] Compared to the three detectors in the original CenterNet model, this invention adds an angle detector (angle prediction branch) to detect angled pores.

[0136] Step 4.7: Introduce angle loss and porosity conductance loss into the loss function; specifically:

[0137] This invention introduces angle loss and porosity conductance loss, enabling the model to better measure pores and calculate porosity conductance. Focal loss is used to calculate the difference between predicted and actual heatmap information. The heatmap loss function is as follows:

[0138]

[0139] Where p and q represent the predicted and actual heatmap values ​​of the image, respectively. i is the index of the pixel position in the feature map, and n is the number of pixels. α and β are two hyperparameters used to control the contribution of each point.

[0140] To minimize the difference between the predicted and actual values, the remaining loss functions are all optimized using the RegL1 loss function. The center point offset loss is as follows:

[0141]

[0142] in and ok These represent the offset information for the predicted and actual k-th instances, respectively.

[0143] The length and width losses are as follows:

[0144]

[0145] in and s k These represent the width and height information of the predicted and actual k-th instances, respectively.

[0146] The angle loss is as follows:

[0147]

[0148] in and θ k These represent the angle information of the predicted and actual k-th instances, respectively.

[0149] The porosity loss is as follows:

[0150]

[0151] in and G k These represent the predicted and actual stomatal conductance information for the k-th instance, respectively. The stomatal conductance is calculated based on the optimal stomatal conductance model.

[0152] The total loss function is the sum of the loss functions mentioned above, and its expression is as follows:

[0153] L = L H +L off +L size +L ang +L con (11)

[0154] Step 4.8: Stomatal segmentation and measurement; specifically:

[0155] The method of this invention can directly locate the air vent position using a rotating bounding frame, which provides a better fit to the air vent compared to traditional target detection methods. In this invention, the length and width of the rotating bounding frame are used as the length and width of the air vent. Figure 8 The experimental results for pore recognition are shown.

[0156] Step 4.9: Calculate the stomatal conductance of maize leaves; specifically:

[0157] Based on the length and width parameters of each stomata in the leaf, the stomatal conductance of the maize leaf is automatically calculated according to formula (1). A regression comparison chart of stomatal conductance values ​​is shown below. Figure 9As shown. The left figure shows the stomatal conductance fitting result of the original CenterNet model, and the right figure shows the stomatal conductance fitting result after adding conductance loss. The smaller the MSE value, the better the R... 2 The larger the value, the better the fit.

[0158] Step 5: Training and testing the deep learning model; the specific steps are as follows:

[0159] Step 5.1: Divide the stomatal images into training, validation, and test sets; specifically:

[0160] The training set is used to teach the model stomatal features and optimize the parameters in the network. The validation set is used to evaluate the model's current generalization ability. The test set is used to test the model's final generalization performance and is not used as a basis for parameter tuning or feature selection. The proportions of the three sets are 70%, 15%, and 15%, respectively.

[0161] Step 5.2: Train the model on the training set; specifically:

[0162] During training, the epoch was set to 20. The learning rate used in the iterations was an exponentially decaying learning rate. The learning rate decayed according to formula (12), and the initial value of gamma was set to 0.95.

[0163] lr=lr×gamma epoch (12)

[0164] Figure 10 The figure shown is a graph illustrating the parameter changes during the training process of the model of this invention;

[0165] Step 5.3: Evaluate the model performance on the test set using the metrics of accuracy, precision, recall, and F1 score.

[0166] In the comparative experiments, the improved CenterNet model of this invention was compared with CFA, ConvNeXt, G-Rep, KFIoU, KLD, Oriented R-CNN, R3Det, ​​ReDet, Rotated FCOS, RetinaNet, and the final results are shown in Table 1. As can be seen from Table 1, the improved CenterNet model generally outperforms other models in terms of precision, recall, and F1-score over the same 20 epochs.

[0167]

[0168] Table 1 Comparison of experimental results

[0169] In summary, this invention provides a method for measuring stomatal conductance in maize leaves based on anchor-free target detection and stomatal conductance calculation. The method improves upon the CenterNet target detection model, using DLA34 as the backbone network, replacing upsampled 2D convolution with deformable convolution, introducing an attention mechanism, and adding an angle prediction branch to automatically identify and measure stomata with tilt angles in maize leaf microscopic images. Simultaneously, a stomatal conductance loss function is introduced, and an optimal stomatal conductance model is used to automatically measure the stomatal conductance of maize leaves.

[0170] In the description of this invention, it should be noted that for directional terms, such as "center," "lateral," and "vertical," the appropriate terms may be used.

[0171] The directions and positional relationships indicated by symbols such as "direction", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", and "counterclockwise" are based on the directions or positional relationships shown in the accompanying drawings and are only for the purpose of facilitating the description of the present invention and simplifying the description. They are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and should not be construed as limiting the specific protection scope of the present invention.

[0172] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification and claims of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or device.

[0173] Note that the above description is merely a preferred embodiment and application of the technical principles of the present invention. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the specific embodiments described herein, and may include many other effective embodiments without departing from the concept of the present invention. The scope of the present invention is determined by the scope of the appended claims.

Claims

1. A method for measuring stomatal activity in maize leaves based on frameless target detection and stomatal conductance calculation, characterized in that: Includes the following steps: Step 1: Obtain images of stomata on maize leaves; Step 2: Preprocess, data augment, and stomatal labeling of maize leaf stomatal images; Step 3: Construct the optimal porosity conductance model; the optimal porosity conductance model is as follows: ； In the formula, The diffusivity of water in air At 25℃ ; The unit of porosity is a unit of · ; To estimate the maximum pore area, the value is taken as... ,in For stomata length; Let be the depth of the stomatal cavity, assumed to be equal to the width of the guard cell; It is the molar volume of air ( ), 25℃, 101.3 Below is ; Step 4: Construct a deep learning model for stomatal identification and measurement; the specific steps are as follows: Step 4.1: Construct an iterative deep aggregation structure; Step 4.2: Construct a hierarchical deep aggregation structure; Step 4.3: Construct the direct learning structure: Combine the iterative deep aggregation structure and the hierarchical deep aggregation structure described above to form the direct learning structure; Step 4.4: Construct an attention mechanism module to selectively acquire the stomatal features of maize leaves; The attention mechanism module combines the channel attention module and the spatial attention module, and simultaneously evaluates and scores both channel attention and spatial attention; Step 4.5: Use deformable convolution instead of 2D convolution during upsampling: First, use a small step downsampling rate to extract target features between different levels, and then use deformable convolution instead of 2D convolution during upsampling to obtain a larger receptive field. Step 4.6: Add an angle prediction branch; an angle detection head was added to detect angled tilted vents; Step 4.7: Introduce angle loss and porosity conductance loss into the loss function; Step 4.8: Pore segmentation and measurement; using the length and width of the rotating bounding frame as the length and width of the pores, the pore positions are located directly through the rotating bounding frame; Step 4.9: Calculate the stomatal conductance of maize leaves: Based on the length and width parameters of each stoma in the leaf, the stomatal conductance of maize leaves is automatically calculated according to formula (1); Step 5: Training and testing the deep learning model; the specific steps are as follows: Step 5.1: Divide the stomatal images into training, validation, and test sets; Step 5.2: Train the model on the training set; Step 5.3: Evaluate model performance on the test set using accuracy, precision, recall, and F1 score metrics.

2. The method for measuring stomatal activity in maize leaves based on anchorless target detection and stomatal conductance calculation according to claim 1, characterized in that: In step 1, the nail polish imprint method is used to obtain images of stomata on corn leaves; the specific steps are as follows: Step 1.1: Apply a layer of clear nail polish to the corn seedling leaves; Step 1.2: Collect nail polish residue; Step 1.3: Prepare a temporary slide for the nail polish imprint; Step 1.4: Observe the temporary slide using a regular optical microscope; Step 1.5: Use ImageView software to capture a clear view as an image of the stomata; Step 1.6: Number the stomatal images and add them to the stomatal image dataset.

3. The method for measuring the stomata of maize leaves based on anchorless target detection and stomatal conductance calculation according to claim 2, characterized in that: Step 2 specifically includes the following steps: Step 2.1: Convert the format and resolution of the maize leaf stomatal image; Step 2.2: Preprocess the stomatal images of maize leaves using geometric distortion correction and grayscale stretching algorithms; Step 2.2: Data augmentation of the preprocessed maize leaf stomatal images was performed using the CloDSA method; Step 2.3: Use labelimg2 software to label the maize leaf stomatal images obtained in Step 2.2; Step 2.4: Convert the label file format to a JSON file that the CenterNet deep learning model can recognize.

4. The method for measuring the stomata of maize leaves based on anchorless target detection and stomatal conductance calculation according to claim 2, characterized in that: The calculation formula for the iterative deep aggregation structure IDA process is as follows: ； Where I represents the entire IDA module. Represents an aggregate node; Indicates an input as and The aggregation node.

5. The method for measuring stomatal activity in maize leaves based on anchorless target detection and stomatal conductance calculation according to claim 2, characterized in that: The calculation of the layered deep aggregation structure HDA process in step 4.2 is as follows: ； in: For aggregation nodes, and They are defined as follows: ； in: It represents a convolutional block.

6. The method for measuring stomatal pores in maize leaves based on anchorless target detection and stomatal conductance calculation according to claim 5, characterized in that: In step 4.2, the input to the aggregation node in the IDA structure is two, while in the hierarchical deep aggregation structure HDA, the input to the aggregation node is two or more. The aggregation node fuses multiple input features to form a single feature and outputs it. To reduce computation, a "convolution-BN-activation function" structure is used to construct the aggregation node, and residual connections are added to it. The expression is as follows: ； in: Indicates the weight. Indicates bias.

7. The method for measuring stomatal pores in maize leaves based on frameless target detection and stomatal conductance calculation according to claim 2, characterized in that: Step 4.7 specifically includes the following steps: Focal loss is used to calculate the difference between the predicted heatmap information and the actual heatmap information; the heatmap loss function is as follows: ； in , These represent the predicted heatmap value and the actual heatmap value of the image, respectively. It is the index of the pixel position in the feature map. It is the number of pixels; and These are two hyperparameters used to control the contribution of each point; To minimize the difference between the predicted and actual values, the remaining loss functions, i.e., the loss calculations for other prediction branches, are all optimized using the RegL1 loss function; the center point offset loss is as follows: ； in and Representing the predicted and actual values ​​respectively. Offset information for each instance; The length and width losses are as follows: ； in and Representing the predicted and actual values ​​respectively. Width and height information for each instance; The angle loss is as follows: ； in and Representing the predicted and actual values ​​respectively. Angle information for each instance; The porosity loss is as follows: ； in and Representing the predicted and actual values ​​respectively. The stomatal conductance information for each instance; the stomatal conductance is calculated based on the optimal stomatal conductance model; the total loss function is the sum of the above loss functions, expressed as follows: 。

Images