A method for monitoring plant growth information

Abstract

A method for monitoring plant growth information, which relates to the technical field of plant monitoring. Aiming at the problem in the prior art that accurate and clear growth information of plants at each stage cannot be obtained, this application solves the problem of low accuracy in distinguishing the growth information of wetland plants at each stage by the traditional naked eye. This application judges the growth information of plants through the leaf area index, combines a neural network, improves the speed of plant monitoring, realizes continuous monitoring of the growth situation of plants, greatly improves the monitoring efficiency, and avoids the problems of low monitoring efficiency, time-consuming and laborious caused by manual visual inspection of the growth status of plants.

Description

Technical Field

[0001] The present invention relates to the technical field of plant monitoring, and specifically relates to a method for monitoring plant growth information. Background Art

[0002] As is well known, the growth of plants is affected by various factors. Especially for crops, a suitable growth environment plays a positive role in plants. Obtaining the growth information of plants, diagnosing their growth status, analyzing their nutrient information, and studying the physiological and ecological laws of plants are of great significance for plant physiological research and guiding agricultural production and planting.

[0003] For a long time, whether it is farmers' planting or experimental fields, etc., monitoring the growth information of plants is basically carried out through on-site observation. Such a discrimination method cannot obtain accurate and clear growth information of plants at each stage. Summary of the Invention

[0004] The purpose of the present invention is: to propose a method for monitoring plant growth information in view of the problem that accurate and clear growth information of plants at each stage cannot be obtained in the prior art.

[0005] The technical solution adopted by the present invention to solve the above technical problems is:

[0006] A method for monitoring plant growth information includes a training step and a monitoring step:

[0007] The specific training step is as follows:

[0008] Training Model 1:

[0009] For the monitored plant, obtain the image of the monitored plant, mark the stem of the monitored plant and the image of the positive leaves of the monitored plant in the image, and finally use the image of the monitored plant as the input and the marked image as the output to train the neural network to obtain the trained Model 1;

[0010] Training Model 2:

[0011] For the monitored plant, obtain the length of the stem of the monitored plant and the corresponding number of leaves according to historical data, and use the length of the stem of the monitored plant as the input and the corresponding number of leaves as the output to train the neural network to obtain the trained Model 2;

[0012] Training Model 3:

[0013] Obtain the leaf area index of the monitored plant and the growth rate of the plant under this leaf area index according to historical data, and use the leaf area index as the input and the growth rate of the plant under this leaf area index as the output to train the neural network to obtain the trained Model 3;

[0014] The monitoring steps are:

[0015] Step 1: Obtain the land area occupied by the monitored plants and the number of monitored plants in the occupied land, and then obtain the average land area occupied by a single monitored plant based on the occupied land area and the number of monitored plants in the occupied land;

[0016] Step 2: Obtain an image of the monitored plant to be identified, and input the image of the monitored plant to be identified into model 1 to obtain the stem of the monitored plant and the front leaf image of the monitored plant in the monitored plant image;

[0017] Step 3: According to the internal parameters of the shooting camera, the area of the positive image of each monitored plant leaf in the monitored plant image to be identified is obtained, and the average area of the positive image of all leaves in the monitored plant image is obtained;

[0018] Step 4: Based on the internal parameters of the camera, the length of the monitored plant stem in the image of the monitored plant to be identified is obtained, and the length of the monitored plant stem is input into model 2 to obtain the corresponding number of leaves;

[0019] Step 5: Multiply the mean value of the positive image area of each monitored plant leaf obtained in step 3 by the number of corresponding leaves obtained in step 4 to obtain the sum of the areas of all leaves of a single plant;

[0020] Step 6: Calculate the leaf area index based on the area of all leaves of a single plant and the average value of the land area obtained in step 1;

[0021] Step 7: Input the leaf area index obtained in step 6 into model 3 to obtain the growth rate, thereby realizing plant growth information monitoring.

[0022] Furthermore, the specific steps of marking the positive leaf image of the monitored plant are:

[0023] Step 1: Obtain the image to be detected and compress it to obtain a compressed image;

[0024] Step 2: Input the compressed image into the trained object detection network for leaf detection;

[0025] The training process of the target detection network is:

[0026] Step 21: Acquire the original image and obtain a compressed image of the original image;

[0027] Step 22: Extract features from the original image and the compressed image respectively to obtain an original image feature map and a compressed image feature map, and use the detection network to detect the original image to obtain a category prediction value in the original image;

[0028] Step 23: Process the original image feature map and the compressed image feature map using Feature map distillation to obtain a compressed image feature map with a distribution similar to that of the original image feature map;

[0029] Step 24: Replace the original compressed image feature map with the compressed image feature map with a distribution similar to that of the original image feature map, and use the detection network to obtain the position and class prediction values of the objects in the compressed image;

[0030] Step 25: Process the class prediction values in the original image and the class prediction values obtained in Step 24 using Logit distillation to obtain class prediction values closer to the class prediction values in the original image;

[0031] Step 26: Use the compressed image of the original image as the input, the positions of the objects in the compressed image obtained in Step 24, and the class prediction values closer to the class prediction values in the original image as the output to train the object detection network.

[0032] Further, the loss function of the Logit distillation is expressed as:

[0033]

[0034] where represents the softmax function, represents the hyperparameter temperature, p t and p s represent the class prediction values of the teacher and the student respectively, and KL represents the KL divergence.

[0035] Further, the loss function of the Feature map distillation is expressed as:

[0036]

[0037] where F T and F S represent the output feature maps of ResNet50 in the teacher and student networks, L represents the number of convolutional layers of ResNet50, C, H, and W represent the number of channels, height, and width of the feature map respectively, f represents the feature map alignment operation, and l, c, h, and w represent the initial channels of the feature map.

[0038] Further, it also includes a step of repairing the monitored plant's positive leaf image obtained in Step 2, specifically:

[0039] Step 2-1: For the damaged positive leaf image, extract the area occupied by the damaged part in the leaf image to obtain a Mask map;

[0040] Step 2-2: Perform dilation processing on the Mask map;

[0041] Steps 2 and 3: Repair the damaged areas in the dilated Mask image to obtain a repaired leaf image.

[0042] Furthermore, if no damaged area is extracted in Step 2.1, the leaf is a leaf with missing contour, and the leaf is processed as a non-normal leaf image.

[0043] Furthermore, the specific steps for repairing the white areas in the Mask image in Step 3 are as follows:

[0044] First, invert the dilated Mask image, then multiply it with the damaged leaf image to obtain a color image with a mask. After that, superimpose the color image with the mask and the dilated Mask image based on channels to obtain a 4-channel image. Finally, input the 4-channel image into the lama model;

[0045] The 4-channel image first undergoes downsampling operation in the lama model, then passes through fast Fourier convolution processing, and finally undergoes upsampling to output the repaired leaf image;

[0046] Among them, during the fast Fourier convolution processing, the input tensor is divided into two parts based on channels and passed through the local branch and the global branch respectively. The local branch is used to extract local information, and the global branch is used to extract global information using fast Fourier convolution;

[0047] Finally, cross-fuse the local information and the global information, and then splice them based on channels to obtain new gray information for the damaged area of the leaf, that is, the repaired area.

[0048] The beneficial effects of the present invention are:

[0049] This application solves the problem of low accuracy in visually distinguishing the growth information of wetland plants at various stages by the naked eye. This application judges the growth information of plants through the leaf area index. This application combines neural networks, which improves the speed of plant monitoring. This application realizes continuous monitoring of the growth of plants, greatly improves the monitoring efficiency, and avoids the problems of low monitoring efficiency, time-consuming and laborious in manually visually observing the growth status of plants. Description of the Drawings

[0050] Figure 1 It is a schematic diagram of stem length and leaf number. Detailed Embodiments

[0051] It should be noted that, without conflict, the various embodiments disclosed in this application can be combined with each other.

[0052] Embodiment 1: Refer to Figure 1 This embodiment will specifically describe a method for monitoring plant growth information, which includes a training step and a monitoring step:

[0053] The specific training step is as follows:

[0054] Train Model 1:

[0055] For the monitored plant, obtain the image of the monitored plant, mark the stem of the monitored plant and the image of the positive leaves of the monitored plant in the image, and finally use the image of the monitored plant as the input and the marked image as the output to train the neural network to obtain the trained Model 1;

[0056] This application first trains Model 1, using the image of the monitored plant and the marked stem and positive leaf images of the monitored plant for training, so that after the test image is input into Model 1, the stem of the monitored plant and the positive leaves of the monitored plant in the image can be obtained.

[0057] Train Model 2:

[0058] For the monitored plant, obtain the length of the stem of the monitored plant and the corresponding number of leaves according to historical data, and use the length of the stem of the monitored plant as the input and the corresponding number of leaves as the output to train the neural network to obtain the trained Model 2;

[0059] Model 2 is trained according to historical data. The purpose is to input the stem of the monitored plant output by Model 1 and the actual length of the stem of the monitored plant obtained using the camera internal parameters into Model 2 to obtain the corresponding number of leaves. In fact, the stem length also reflects the growth information of the plant, and this application obtains the leaf area index through the number of leaves corresponding to the stem length at this growth rate. The data of Model 2 is as Figure 1 shown.

[0060] Train Model 3:

[0061] According to historical data, obtain the leaf area index of the monitored plant and the growth rate of the plant under this leaf area index, and use the leaf area index as the input and the growth rate of the plant under this leaf area index as the output to train the neural network to obtain the trained Model 3;

[0062] Model 3 monitors the growth rate of the plant according to the growth rate of the plant under each leaf area index.

[0063] The monitoring step is as follows:

[0064] Step 1: Obtain the land area occupied by the monitored plants and the number of the measured plants within the occupied land, and then obtain the average value of the land area occupied by a single monitored plant based on the occupied land area and the number of the measured plants within the occupied land;

[0065] Step 2: Obtain the image of the monitored plant to be recognized, and input the image of the monitored plant to be recognized into Model 1 to obtain the stem of the monitored plant and the front-facing leaf image of the monitored plant in the image of the monitored plant;

[0066] Step 3: According to the internal parameters of the camera, obtain the area of the front-facing image of each leaf of the monitored plant in the image of the monitored plant to be recognized, and obtain the average value of the areas of the front-facing images of each leaf of the monitored plant;

[0067] Step 4: According to the internal parameters of the camera, obtain the length of the stem of the monitored plant in the image of the monitored plant to be recognized, and input the length of the stem of the monitored plant into Model 2 to obtain the corresponding number of leaves;

[0068] Step 5: Multiply the average value of the areas of the front-facing images of each leaf of the monitored plant obtained in Step 3 by the corresponding number of leaves obtained in Step 4 to obtain the total area of all leaves;

[0069] Step 6: Obtain the leaf area index based on the total area of all leaves and the average value of the occupied land area obtained in Step 1;

[0070] Step 7: Input the leaf area index obtained in Step 6 into Model 3 to obtain the growth rate, thereby realizing the monitoring of plant growth information.

[0071] The leaf area index (LAI), also known as the leaf area coefficient, refers to the multiple of the total area of plant leaves per unit land area to the land area. That is: leaf area index = total leaf area / land area. In field experiments, the leaf area index (LAI) is an important indicator reflecting the growth status of plant populations, and its size is directly and closely related to the final yield.

[0072] This application first obtains the images of the monitored plants through a camera, and the actual sizes of each plant part in the images can be obtained according to the internal parameters of the camera.

[0073] The images obtained in this application include leaf images in various directions and postures. This application only identifies the forward-facing leaves in the current image and obtains the average value of the areas of the forward-facing leaves. When acquiring images of leaves, due to the large number of leaves and the variety of leaf images, it is not conducive to the extraction of leaf images. This application first obtains the stem length of the plant and the corresponding number of leaves based on historical data, trains a neural network with this, then obtains the forward-facing leaves in each image, and obtains the average value of the areas of the forward-facing leaves in the image, thereby reducing errors and solving the problem that it is not easy to obtain the forward-facing area of leaves in an image.

[0074] Specific Embodiment 2: This embodiment is a further description of Specific Embodiment 1. The difference between this embodiment and Specific Embodiment 1 is that the specific steps for marking the forward-facing leaf images of the monitored plant are as follows:

[0075] Step 1: Obtain the image to be detected and compress the image to be detected to obtain the compressed image;

[0076] Step 2: Input the compressed image into the trained object detection network for leaf detection;

[0077] The training process of the object detection network is as follows:

[0078] Step 21: Obtain the original image and obtain the compressed image of the original image;

[0079] Step 22: Extract features from the original image and the compressed image respectively to obtain the original image feature map and the compressed image feature map, and use the detection network to detect the original image to obtain the class prediction value in the original image;

[0080] Step 23: Use Feature map distillation to process the original image feature map and the compressed image feature map to obtain a compressed image feature map similar to the distribution of the original image feature map;

[0081] Step 24: Replace the original compressed image feature map with the compressed image feature map similar to the distribution of the original image feature map, and use the detection network to obtain the position and class prediction value of the object in the compressed image;

[0082] Step 25: Use Logit distillation to process the class prediction value in the original image and the class prediction value obtained in Step 24 to obtain a class prediction value close to the class prediction value in the original image (this is actually the role of knowledge distillation, making the prediction values gradually approach, and thus completing the "teaching" of the teacher network to the student network);

[0083] Step 26: Use the compressed image of the original image as the input, the positions of the objects in the compressed image obtained in Step 24, and the class prediction values close to the class prediction values in the original image as the output to train the object detection network.

[0084] This application utilizes knowledge distillation, where the object detection network is the student network trained in this application, and the detection network in Step 22 is the teacher network.

[0085] Traditional networks use the original-sized images, resulting in a very large computational load. Using thumbnail images, on the other hand, would result in loss of feature information. Therefore, although this application uses thumbnail images for calculation, it also uses knowledge distillation to retain feature information, which can not only improve computational efficiency but also avoid a decrease in accuracy caused by feature loss.

[0086] Specific Embodiment 3: This embodiment is a further elaboration of Specific Embodiment 2. The difference between this embodiment and Specific Embodiment 2 is that the loss function of the Logit distillation is expressed as:

[0087]

[0088] where, represents the softmax function, represents the hyperparameter temperature, p t and p s represent the class prediction values of the teacher and the student respectively, and KL represents the KL divergence.

[0089] Specific Embodiment 4: This embodiment is a further elaboration of Specific Embodiment 3. The difference between this embodiment and Specific Embodiment 3 is that the loss function of the Feature map distillation is expressed as:

[0090]

[0091] where, F T and F S represent the output feature maps of ResNet50 in the teacher and student networks, L represents the number of convolutional layers of ResNet50, C, H, and W represent the number of channels, height, and width of the feature map respectively, f represents the feature map alignment operation, and l, c, h, and w represent the initial channels of the feature map.

[0092] Specific Embodiment 5: This embodiment is a further elaboration of Specific Embodiment 1. The difference between this embodiment and Specific Embodiment 1 is that it further includes a step of repairing the positive leaf image of the monitored plant obtained in Step 2, specifically:

[0093] Step 2-1: For the damaged forward blade image, extract the area occupied by the damaged part in the blade image to obtain a Mask image;

[0094] Step 2-2: Perform dilation processing on the Mask image;

[0095] Step 2-3: Repair the damaged area in the dilated Mask image to obtain a repaired blade image.

[0096] Specific Embodiment 6: This embodiment is a further description of Specific Embodiment 5. The difference between this embodiment and Specific Embodiment 5 is that if no damaged area is extracted in Step 2-1, the blade is a blade with missing contour, and the blade is processed as a non-forward blade image.

[0097] The purpose of this application is to obtain the total blade area by obtaining the average value of the forward blade area and the number of leaves of a single plant in the same historical period. For damaged images, there are two cases. One is that the middle of the blade is damaged but the contour is intact, and in this case, the damaged area can be extracted through the solution in Specific Embodiment 5. The second case is that the contour of the blade is missing, and in this case, the blade is deleted because this application is to obtain the average value of forward blades, and the blade with missing contour will cause errors in the obtained average value. Therefore, the blade is deleted to reduce errors.

[0098] Specific Embodiment 7: This embodiment is a further description of Specific Embodiment 6. The difference between this embodiment and Specific Embodiment 6 is that the specific steps for repairing the white area in the Mask image in Step 3 are as follows:

[0099] First, invert the dilated Mask image, then multiply it with the damaged blade image to obtain a color image with a mask. After that, superimpose the color image with the mask and the dilated Mask image based on the channels to obtain a 4-channel image. Finally, input the 4-channel image into the lama model;

[0100] The 4-channel image first undergoes a downsampling operation in the lama model, then passes through a fast Fourier convolution process, and finally undergoes upsampling to output a repaired image without a red frame;

[0101] Among them, during the fast Fourier convolution process, the input tensor is divided into two parts based on the channels and passed through the local branch and the global branch respectively. The local branch is used to extract local information, and the global branch is used to extract global information using the fast Fourier convolution;

[0102] Finally, the local information and the global information are cross-fused and then spliced based on the channels to obtain new grayscale information for the damaged area of the blade, that is, the repaired area.

[0103] The original RGB image is used to extract the red frame to obtain the Mask image. There are two ways to extract the red frame: direct extraction from the RGB channels and extraction in the HSV color space. In the red channel, there are obvious gray value features. By setting the threshold Th, the area where the red frame is located can be roughly extracted as a white Mask image. First, convert the original RGB image to the HSV color space, and then construct a mask for the red pixels within the red threshold range in the HSV space to obtain the position of the red frame. Although the red frame cannot be clearly segmented in HSV, the position of the red frame can be accurately extracted. Although the displayed red color is imperfect and blurred, its advantage is that it will not display non-red content.

[0104] The Mask image is dilated to obtain the dilated Mask image. Dilation is an operation to find local maxima. The kernel B is convolved with the graph, that is, the maximum value of the pixel points in the area covered by the kernel B is calculated, and this maximum value is assigned to the pixel specified by the reference point. This will cause the highlighted area in the image to gradually grow. The dilated Mask image contains the complete red frame area.

[0105] The white area in the dilated Mask image is repaired using image processing methods or deep learning methods. The image processing method uses the FMM algorithm. The deep learning method uses the lama model. The repaired color image is converted to a grayscale image through color space conversion and can be used for training.

[0106] The FMM algorithm is used to repair and remove the red frame in the image processing method. The idea behind the FMM algorithm is to first process the pixel points on the edge of the area to be repaired, and then gradually push in layer by layer until all pixel points are repaired. For a point p to be repaired, we need to calculate the new gray value of point p using all the points in the neighborhood of this point. A weight function is used to determine which pixel values have a greater impact on the new pixel value and which have a smaller impact. The Fast Marching Method (FMM) is used to determine the order of the pixels to be repaired.

[0107] Based on the obvious gray value feature of the red frame in the red channel of the image to be repaired in this application, by setting the threshold Th, the area where the red frame is located can be roughly extracted as a white Mask image. After morphological dilation with a large kernel, a training mask mask for the final lama to be repaired is obtained. The main innovation points of lama are: proposing a new repair network structure, using fast Fourier convolution, having a wide image reception field, a high receptive field perception loss, and a relatively large training mask (mask), which can effectively improve the performance potential of the first two components. It can be well generalized to images with a higher resolution than during training, and achieve performance comparable to the benchmark with a lower number of parameters and computational cost.

[0108] Main ideas of the LaMa algorithm: 1. Use fast Fourier convolution (FFC) to increase the receptive field to improve the quality of image inpainting; 2. The bias in FFC makes the network more generalizable (it can also inpaint high-resolution images during low-resolution image training); 3. Use a larger mask for training to further improve the performance of the network. The lama model architecture is as follows: it includes a black-and-white mask image and an original image; after covering the masked image on the original image and inputting it into the Inpainting network, it is first downsampled to a low resolution, then passes through several fast Fourier convolution FFC residual blocks, and finally outputs an upsampled result to generate a high-resolution inpainted image. This feature of FFC improves the perceptual quality and network parameter efficiency. The inductive bias of FFC enables the network to obtain high resolution, which does not exist during training. FFC reduces the amount of training data and computational cost required.

[0109] Embodiment:

[0110] Obtain the land area occupied by the measured plant. Since this application is targeted at a single plant, the mean value of the land area occupied by a single plant is obtained here. If multiple plants are targeted during implementation, the leaf area index of multiple plants can also be obtained through the technical solution of this application, and then the growth rate of multiple plants can be obtained.

[0111] The following is an explanation for a single plant: First, obtain the image of the measured plant, then use Model 1 to obtain the image of the stem in the image, and according to the camera internal parameters, the actual stem length of the detected plant is 10 cm. Then, input the stem length into Model 2 to obtain the number of leaves under the same growth rate in the same historical period, which is 30. The number of positive leaves in the image of the measured plant is 6, among which 2 are damaged but have intact outlines, and 1 has a missing outline. Therefore, 5 positive leaves are obtained after inpainting. The sum of the areas of the 5 positive leaves is 140 cm 2 , so the mean leaf area is 28 cm 2 . The land area occupied by a single plant is 900 cm 2 . Then 28 cm 2 ×30 is the sum of the leaf areas, and the ratio of the sum of the areas to the occupied land area is the leaf area index. Finally, obtain the growth rate according to Model 3.

[0112] It should be noted that the specific implementation manners are only explanations and illustrations of the technical solutions of the present invention, and the scope of the right protection cannot be limited thereby. Any changes that are only partial based on the claims and specifications of the present invention should still fall within the protection scope of the present invention.

Claims

1. A method for monitoring plant growth information, characterized in that It includes a training step and a monitoring step: The specific training step is as follows: Train Model 1: For the monitored plant, obtain the image of the monitored plant, mark the stem of the monitored plant and the positive leaf image of the monitored plant in the image. Finally, use the image of the monitored plant as the input and the marked image as the output to train the neural network, and obtain the trained Model 1; Train Model 2: For the monitored plant, obtain the length of the stem of the monitored plant and the corresponding number of leaves according to the historical data, and use the length of the stem of the monitored plant as the input and the corresponding number of leaves as the output to train the neural network, and obtain the trained Model 2; Train Model 3: Obtain the leaf area index of the monitored plant and the growth rate of the plant under this leaf area index according to the historical data, and use the leaf area index as the input and the growth rate of the plant under this leaf area index as the output to train the neural network, and obtain the trained Model 3; The monitoring step is as follows: Step 1: Obtain the land area occupied by the monitored plant and the number of the monitored plants in the occupied land. Then, obtain the average value of the land area occupied by a single monitored plant according to the occupied land area and the number of the monitored plants in the occupied land; Step 2: Obtain the image of the monitored plant to be recognized, and input the image of the monitored plant to be recognized into Model 1 to obtain the stem of the monitored plant and the positive leaf image of the monitored plant in the image of the monitored plant; Step 3: According to the internal parameters of the shooting camera, obtain the area of the positive image of each leaf of the monitored plant in the image of the monitored plant to be recognized, and obtain the average value of the areas of all the positive leaf images in the image of the monitored plant; Step 4: According to the internal parameters of the shooting camera, obtain the length of the stem of the monitored plant in the image of the monitored plant to be recognized, and input the length of the stem of the monitored plant into Model 2 to obtain the corresponding number of leaves; Step 5: Multiply the average value of the areas of the positive images of each leaf of the monitored plant obtained in Step 3 by the corresponding number of leaves obtained in Step 4 to obtain the sum of the areas of all the leaves of a single plant; Step 6: Obtain the leaf area index according to the sum of the areas of all the leaves of a single plant and the average value of the occupied land area obtained in Step 1; Step 7: Input the leaf area index obtained in Step 6 into Model 3 to obtain the growth rate, thereby realizing the monitoring of plant growth information; The specific steps for marking the positive leaf image of the monitored plant are as follows: Step 1: Obtain the image to be detected, and compress the image to be detected to obtain the compressed image; Step 2: Input the compressed image into the trained object detection network for leaf detection; The training process of the object detection network is as follows: Step 21: Obtain the original image and obtain the compressed image of the original image; Step 22: Extract features from the original image and the compressed image respectively to obtain the original image feature map and the compressed image feature map, and use the detection network to detect the original image to obtain the class prediction value in the original image; Step 23: Use Feature map distillation to process the original image feature map and the compressed image feature map to obtain a compressed image feature map with a distribution similar to that of the original image feature map; Step 24: Replace the original compressed image feature map with a compressed image feature map similar to the distribution of the original image feature map, and use the detection network to obtain the position and class prediction values of the objects in the compressed image; Step 25: Use Logit distillation to process the class prediction values in the original image and the class prediction values obtained in Step 24 to obtain class prediction values closer to those in the original image; Step 26: Use the compressed image of the original image as the input, the positions of the objects in the compressed image obtained in Step 24, and the class prediction values closer to those in the original image as the output to train the object detection network.

2. The method for monitoring plant growth information according to claim 1, characterized in that The loss function of the Logit distillation is expressed as: Among them, represents the softmax function, represents the hyperparameter temperature, p t and p s represent the class prediction values of the teacher and the student respectively, and KL represents the KL divergence.

3. The method for monitoring plant growth information according to claim 2, characterized in that The loss function of the Feature map distillation is expressed as: Among them, F T and F S represent the output feature maps of ResNet50 in the teacher and student networks. L represents the number of convolutional layers of ResNet50. C, H, and W represent the number of channels, height, and width of the feature maps respectively. f represents the feature map alignment operation, l c, h, and w represent the initial channels of the feature maps.

4. A method for monitoring plant growth information according to claim 1, characterized in that It further includes a step of repairing the monitored plant's positive leaf image obtained in Step 2, specifically: Step 2-1: For the damaged positive leaf image, extract the area occupied by the damaged part in the leaf image to obtain a Mask image; Step 2-2: Perform dilation processing on the Mask image; Step 2-3: Repair the damaged area in the dilated Mask image to obtain a repaired leaf image.

5. The method for monitoring plant growth information according to claim 4, wherein If no damaged area is extracted in Step 2-1, then this leaf is a leaf with missing contour, and this leaf is processed as a non-positive leaf image.

6. The method for monitoring plant growth information according to claim 4, wherein The specific steps for repairing the white area in the Mask image in Step 3 are: First, invert the dilated Mask image, then multiply it with the damaged leaf image to obtain a color image with a mask. After that, superimpose the color image with the mask and the dilated Mask image based on the channels to obtain a 4-channel image. Finally, input the 4-channel image into the lama model; The 4-channel image first undergoes a downsampling operation in the lama model, then passes through fast Fourier convolution processing, and finally undergoes upsampling to output the repaired leaf image; Among them, during the fast Fourier convolution processing, the input tensor is divided into 2 parts based on the channels, and they pass through the local branch and the global branch respectively. The local branch is used to extract local information, and the global branch is used to extract global information using fast Fourier convolution; Finally, the local information and the global information are cross-fused and then concatenated based on the channels to obtain new gray information for the damaged area of the leaf, that is, the repaired area.

Images