Method for evaluating the planting characteristics of a sugarcane transplanter

Abstract

The application discloses a kind of evaluation methods of planting characteristics of sugarcane transverse planter, comprising the following steps: (1) using unmanned aerial vehicle to obtain sugarcane plant image in cane field, the data set is obtained by processing sugarcane plant image, data set includes training set and test set;(2) picture data set is marked;(3) real density map is obtained;(4) model improvement;(5) using improved CSRNet neural network to the feature extraction of training set mark file, generates predicted density map;(6) model training;(7) the image of test set that is marked is input into trained network model, generates predicted density map, and the pixel point probability value of density map is summed, to obtain predicted quantity.The application method can quickly and accurately extract crop planting area, stage crop growth detection and carry out crop yield estimation.

Description

Technical Field

[0001] This invention relates to the field of intelligent sugarcane planting technology, and in particular to a method for evaluating the planting characteristics of a sugarcane horizontal planting machine. Background Technology

[0002] Currently, most sugarcane planting machines developed and put into use in China are real-time seed-cutting sugarcane planting machines. These machines can sequentially complete processes such as furrowing, fertilizing, cutting seeds, covering with soil, and laying mulch. However, these machines require manual feeding of sugarcane seeds into the seed cutter for real-time cutting and planting. This results in high workload for operators, uneven feeding speed and quantity, excessive seed usage, and a tendency to miss seeds. The seed usage is approximately 0.8-1.2 tons per acre. Furthermore, the real-time feeding and cutting of whole sugarcane stalks reduces planting efficiency and leads to uneven planting density. Traditional sugarcane planting involves placing sugarcane longitudinally along the furrows, but this method requires more seeds, results in shallower effective seed depth, lower seedling emergence per unit area, and makes the sugarcane susceptible to lodging.

[0003] The applicant has developed a pre-cut seeding double-bud sugarcane horizontal planting machine and applied for a patent. Patent application publication number CN 115413435 A discloses a sugarcane horizontal planting machine with continuous seed supply and precision planting capabilities. This machine can complete multiple planting processes in one operation, including double-row ditching, fertilization, seeding, replanting, and covering with soil. A series of studies have been conducted on various subsystems such as the deep ditching system, continuous seed supply system, horizontal seeding system, precision fertilization system, and real-time replanting system. It achieves precise planting of sugarcane using pre-cut seeding, improving production efficiency, reducing labor intensity, lowering costs, and increasing sugarcane yield per acre. Pre-cut seeding ensures that sugarcane is planted "deeply and buried shallowly," requiring less seed but resulting in a high germination rate, a large number of effective plants, and reduced lodging. Clearly, the pre-cut seeding double-bud sugarcane horizontal planting machine has a significant advantage in sugarcane quality compared to other planting methods (such as real-time seeding sugarcane planting machines).

[0004] Traditional sugarcane yield assessment relies primarily on manual on-site surveys in sugarcane fields to determine differences in sugarcane growth status (seed quantity, budding rate, lodging resistance, etc.) and yield. Previously, evaluating the planting quality of different methods and comparing growth status relied on random manual sampling. This method is not only time-consuming and labor-intensive, but also inefficient, as it only samples specific areas and cannot eliminate the possibility of sampling errors, failing to observe the overall growth status. Therefore, a more comprehensive and efficient monitoring method is needed to assess the planting quality of pre-cut seed double-bud sugarcane horizontal planting machines.

[0005] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a method for evaluating the planting characteristics of a sugarcane horizontal planting machine, and to conduct quality assessments of different sugarcane planting methods.

[0007] To achieve the above objectives, the present invention provides a method for evaluating the planting characteristics of a sugarcane horizontal planting machine, comprising the following steps:

[0008] (1) Use drones to acquire images of sugarcane plants in sugarcane fields, process the sugarcane plant images to obtain a dataset, which includes a training set and a test set;

[0009] (2) Label the image dataset from step (1) to obtain the label files that correspond one-to-one with the training set and the test set, and obtain the real coordinate information of the sugarcane plants.

[0010] (3) Generate a ground truth map based on the annotation file of the training set in step (2). Then, perform two-dimensional Gaussian blur processing on the ground truth map to obtain the true density map. The summation of the density maps is the number of sugarcane cores.

[0011] (4) Improved using CSRNet neural network as model. CSRNet neural network mainly includes front-end network and back-end network. A channel attention module is introduced between the front-end network and the back-end network. The front-end network mainly performs two-dimensional feature extraction on the input image. The channel attention module adjusts the feature map channel by channel to reduce the error estimation of the background. The features extracted by the channel attention module are fused and sent to the back-end network. The back-end network is used to generate density estimation map.

[0012] (5) Use the improved CSRNet neural network to extract features from the training set annotation file of step (2) and generate a prediction density map;

[0013] (6) Calculate the density map loss using the real density map of the training set in step (3) and the predicted density map in step (5), introduce an adaptive loss function into the model for calculation, and then train the improved neural network model to obtain the model with the best performance and save it.

[0014] (7) Input the test set images labeled in step (2) into the trained network model to generate a prediction density map, and then sum the pixel probability values ​​of the density map to obtain the prediction number.

[0015] Preferably, in the above technical solution, the ratio of the training set to the test set in the dataset of step (1) is 3:1.

[0016] Preferably, in the above technical solution, the annotation in step (2) is to use the labelme annotation tool to label the dataset. Each sugarcane plant image after labeling will generate a JSON annotation file, which contains the coordinate information of each marked point, that is, the real location of each sugarcane plant. Then the coordinate information is converted into mat format for storage.

[0017] Preferably, in the above technical solution, the method for obtaining the true density map in step (3) includes:

[0018] A ground truth map with annotations is generated using a .mat file, and then the ground truth map is subjected to a two-dimensional Gaussian blur. When marking a sugarcane core, the center position of the sugarcane core is denoted as x. i Then the marked point is represented as δ(xx) i Therefore, a sugarcane image containing N sugarcane pith labels can be represented by the function H(x):

[0019]

[0020] Performing a two-dimensional Gaussian convolution on H(x) diffuses the sugarcane core into a range, resulting in a continuous density map, i.e.:

[0021]

[0022] Where F(x) represents the generated density map; G(x) is a normalized two-dimensional Gaussian kernel; * represents convolution operation; and the diffusion range of the sugarcane core is positively correlated with the average distance of its K nearest neighbor cores. This indicates that at a distance x from the sugarcane core i The average distance between the K nearest sugarcane cores, β is a weight; G(x) is normalized, and the summation over the density map is the number of sugarcane cores.

[0023] Preferably, in the above technical solution, in step (4) the CSRNet neural network, the front-end network is a feature extraction structure composed of a VGG16 backbone network; the VGG16 with the fully connected layers removed is used as the backbone network of the model, and the size of the output image is 1 / 8 of the original input image; the back-end network is a dilated convolutional layer, which is used to expand the receptive field while keeping the image resolution unchanged, and generate a high-quality density map.

[0024] Preferably, in the above technical solution, the first 13 layers of the VGG16 network are used as the front end, including 10 convolutional layers and 3 pooling layers. There are 3 pooling layers with filter = 2x2 and stride = 2, and the convolutions all use convolutional layers with padding = 1, stride = 1, and filter = 3x3. At the same time, the fully connected layers are removed to adapt to changes in the size of the input image.

[0025] Preferably, in the above technical solution, the formula for the dilated convolutional layer of the backend network is as follows:

[0026]

[0027] x(m,n) is the input image, w(i,j) is the filter, and r is the dilation rate; when r=1, dilated convolution is just a normal convolution operation.

[0028] Preferably, in the above technical solution, step (4) the channel attention module is used to calculate a set of probability values ​​to establish a dependency between channels, and then perform a weighted summation on all channel mappings to update the original channels;

[0029] The channel attention module processes a feature map of size C×H×W through a 1×1 convolution, resulting in two feature maps: C×HW and HW×C. C1 is C×HW, C2 is HW×C, and C3 is C×HW. First, the correlation between channels is calculated. Then, C1 and C2 matrices are multiplied and softmax normalized to obtain a channel attention feature map C5 of size C×C. This is then multiplied by C3 to obtain an attention map of size C×H×W. Finally, a weighted sum is performed with the original image after the 1×1 convolution. The channel attention process can be represented as:

[0030]

[0031] Indicate C i The value at position j, C j The value at j represents the input image, and μ represents the learnable parameter.

[0032] Preferably, in the above technical solution, the training method in step (6) includes:

[0033] 1) The density map loss of the training set can be calculated by comparing the true density map with the predicted density map of the training set.

[0034] 2) Sum the pixel probability values ​​of the actual density map and the predicted density map to obtain the actual number and predicted number of sugarcane plants, and then calculate the quantity loss of the two.

[0035] 3) The two losses are weighted and summed, and the model begins training, gradually causing the loss value to converge; this process introduces an adaptive loss function for calculation.

[0036] Preferably, in the above technical solution, step 3) of introducing the adaptive loss function includes: using the loss function, the formula of which is as follows:

[0037]

[0038] Where θ represents the network's weight parameters, N represents the training batch size; θ represents all parameters of CSRNet; X i This represents the input image; Represents the ground truth value;

[0039] However, for slightly sparse sugarcane samples, the density map loss is usually very small. Therefore, the small loss value for sparse sugarcane samples is easily overwhelmed by the large loss value for dense sugarcane samples, leading to unfair treatment of sparse sugarcane samples during the training phase. This leads to another adaptive loss function:

[0040]

[0041] Wherein, C(X) i ;θ) and This involves predicting the number of sugarcane cores and the actual number of cores. When combining density map loss and plant number loss, these two losses should be on the same order of magnitude; otherwise, the larger loss will dominate, weakening the impact of the smaller loss. Therefore, to balance the two losses, an adaptive loss function is introduced, which can be expressed as:

[0042] L==α*L D +β*L c

[0043] Where α and β are weighting factors; instead of using fixed values, α and β are set to vary with L. D and L c Adaptive weights that vary according to the ratio between them; adaptive weights make L D and L c The values ​​are always on the same order of magnitude, allowing the network to treat the two losses fairly.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] (1) The method for evaluating the planting characteristics of a sugarcane horizontal planting machine of the present invention involves the use of a drone to first capture large-scale images of sugarcane, then transmitting the images back, and finally performing relevant machine vision processing for further comprehensive analysis and research. This process enables observation in large-scale scenarios and is more flexible and efficient than manual monitoring, and can be applied to the evaluation of planting quality in pre-cut seed double-bud sugarcane horizontal planting machines. Compared with traditional statistical methods, it largely avoids interference from human factors, effectively saving time, manpower, material resources, and financial resources, and can quickly and accurately extract crop planting area, monitor crop growth, and estimate crop yield.

[0046] (2) Convolutional neural networks estimate the corresponding density map from the input image, integrate the density map, and finally calculate the target number. When sugarcane is in the seedling stage, the emergence rate can be calculated based on the density map and the number of plants, combined with the amount of seed used. During the inter-cultivation period, the same method can be used to obtain data on various aspects of sugarcane growth. Finally, at the maturity stage, which is the sugarcane harvesting period, no manual yield measurement is required; instead, the sugarcane yield data for the year can be obtained based on the density map and the specific number of plants at the current stage. Dynamic monitoring is carried out in stages during the sugarcane growth process to obtain growth data for each stage, so that we can have a more intuitive understanding of the planting quality assessment of the pre-cut seed double-bud segment sugarcane horizontal planting machine. Attached Figure Description

[0047] Figure 1 This is a flowchart of the training process for the calculation algorithm for estimating sugarcane plant size based on the evaluation method of sugarcane horizontal planting machine according to the present invention.

[0048] Figure 2 This is a test flowchart of the algorithm for estimating sugarcane plant growth using the evaluation method for the planting characteristics of a sugarcane horizontal planter according to the present invention.

[0049] Figure 3 This is a diagram of the CSRNet network architecture according to the method of the present invention;

[0050] Figure 4 This is a flowchart of the improved network model according to the method of the present invention;

[0051] Figure 5 This is a framework diagram of the spatial attention module in the method according to the present invention;

[0052] Figure 6 This is the original image of a sugarcane plant captured by a drone;

[0053] Figure 7 Is it using labelme to Figure 5 Mark the sugarcane plants in the middle section;

[0054] Figure 8This is the original image of the sugarcane plant used for the test;

[0055] Figure 9 yes Figure 8 The generated true density map;

[0056] Figure 10 yes Figure 8 and Figure 9 The generated predicted density map;

[0057] in, Figure 9 and 10 The brighter the color in a region, the higher the density of the sugarcane core; conversely, the darker the color, the sparser the sugarcane core. Detailed Implementation

[0058] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings, but it should be understood that the scope of protection of the present invention is not limited to the specific embodiments.

[0059] Unless otherwise expressly stated, throughout the specification and claims, the term "comprising" or its variations such as "including" or "comprises" shall be understood to include the stated elements or components without excluding other elements or other components.

[0060] This application focuses on a sugarcane field planted by a pre-cutting horizontal planting machine in Guangxi Agricultural Science and Technology New City, Fusui County, Chongzuo City, Guangxi Zhuang Autonomous Region. The sugarcane planted in the experimental field is Zhongzhe No. 9, which is the first year of ratooning planted by the horizontal planting machine. The planting area is 42 mu (approximately 2.8 hectares), with a row spacing of 140 cm and an average of 15 plants per meter. The growth cycle is 7 months.

[0061] Experimental data was collected on October 29, 2022. All experiments used a DJI Phantom 4 Pro 2.0 drone to photograph sugarcane under clear weather conditions. The flight altitude was 10 to 20 meters. The sugarcane planter used was a pre-cut seed double-bud sugarcane horizontal planter, which is patented (patent publication number CN 115413435 A). The evaluation method in this application is not limited to evaluating this device; it can also be used to evaluate other methods of sugarcane planting.

[0062] like Figure 1-8 As shown, a method for evaluating the planting characteristics of a sugarcane horizontal planting machine includes the following steps:

[0063] I. Data Collection

[0064] Use drones to collect visible light images of the sugarcane canopy during the sugarcane's maturity period (growth cycle of about seven months) as the input raw images; take a rectangular quadrat that is not adjacent to the edge, specifically: take 10 rows, each row is 11 meters long, and take a whole image (used to calculate the number of effective stems per acre).

[0065] II. Image Preprocessing

[0066] Blurred sugarcane images were removed, and the images after adaptive contrast enhancement were cropped, resulting in a total of 801 sugarcane images. From the acquired original images, several images were preprocessed and then divided into a test set and a training set at a 1:3 ratio, with 206 images in the test set and 595 images in the training set, and the two sets having no overlap.

[0067] III. Annotated Datasets

[0068] A labeling tool called Labelme was used to label the sugarcane plant data. Each labeled sugarcane plant image generates a JSON label file. This file contains the coordinates of each labeled point, i.e., the actual location of each sugarcane plant. Label files corresponding one-to-one with the training and test sets were obtained, capturing the actual coordinates of the sugarcane plants. Finally, this coordinate information was converted to MAT format for storage.

[0069] Using the annotation file from the training set in step two, a ground truth map with basic ground truth points is generated using a .mat file. Then, the ground truth map is subjected to a two-dimensional Gaussian blur. When marking a sugarcane core, the center position of the sugarcane core is denoted as x. i Then the marked point is represented as δ(xx) i Therefore, a sugarcane image containing N sugarcane pith labels can be represented by the function H(x):

[0070]

[0071] Performing a two-dimensional Gaussian convolution on H(x) diffuses the sugarcane core into a range, resulting in a continuous density map, i.e.:

[0072]

[0073] Where F(x) represents the generated density map; G(x) is a normalized two-dimensional Gaussian kernel; * represents convolution operation; and the diffusion range of the sugarcane core is positively correlated with the average distance of its K nearest neighbor cores. This indicates that at a distance x from the sugarcane core i The average distance between the K nearest sugarcane hearts, with β as a weight. Since G(x) is normalized, the summation over the density map gives the number of sugarcane hearts.

[0074] IV. Improved Network Model

[0075] An improvement is made using the CSRNet neural network as the model. The CSRNet neural network mainly consists of a front-end network and a back-end network. A channel attention module is introduced between the front-end network and the back-end network. The front-end network mainly performs two-dimensional feature extraction on the input image. The channel attention module adjusts the feature map channel by channel to reduce the error estimation of the background. The features extracted by the channel attention module are fused and fed into the back-end network. The back-end network is used to generate a density estimation map.

[0076] The front-end network is a feature extraction structure composed of a VGG16 backbone network; the back-end network is a dilated convolutional layer used to generate a density estimation map.

[0077] CSRNet uses a VGG16 network with fully connected layers removed as its backbone. The output image size is 1 / 8 of the original input image. As the number of convolutional layers increases, the output image also becomes smaller, which increases the difficulty of generating density maps. Therefore, CSRNet uses a dilated convolutional neural network as its backend network to expand the receptive field while maintaining the image resolution, in order to generate high-quality density maps.

[0078] CSRNet uses the first 13 layers of a pre-trained VGG-16 network as its front end, which includes 10 convolutional layers and 3 pooling layers (max-pooling). There are 3 pooling layers with filter = 2x2 and stride = 2, and all convolutions use padding = 1, stride = 1, and filter = 3x3. At the same time, fully connected layers are removed to accommodate changes in input image size.

[0079] CSRNet uses dilated convolution layers (or dilated convolution layers) in the backend. Their purpose is to achieve a larger receptive field without pooling, thus avoiding the accuracy loss caused by pooling. In convolutional neural networks, the quality of feature extraction depends on the size of the receptive field. Larger kernels result in a larger receptive field, capturing more information from the image and extracting better features. However, simply enlarging the kernel significantly increases computation, hindering the construction of deep networks and degrading network performance. Furthermore, input images typically undergo pooling, which can expand the receptive field while reducing computation. However, pooling can lose some important image features. Dilated convolution is a compromise, expanding the receptive field without increasing network parameters.

[0080] The most crucial part of our network is the dilated convolution layer, as shown in the following formula:

[0081]

[0082] x(m,n) is the input image, w(i,j) is the filter, and r is the dilation rate. When r=1, dilated convolution is simply a regular convolution operation.

[0083] The attention mechanism originates from research on human vision. In cognitive science, due to the bottleneck of information processing, humans selectively focus on a portion of all information while ignoring other visible information. To make efficient use of limited visual information processing resources, humans need to select specific parts of the visual region and then concentrate on them. For example, when reading, people typically only pay attention to and process a small number of words. Similar to human visual attention, the attention mechanism in deep learning also uses attention models to acquire more important features from the target and ignore ineffective information that does not contribute significantly to the current task. The core goal of the attention model is to select the most critical information for the current target from a large amount of information.

[0084] By introducing a channel attention model, a set of probability values ​​can be calculated to establish a dependency between channels. Then, a weighted summation of all channel mappings can be performed to update the original channels. The channel attention model can extract foreground features more efficiently, avoiding errors in estimation caused by background.

[0085] The channel attention module takes a feature map of size C×H×W and performs a 1×1 convolution to obtain two feature maps, C×HW and HW×C, where C1 is C×HW, C2 is HW×C, and C3 is C×HW. First, the correlation between channels is calculated. Then, the multiplication of matrices C1 and C2 is followed by softmax normalization to obtain a channel attention feature map C5 of size C×C. This is then multiplied by C3 to obtain an attention map of size C×H×W. Finally, a weighted sum is performed with the original image after the 1×1 convolution. The channel attention process can be represented as:

[0086]

[0087] Indicate C i The value at position j, C j Let represent the value of the input image at position j, and μ represent the learnable parameters. (Spatial attention module) Figure 6 As shown.

[0088] V. Model Training

[0089] In terms of network training weight initialization, the first 10 layers of the VGG-16 network with extracted basic features were initialized using the pre-trained model parameters, the standard deviation of the Gaussian kernel was initialized to 0.01, the optimizer was Adam, the batch size during the training phase was set to 1, and the number of iterations was 400.

[0090] During training, the input training set images are first fed into the network for feature extraction, generating predicted density maps. The size of each density map is consistent with the size of the corresponding input image. Then, based on the point coordinate files generated from the labeled data, a fixed Gaussian kernel is applied to each coordinate point to generate the corresponding true density map. The training methods include:

[0091] 1) The density map loss of the training set can be calculated by comparing the true density map with the predicted density map of the training set.

[0092] 2) Sum the pixel probability values ​​of the actual density map and the predicted density map to obtain the actual number and predicted number of sugarcane plants, and then calculate the quantity loss of the two.

[0093] 3) The two losses are weighted and summed, and the model begins training, gradually allowing the loss values ​​to converge; this process incorporates an adaptive loss function for calculation. Then, the improved neural network model is trained to obtain the optimal-performing model, which is then saved.

[0094] The commonly used loss function is Euclidean, and its formula is as follows:

[0095]

[0096] Where θ represents the network's weight parameters, N represents the training batch size; θ represents all parameters of CSRNet; X i This represents the input image; Represents the ground truth value.

[0097] However, for slightly sparse sugarcane samples, the density map loss is usually very small. Therefore, the small loss value for sparse sugarcane samples is easily overwhelmed by the large loss value for dense sugarcane samples, leading to unfair treatment of sparse sugarcane samples during the training phase. This leads to another adaptive loss function:

[0098]

[0099] Where C(X) i ;θ) and This involves predicting the number of sugarcane cores and the actual number of cores. When combining density map loss and plant number loss, these two losses should be on the same order of magnitude. Otherwise, the larger loss will dominate, weakening the impact of the smaller loss. Therefore, to balance the two losses, an adaptive loss function is introduced, which can be expressed as:

[0100] L==α*L D +β*L c

[0101] Where α and β are weighting factors. Instead of using fixed values, α and β are set to vary with L. D and L c The adaptive weights vary according to the ratio between them. The adaptive weights make L... D and L c The values ​​are always on the same order of magnitude, thus enabling the network to treat the two losses fairly.

[0102] VI. Testing and Experimentation

[0103] The test images from the test set are input into the trained model to generate a high-quality prediction density map. Then, the pixel probability values ​​of the density map are summed to obtain the number of predictions.

[0104] The image quality obtained after training in this application is as follows: Figure 9 , Figure 10 As shown.

[0105] After determining the number of sugarcane plants in the sampling plot (i.e., the total number of effective stems surveyed), the number of effective stems per meter is calculated using the formula: Number of effective stems per meter = Total number of effective stems surveyed ÷ (Survey area, i.e., number of rows × row length). Then, the number of effective stems per mu (unit of land area) is calculated from the number of effective stems per meter using the formula: Number of effective stems per mu = Number of effective stems per meter × 666.67 ÷ Average row spacing.

[0106] The foregoing description of specific exemplary embodiments of the invention is for illustrative and explanatory purposes. These descriptions are not intended to limit the invention to the precise forms disclosed, and it will be apparent that many changes and variations can be made in accordance with the foregoing teachings. The exemplary embodiments were chosen and described in order to explain the specific principles of the invention and its practical application, thereby enabling those skilled in the art to implement and utilize various different exemplary embodiments of the invention, as well as various different choices and variations. The scope of the invention is intended to be defined by the claims and their equivalents.

Claims

1. A method for evaluating the planting characteristics of a sugarcane horizontal planting machine, characterized in that, Includes the following steps: (1) Use drones to acquire images of sugarcane plants in sugarcane fields, process the sugarcane plant images to obtain a dataset, which includes a training set and a test set; (2) Label the image dataset from step (1) to obtain the label files that correspond one-to-one with the training set and the test set, and obtain the real coordinate information of the sugarcane plants. (3) Generate a ground truth map based on the annotation file of the training set in step (2). Then, perform two-dimensional Gaussian blur processing on the ground truth map to obtain the true density map. The summation of the density maps is the number of sugarcane cores. (4) Improved using CSRNet neural network as model. CSRNet neural network mainly includes front-end network and back-end network. A channel attention module is introduced between the front-end network and the back-end network. The front-end network mainly performs two-dimensional feature extraction on the input image. The channel attention module adjusts the feature map channel by channel to reduce the error estimation of the background. The features extracted by the channel attention module are fused and sent to the back-end network. The back-end network is used to generate density estimation map. (5) Use the improved CSRNet neural network to extract features from the training set annotation file of step (2) and generate a prediction density map; (6) Calculate the density map loss using the real density map of the training set in step (3) and the predicted density map in step (5), introduce an adaptive loss function into the model for calculation, and then train the improved neural network model to obtain the model with the best performance and save it. (7) Input the test set images labeled in step (2) into the trained network model to generate a prediction density map, and then sum the pixel probability values ​​of the density map to obtain the prediction number.

2. The method for evaluating the planting characteristics of a sugarcane horizontal planting machine according to claim 1, characterized in that, In step (1), the ratio of the training set to the test set in the dataset is 3:

1.

3. The method for evaluating the planting characteristics of a sugarcane horizontal planting machine according to claim 1, characterized in that, The annotation in step (2) involves using the labelme annotation tool to label the dataset. Each sugarcane plant image after labeling will generate a JSON annotation file containing the coordinate information of each labeled point, i.e., the actual location of each sugarcane plant. The coordinate information is then converted into MAT format for storage.

4. The method for evaluating the planting characteristics of a sugarcane horizontal planting machine according to claim 1, characterized in that, Step (3) involves obtaining the true density map using the following methods: A ground truth map with annotations is generated using a .mat file, and then the ground truth map is subjected to a two-dimensional Gaussian blur. When marking a sugarcane core, the center position of the sugarcane core is denoted as x. i Then the marked point is represented as δ(xx) i Therefore, a sugarcane image containing N sugarcane pith labels can be represented by the function H(x): Performing a two-dimensional Gaussian convolution on H(x) diffuses the sugarcane core into a range, resulting in a continuous density map, i.e.: Where F(x) represents the generated density map; G(x) is a normalized two-dimensional Gaussian kernel; * represents convolution operation; and the diffusion range of the sugarcane core is positively correlated with the average distance of its K nearest neighbor cores. This indicates that at a distance x from the sugarcane core i The average distance between the K nearest sugarcane cores, β is a weight; G(x) is normalized, and the summation over the density map is the number of sugarcane cores.

5. The method for evaluating the planting characteristics of a sugarcane horizontal planting machine according to claim 1, characterized in that, Step (4) The front-end network of the CSRNet neural network is a feature extraction structure composed of a VGG16 backbone network; VGG16 with the fully connected layers removed is used as the backbone network of the model, and the size of the output image is 1 / 8 of the original input image; The backend network consists of dilated convolutional layers, which are used to expand the receptive field while maintaining the image resolution and generating high-quality density maps.

6. The method for evaluating the planting characteristics of a sugarcane horizontal planting machine according to claim 5, characterized in that, The first 13 layers of the VGG16 network were used as the front end, including 10 convolutional layers and 3 pooling layers. There were 3 pooling layers with filter=2x2 and stride=2, and the convolutions were all convolutional layers with padding=1, stride=1, and filter=3x3. At the same time, fully connected layers were removed to adapt to changes in the size of the input image.

7. The method for evaluating the planting characteristics of a sugarcane horizontal planting machine according to claim 5, characterized in that, The formula for the dilated convolutional layer in the backend network is as follows: x(m,n) is the input image, w(i,j) is the filter, and r is the dilation rate; when r=1, dilated convolution is just a normal convolution operation.

8. The method for evaluating the planting characteristics of a sugarcane horizontal planting machine according to claim 1, characterized in that, Step (4) The channel attention module is used to calculate a set of probability values ​​to establish a dependency between channels, and then to perform a weighted summation of all channel mappings to update the original channels; The channel attention module processes a feature map of size C×H×W through a 1×1 convolution, resulting in two feature maps: C×HW and HW×C. C1 is C×HW, C2 is HW×C, and C3 is C×HW. First, the correlation between channels is calculated. Then, C1 and C2 matrices are multiplied and softmax normalized to obtain a channel attention feature map C5 of size C×C. This is then multiplied by C3 to obtain an attention map of size C×H×W. Finally, a weighted sum is performed with the original image after the 1×1 convolution. The channel attention process can be represented as: C represents i The value at position j, C j The value at j represents the input image, and μ represents the learnable parameter.

9. The method for evaluating the planting characteristics of a sugarcane horizontal planting machine according to claim 1, characterized in that, The training methods in step (6) include: 1) The density map loss of the training set can be calculated by comparing the true density map with the predicted density map of the training set. 2) Sum the pixel probability values ​​of the actual density map and the predicted density map to obtain the actual number and predicted number of sugarcane plants, and then calculate the quantity loss of the two. 3) The two losses are weighted and summed, and the model begins training, gradually causing the loss value to converge; this process introduces an adaptive loss function for calculation.

10. The method for evaluating the planting characteristics of a sugarcane horizontal planting machine according to claim 9, characterized in that, Step 3) Methods for introducing an adaptive loss function include: the loss function used, the formula of which is as follows: Where θ represents the network's weight parameters, N represents the training batch size; θ represents all parameters of CSRNet; X i This represents the input image; Represents the ground truth value; Introducing another adaptive loss function: Wherein, C(X) i ;θ) and This involves predicting the number of sugarcane cores and the actual number of cores. When combining density map loss and plant number loss, these two losses should be on the same order of magnitude; otherwise, the larger loss will dominate, weakening the impact of the smaller loss. Therefore, to balance the two losses, an adaptive loss function is introduced, which can be expressed as... L==α*L D +β*L c Where α and β are weighting factors; instead of using fixed values, α and β are set to vary with L. D and L c Adaptive weights that vary according to the ratio between them; adaptive weights make L D and L c The values ​​are always on the same order of magnitude, allowing the network to treat the two losses fairly.

Images