End-to-end seedling strip identification method and device based on deep learning instance segmentation

Abstract

The invention discloses an end-to-end seedling strip identification method and device based on deep learning instance segmentation, and belongs to the technical field of computer vision and artificial intelligence. The method comprises the steps of segmenting seedling strip pixels and a background based on a segmentation network, and segmenting different seedling strips in a high-dimensional space; feature fusion is carried out on the segmentation result of the seedling band pixel and background segmentation and the segmentation results of different seedling bands in a high-dimensional space in a pixel space to obtain a feature fusion representation, the feature fusion representation is processed based on a position coding matrix to obtain a final feature representation, and the final feature representation is processed based on a block classification head network to obtain a final feature representation; endowing each pixel block represented by the final features with a corresponding seedling zone ID (Identity); and processing the seedling strips with the same type of seedling strip IDs by adopting a least square method, and setting the number of times of fitting a polynomial to obtain seedling strip parameters. The method is high in precision and generalization, and seedling strip identification can be realized without a high-requirement manual annotation data set.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision and artificial intelligence technology, and specifically relates to an end-to-end seedling strip recognition method and device based on deep learning instance segmentation. Background Art

[0002] Agriculture provides material resources for human production and life, and is the foundation of human survival and development. Since the 21st century, artificial intelligence technology has made rapid progress, which has promoted the development of all walks of life towards unmanned and intelligent. In order to realize intelligent agriculture, the accuracy and real-time performance of field automatic navigation technology are extremely important. As a landmark reference of the farmland environment, crop rows can guide the movement and operation of agricultural robots. An effective seedling belt recognition algorithm can greatly improve the accuracy of agricultural robot navigation and operation. It can provide spatial constraints for subsequent precision fertilization, precision weeding, and precise control of intelligent agricultural machinery, thereby achieving the beneficial effects of saving fertilizers and pesticides and increasing per-acre yields, effectively reducing labor intensity, reducing environmental pollution, reducing production costs, and improving resource utilization and agricultural product yield and quality.

[0003] At present, the methods for seedling belt recognition can be roughly divided into two methods. The first is a single-stage method, which directly extracts the parameters of the seedling belt row through a neural network, without the need to obtain the seedling belt points in advance. The seedling belt recognition task is usually regarded as a regression task of the seedling belt line. By setting a variety of anchor frame lines, the precise seedling belt recognition task is finally completed. However, this method is limited by the setting method of the anchor frame, and the detected seedling belt can only be linear, and it is powerless for curved seedling belts; the second is a two-stage method, which obtains the seedling belt points in advance, and then clusters the seedling belt points to obtain the seedling belt rows of each category.

[0004] The two-stage method can be roughly divided into two types. One is to obtain the crop center first and then obtain the seedling strip row. The other is to obtain the crop row directly without obtaining the crop center. However, both methods need to use the least squares method or Hough transform to extract the seedling strip parameters to obtain the final seedling strip parameters. The way to obtain the crop center can be summarized into two types. One is to use traditional image processing methods to extract the crop center, but this method has poor adaptability and cannot face the complex and changeable farmland environment. In addition, the hyperparameters need to be specifically set by experts, which restricts its subsequent development. Second, the YOLO algorithm is used to extract the center of the crop frame. Although this method has strong anti-interference ability, it still cannot get rid of the clustering algorithm with high time complexity and cannot guarantee real-time performance. In addition, the method that does not require the acquisition of the crop row center is basically based on the semantic segmentation method (U-Net, MobileNetV2 + DeepLabV3, ENet, etc.). This method usually relies on a large number of high-demand manually labeled seedling strip data sets, and it is also unable to get rid of the time-consuming clustering algorithm. Summary of the invention

[0005] To solve the above technical problems, the present invention provides an end-to-end seedling belt recognition method and device based on deep learning instance segmentation, which uses a more flexible two-stage method, takes seedling belt row recognition as an instance segmentation task, distinguishes seedling belts from background information at the pixel level, and realizes end-to-end extraction of seedling belt points without the use of clustering algorithms. It is a high-precision, highly generalized, flexible, real-time seedling belt recognition method that does not require high requirements for manually labeled data sets.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0007] An end-to-end seedling band identification method based on deep learning instance segmentation, the method comprising:

[0008] Step 1: Segmenting the seedling strip pixels from the background based on a segmentation network, and segmenting different seedling strips in a high-dimensional space, wherein the segmentation network is trained using seedling strip data that is not precisely labeled;

[0009] Step 2: feature fusion is performed in the pixel space on the segmentation result of the seedling strip pixel and the background segmentation result and the segmentation result of different seedling strips in the high-dimensional space to obtain a feature fusion representation, the feature fusion representation is processed based on the position encoding matrix to obtain a final feature representation, the final feature representation is processed based on the block classification head network, and each pixel block of the final feature representation is assigned a corresponding seedling strip ID;

[0010] Step 3: Use the least squares method to process the seedling belts with the same type of seedling belt ID, and obtain the seedling belt parameters by setting the degree of the fitting polynomial.

[0011] On the other hand, the present invention provides an end-to-end seedling band identification device based on deep learning instance segmentation, comprising:

[0012] A segmentation module, used for segmenting seedling strip pixels from background based on a segmentation network, and segmenting different seedling strips in a high-dimensional space, wherein the segmentation network is trained using seedling strip data that is not precisely labeled;

[0013] The encoding module is used to perform feature fusion in the pixel space on the segmentation results of the seedling strip pixels and the background and the segmentation results of different seedling strips in the high-dimensional space to obtain a feature fusion representation, process the feature fusion representation based on the position encoding matrix to obtain a final feature representation, process the final feature representation based on the block classification head network, and assign a corresponding seedling strip ID to each pixel block of the final feature representation;

[0014] The calculation module is used to process the seedling belts with the same type of seedling belt ID using the least square method, and obtain the seedling belt parameters by setting the degree of the fitting polynomial.

[0015] In a third aspect, the present invention provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned end-to-end seedling band identification method based on deep learning instance segmentation.

[0016] In a fourth aspect, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enables the processor to implement the aforementioned end-to-end seedling band identification method based on deep learning instance segmentation.

[0017] The beneficial effects of the present invention are:

[0018] It has strong generalization ability and uses deep neural network method to cope with complex and changeable farmland environment without setting various complicated clustering algorithm hyperparameters.

[0019] High flexibility, the type of extraction strip (straight line, curve, etc.) can be set according to the task objectives;

[0020] Strong real-time performance, successfully abandoning the clustering algorithm, greatly reducing time complexity, and extracting seedling points end-to-end;

[0021] The labor cost is low, and the weakly supervised training paradigm is used. The labeled data set does not need to specify the seedling belt category, and does not need to be accurate to the pixel point;

[0022] High precision, introducing connected branches, learning the correlation between pixels, and avoiding the generation of broken seedling belts. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a principle diagram of an end-to-end seedling band identification method based on deep learning instance segmentation of the present invention;

[0024] Figure 2 Input the original image and annotation map to the network, where (a) is the network input image, (b) is the seedling background segmentation annotation map, (c) is the embedding segmentation annotation map, and (d) is the connectivity segmentation annotation map;

[0025] Figure 3 This is the network structure diagram for the seedling strip extraction and separation task;

[0026] Figure 4 Instantiate the network structure diagram for the seedling belt;

[0027] Figure 5 The intermediate results of the seedling strip extraction and separation task are shown in Figure 1, where (a) is the seedling strip background segmentation map and (b) is the high-dimensional embedding segmentation map.

[0028] Figure 6It is an intermediate result image of feature fusion and position encoding enhancement in pixel space;

[0029] Figure 7 This is the result image after extracting the seedling strip parameters and mapping it back to the original image;

[0030] Figure 8 The following are the experimental results under different environments. DETAILED DESCRIPTION

[0031] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0032] The present invention adopts an instance-based segmentation method, uses imprecisely labeled seedling belt data to train the model, abandons the clustering algorithm (such as DBSCAN) introduced by traditional semantic segmentation, greatly improves the algorithm running speed, and the inference speed on a single 2080Ti is greater than 100FPS. It is a high-precision, generalized, flexible, real-time, and end-to-end seedling belt recognition method that does not require high-demand manually labeled data sets. It makes up for the problem that the previous single-stage algorithm cannot obtain flexible seedling belt lines, solves the high latency problem caused by the introduction of clustering algorithms in the two-stage algorithm, and alleviates the current situation that the previous algorithm needs to accurately label data sets.

[0033] like Figure 1 As shown, the method includes: the first stage (seedling band extraction and separation) is completed in the feature space, the purpose of which is to separate the background from the seedling band and distinguish different seedling bands. The second stage (seedling band instantiation) is completed in the pixel space, the purpose of which is to give the previously extracted and separated seedling band position ID, and to use a pixel block-based classification head to speed up the network reasoning speed. Through these two stages, the seedling band points are finally obtained end-to-end, and the final seedling band parameters can be obtained by using a fitting algorithm such as the least squares method.

[0034] like Figure 2 As shown in the figure, this paper adopts an instance segmentation-based method and uses the imprecisely labeled seedling belt data to train the model. At the same time, in order to reduce the workload of labeling, this method does not need to manually label the categories, but divides them according to the connectivity of the seedling belt. Figure 2 The actual network input seedling belt image shown in (a) is first used to click on the seedling belt center in turn to obtain the connected seedling belt lines, as shown in Figure 2 The seedling belt background segmentation and annotation map shown in (b) is then obtained according to the connectivity of the seedling belt, that is, Figure 2 The embedded segmentation annotation diagram shown in (c) is shown in Figure 1. At the same time, the connected branch is introduced in the seedling strip extraction and separation, and the annotations required are as follows: Figure 2 The connected segmentation annotation map shown in (d) is constructed by giving a seedling belt background segmentation map. , and perform connectivity analysis on each pixel to construct a directional connectivity annotation graph. , if its value is 1 (seedling point), check whether the points separated by 2 pixels (Manhattan distance) in the 8 main directions (horizontal, vertical and diagonal directions) are also 1. If so, the connectivity annotation of this direction is set to 1, otherwise it is set to 0; if the pixel itself is 0 (i.e. background pixel), the connectivity annotations in all directions are 0. Finally, 8 binary annotation maps are obtained, each of which corresponds to the connectivity information in one direction. This method can effectively capture local structures at a larger scale, enhance the expression ability of connected areas, and provide a basis for further tasks such as boundary optimization and connected area analysis. The formula is as follows:

[0035] ,

[0036] In the above formula, d represents 8 main directions (horizontal, vertical and diagonal directions), and Represents the pixel offset value, which is determined according to the direction. Thus, the construction of the weakly supervised seedling belt dataset is completed.

[0037] Specifically, each step is introduced below:

[0038] Step 1, extracting and separating the seedling strips; Figure 3 The figure shows the network structure of the seedling strip extraction and separation task, which uses an efficient segmentation network (such as BiseNetV2, PIDnet, STDC) as the backbone network and constructs the task head through convolutional layers and upsampling layers. The goal of this task can be divided into two parts: one is to effectively segment the seedling strip pixels from the background, and the other is to accurately distinguish different seedling strips in high-dimensional space.

[0039] In order to effectively segment the seedling strip pixels from the background and solve the problem of misidentification of broken seedling strip blocks in the process of segmenting the seedling strip pixels from the background, two branches are introduced: the segmentation branch and the connection branch. The two branches have the same structure and are implemented by combining the backbone network with the task head. Figure 2 The parameters of the seedling belt background segmentation annotation information in (b) are used for parameter training. In order to ensure the continuity of the seedling belt, a connected branch is introduced to prevent the generation of broken seedling belt blocks and further optimize the segmentation results. The connected branch is based on Figure 2 (d) The connectivity segmentation annotation information is used for training, thereby effectively improving the integrity of the seedling belt.

[0040] To accurately distinguish different seedlings in high-dimensional space, we still use the backbone network plus task head structure, introduce the embedding branch, and use Figure 2The loss function of this task aims to maximize the inter-class difference and minimize the intra-class difference, thus ensuring that different seedling bands can be effectively separated in high-dimensional space.

[0041] Specifically, Figure 3 As shown, the segmentation network used by the backbone network (such as BiseNetV2, PIDnet, STDC) aims to extract Figure 2 The features (C1×H / 8×W / 8) of the original input image (3×H×W) of the network shown in (a). The task head consists of two convolutional layers and a linear interpolation layer, and its main function is to map the image from the feature space to the pixel space. The first convolutional layer is responsible for extracting local features and enhancing spatial information to capture richer contextual relationships; the second convolutional layer is used for channel transformation to match the number of annotated channels to ensure that the network output is aligned with the target label. Finally, the linear interpolation layer restores the size of the feature map output by the two convolutional layers to the original input image size to facilitate the calculation of losses and ensure accurate supervised learning during training. For different tasks, the output of the task head has different numbers of channels. For Head1 corresponding to the segmentation branch, its main purpose is to extract the seedling band, so its number of channels is 2 (whether it belongs to the seedling band point); for Head2 corresponding to the connected branch, its main purpose is to enhance the connectivity information of the seedling band, specifically to check whether the 8 directions of the pixel adjacent to the pixel are seedling band points, so its number of channels is 8. For the Head3 branch corresponding to the embedding branch, the number of channels depends on the embedding dimension. The number of embedding dimensions faces the balance of "accuracy-complexity". After many experiments, the embedding dimension is finally selected as 4. For the three task heads, three corresponding loss functions are introduced to train the network based on the error back propagation algorithm. The task heads corresponding to the segmentation branch and the connected branch use the category-unbalanced cross entropy loss function (the loss functions of the segmentation branch and the connected branch are marked with Loss1 and Loss2), and the formula is as follows:

[0042] ,

[0043] Where C represents the number of categories, Indicates category The weight of Indicates category The label value of Represents the probability value of the model output. The weight calculation formula is as follows:

[0044] ,

[0045] in, Indicates category The constant 1.02 is used to avoid the abnormal situation where the numerator is 0.

[0046] Loss3 is the embedding loss, which can be summarized into three parts: the first is the variance loss, the smaller the variance value of the same seedling line element, the better. The second is the distance loss, the distance between different seedling line elements should be as far as possible. At the same time, in order to avoid overfitting, regularization loss is added as the third part. The three parts are used together as a comprehensive loss function.

[0047] Step 2: Instantiation of seedling strips: The goal of the seedling strip instantiation stage is to assign position IDs to the previously separated seedling strips. First, the output results of the seedling strip segmentation branch and the seedling strip embedding branch are multiplied pixel by pixel in the pixel space to obtain a feature fusion representation. In order to introduce position feature information, a position encoding matrix is ​​constructed, and the definition of the matrix is ​​shown in the following formula:

[0048] ,

[0049] Finally, the feature fusion representation is added to the position encoding matrix to obtain the final feature representation. This feature representation is then input into the block classification head network composed of stacked convolutional layers, and the network assigns a corresponding seedling band ID to each pixel block. Through steps 1 and 2, the probability of each category of each 8×8 pixel block in a series of images can be obtained. By setting a threshold, a corresponding seedling band ID is assigned to each pixel block whose probability exceeds the threshold.

[0050] like Figure 4, the seedling belt instantiation involves two parts: feature fusion and block classification head network. Feature fusion is performed in pixel space, which is intuitive, highly interpretable, and easy to add preprocessing settings (such as the selection of ROI area). The purpose of introducing position encoding is that position encoding symbolizes the prior information of seedling belts to a certain extent, and the use of position encoding can better distinguish seedling belts; on the other hand, the introduction of position encoding can concretize the meaning of seedling belt categories, and seedling belt categories represent the order of seedling belt positions. This avoids the difficulty of seedling belt tracking that may be encountered in subsequent applications. The ultimate goal of the algorithm is to extract seedling belt points end-to-end and fit the relevant parameters of the seedling belt line based on the seedling belt points. In order to balance the "accuracy-complexity" problem, a block classification architecture is used to predict the seedling belt category to which 8×8 pixels belong. In order to improve the reasoning speed of the network, a fully convolutional network (FCN) is used to complete the classification task. Among them, the block classification head network consists of three convolutional layers with a stride of 2 and a convolution kernel of 3×3 and one convolutional layer with a stride of 1 and a convolution kernel of 1×1. The first three 3×3 convolutional layers are mainly used to gradually extract deep features. At the same time, by setting the step size to 2, the feature map size is halved layer by layer, thereby reducing the amount of calculation and improving the computational efficiency of the model. The last 1×1 convolutional layer is responsible for channel mapping and feature fusion to adapt to the final classification task. The operation of halving the size of the feature map three times in a row can effectively reduce the number of seedling points that need to be processed later, thereby significantly reducing the computational overhead, improving the efficiency of fitting the seedling parameters, and ensuring the accuracy of classification. Experiments have shown that after feature fusion and position encoding, the block classification head network can quickly and accurately extract seedling points end-to-end after training.

[0051] Step 3, extracting seedling belt parameters;

[0052] The least square method is used to process the seedling belt points with the same category ID. By setting the degree of the fitting polynomial, various types of seedling belt parameters can be flexibly obtained.

[0053] Example

[0054] The method is verified based on the present invention according to actual application.

[0055] Step 1, extracting and separating the seedling strips;

[0056] After labeling the dataset, randomly sample from the original dataset and divide it into training set, test set, validation set, etc. according to the ratio of 8:1:1. Input the images into two sub-branches respectively to separate the seedling strip pixels from the background (e.g. Figure 5 (a)), and the two tasks of separating different seedling bands in high-dimensional space (as shown in Figure 5In order to avoid the appearance of broken seedling strip blocks in the step of separating the seedling strip pixels from the background, a connected branch is introduced to learn the correlation between pixels.

[0057] Step 2, instantiation of seedling belt;

[0058] In this stage, the pixels of the seedling strip are separated from the background (e.g. Figure 5 (a)), and the task of separating different seedling bands in high-dimensional space (as shown in Figure 5 The result obtained by position encoding is fused in pixel space as shown in (b). Figure 6 The result is shown in Figure 1. After that, a block classification head network is used to assign a specific category (background, specific seedling belt category) to each pixel block to finally complete the seedling belt recognition task. So far, the seedling belt points have been extracted end-to-end through deep learning.

[0059] Step 3, extracting seedling belt parameters;

[0060] After the strips are instantiated, similar strips are integrated together, the strip parameters are calculated using the least squares method, and then remapped back to the original image. The final running effect of this method is as follows Figure 7 and Figure 8 As shown in the figure, the seedling strip curve fitted by quadratic polynomial is highly similar to the annotation style. For easy observation, the fitting parameters are displayed in the upper left corner of the image to facilitate comparison and analysis. At the same time, in order to verify the adaptability of this method in different environments, a variety of scene tests were carried out (such as Figure 8 As shown in the figure, from left to right, there are complex situations such as partial seedling strip missing, shadow occlusion, small seedling strip size, and sparse seedling strip. The experimental results show that this method can maintain a stable fitting effect under different challenging conditions, proving its good robustness and universality.

[0061] This method focuses on the practical problems of low navigation accuracy, high seedling injury rate and low real-time performance in the rugged farmland environment. The single-stage method extracts the parameters of the curved seedling belt, and the two-stage method based on segmentation has poor real-time performance. An end-to-end seedling belt navigation algorithm based on instance segmentation is proposed. The improved STDC-Net combines position encoding, feature block classification and connected branches to obtain continuous seedling belt lines. The recognition accuracy of CRDLD in the public dataset is 96%, and the FPS reaches 122 frames on a single 2080Ti.

[0062] On the other hand, the present invention provides an end-to-end seedling band identification device based on deep learning instance segmentation, and each module included in the device can implement each step of the aforementioned method, specifically including:

[0063] A segmentation module, used for segmenting seedling strip pixels from background based on a segmentation network, and segmenting different seedling strips in a high-dimensional space, wherein the segmentation network is trained using seedling strip data that is not precisely labeled;

[0064] The encoding module is used to perform feature fusion in the pixel space on the segmentation results of the seedling strip pixels and the background and the segmentation results of different seedling strips in the high-dimensional space to obtain a feature fusion representation, process the feature fusion representation based on the position encoding matrix to obtain a final feature representation, process the final feature representation based on the block classification head network, and assign a corresponding seedling strip ID to each pixel block of the final feature representation;

[0065] The calculation module is used to process the seedling belts with the same type of seedling belt ID using the least square method, and obtain the seedling belt parameters by setting the degree of the fitting polynomial.

[0066] In a third aspect, the present invention provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned end-to-end seedling band identification method based on deep learning instance segmentation.

[0067] In a fourth aspect, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enables the processor to implement the aforementioned end-to-end seedling band identification method based on deep learning instance segmentation.

[0068] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An end-to-end seedling band identification method based on deep learning instance segmentation, characterized in that: The method comprises: Step 1: Segmenting the seedling strip pixels from the background based on a segmentation network, and segmenting different seedling strips in a high-dimensional space, wherein the segmentation network is trained using seedling strip data that is not precisely labeled; Step 2: feature fusion is performed in the pixel space on the segmentation result of the seedling strip pixel and the background segmentation result and the segmentation result of different seedling strips in the high-dimensional space to obtain a feature fusion representation, the feature fusion representation is processed based on the position encoding matrix to obtain a final feature representation, the final feature representation is processed based on the block classification head network, and each pixel block of the final feature representation is assigned a corresponding seedling strip ID; Step 3: Use the least squares method to process the seedling belts with the same type of seedling belt ID, and obtain the seedling belt parameters by setting the degree of the fitting polynomial.

2. According to the end-to-end seedling band identification method based on deep learning instance segmentation according to claim 1, it is characterized in that: The step 1 includes the segmentation network segmenting the seedling band pixels from the background based on the segmentation branch and the connected branch, and segmenting different seedling bands in a high-dimensional space based on the embedding branch; wherein the segmentation branch uses the seedling band background segmentation annotation information for parameter training, the connected branch uses the connectivity segmentation annotation information for training, and the embedding branch uses the embedded segmentation annotation information for training.

3. The end-to-end seedling band identification method based on deep learning instance segmentation according to claim 2 is characterized in that: The segmentation branch, the connected branch, and the embedding branch all adopt a backbone network combined with a task head. The backbone network is used to extract the features of the original input image, and the task head is used to map the features of the original input image to the pixel space.

4. The end-to-end seedling band identification method based on deep learning instance segmentation according to claim 3 is characterized in that: The task head includes two convolutional layers and a linear interpolation layer connected in sequence, wherein the first convolutional layer is used to extract local features of the original input image and enhance spatial information, the second convolutional layer is used for channel transformation to match the labeled number of channels to ensure that the network output is aligned with the target label, and the linear interpolation layer is used to restore the size of the feature maps output by the two convolutional layers to the size of the original input image.

5. The end-to-end seedling band identification method based on deep learning instance segmentation according to claim 3 is characterized in that: The task heads corresponding to the segmentation branch and the connected branch use the category-imbalanced cross entropy loss function; the task head corresponding to the embedding branch uses variance loss, distance loss, and regularization loss as the comprehensive loss function; the cross entropy loss function and the comprehensive loss function train the segmentation network based on the error back propagation algorithm.

6. The end-to-end seedling band identification method based on deep learning instance segmentation according to claim 1 is characterized in that: The step 2 comprises: The segmentation results of the seedling strip pixels and the background are multiplied pixel by pixel in the pixel space with the segmentation results of different seedling strips in the high-dimensional space to obtain the feature fusion representation; Add the feature fusion representation to the position encoding matrix to obtain the final feature representation; The final feature representation is input into a block classification head network composed of stacked convolutional layers. The block classification head network calculates the probability of the category of each pixel block in the final feature representation, sets a threshold, and assigns a corresponding seedling band ID to each pixel block whose probability exceeds the threshold.

7. The end-to-end seedling band identification method based on deep learning instance segmentation according to claim 6 is characterized in that: The block classification head network includes three sequentially connected convolution layers with a stride of 2 and a convolution kernel of 3×3 and one convolution layer with a stride of 1 and a convolution kernel of 1×1.

8. An end-to-end seedling band identification device based on deep learning instance segmentation, characterized in that: include: A segmentation module, used for segmenting seedling strip pixels from background based on a segmentation network, and segmenting different seedling strips in a high-dimensional space, wherein the segmentation network is trained using seedling strip data that is not precisely labeled; The encoding module is used to perform feature fusion in the pixel space on the segmentation results of the seedling strip pixels and the background and the segmentation results of different seedling strips in the high-dimensional space to obtain a feature fusion representation, process the feature fusion representation based on the position encoding matrix to obtain a final feature representation, process the final feature representation based on the block classification head network, and assign a corresponding seedling strip ID to each pixel block of the final feature representation; The calculation module is used to process the seedling belts with the same type of seedling belt ID using the least square method, and obtain the seedling belt parameters by setting the degree of the fitting polynomial.

9. An electronic device, characterized in that: include: one or more processors; A memory for storing one or more programs; Wherein, when one or more programs are executed by the one or more processors, the one or more processors implement the end-to-end seedling band identification method based on deep learning instance segmentation as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: Executable instructions are stored thereon, which, when executed by a processor, enable the processor to implement the end-to-end seedling band identification method based on deep learning instance segmentation as described in any one of claims 1-7.

Images