Parcel segmentation method, device, apparatus and medium

Abstract

The application provides a plot segmentation method, device, equipment and medium, and relates to the field of remote sensing images, comprising: obtaining the maximum pooling layer output and the encoder output feature after down-sampling the target image according to the upper-layer encoder, and traversing all levels of the encoder until the maximum pooling layer output and the encoder output feature are obtained; updating the encoder output feature according to the high-dimensional space feature to obtain the decoder connection input; inputting the decoder connection input of the bottom layer and the up-sampling result of the bottom-layer encoder into the bottom-layer decoder, taking the output result of the bottom-layer decoder as the up-sampling result of the decoder of the previous level, traversing all levels of the decoder until the target image segmentation result output by the upper-layer decoder is obtained. The application can essentially solve the problem of salt and pepper, simultaneously perform the three tasks of the range, boundary and distance to the plot boundary of the planting plot, realize the conversion from the instance segmentation to the semantic segmentation, and improve the precision of plot segmentation.

Description

Technical Field

[0001] This invention relates to the field of remote sensing imagery, and more particularly to a method, apparatus, device, and medium for land parcel segmentation. Background Technology

[0002] The smallest unit of classification in remote sensing imagery is the pixel. Each pixel has different temporal spectral characteristics, so the types can be distinguished based on the temporal spectral characteristics of each pixel. Due to the high heterogeneity of the ground, even the temporal spectral characteristics at the pixel level within a single plot are not the same. Differences in crop growth, soil background, phenology, etc., make pixel-based crop classification results prone to salt-and-pepper noise, affecting the final crop classification results.

[0003] Currently, pixel-oriented classification results are commonly used to reduce the impact of salt-and-pepper noise on crop classification results, but the above methods cannot fundamentally solve the problem of salt-and-pepper noise generation. Summary of the Invention

[0004] This invention provides a method, apparatus, equipment, and medium for plot segmentation to address the current technical shortcomings that cannot fundamentally solve the problem of salt and pepper noise. This invention proposes a method to effectively avoid salt and pepper noise, which is more consistent with real crop planting conditions, and makes the crop classification results after plot segmentation more closely resemble real application scenarios.

[0005] In a first aspect, the present invention provides a method for dividing land parcels, comprising:

[0006] Based on the downsampled target image from the upper encoder, obtain the downsampled max pooling layer output and encoder output features. Iterate through all encoder levels until the corresponding max pooling layer output and encoder output features for each level are obtained.

[0007] Update the encoder output features corresponding to each level based on the high-dimensional space features to obtain the decoder connection input corresponding to each level;

[0008] After the input of the lower-level decoder is connected to the input and the upsampling result of the lower-level encoder, the output result of the lower-level decoder is used as the upsampling result of the upper-level decoder. The process is repeated for all levels of decoders until the target image segmentation result output by the upper-level decoder is obtained.

[0009] The high-dimensional spatial features are determined based on the encoder output features, high-level features, and number of channels corresponding to each level;

[0010] The target image segmentation result is determined after segmenting the target image based on the land parcel boundary features, land parcel range features, and land parcel boundary distance features.

[0011] According to the land parcel segmentation method provided by the present invention, the step of traversing all levels of encoders until the max pooling layer output and encoder output features corresponding to each level are obtained includes:

[0012] Based on the encoder output of the next level, downsample the output of the current level's max pooling layer to obtain the output of the next level's max pooling layer and the encoder output features of the next level;

[0013] Iterate through all encoder levels until you obtain the max pooling layer output and encoder output features corresponding to each level.

[0014] According to a land parcel segmentation method provided by the present invention, the step of updating the encoder output features corresponding to each level based on high-dimensional spatial features to obtain the decoder connection input corresponding to each level includes:

[0015] The higher-level features are determined based on the width, height, and preset multiple of the encoder output features corresponding to each level;

[0016] The first convolutional layer is determined by processing the encoder output features and the number of feature channels based on the linear rectification function;

[0017] The second convolutional layer is determined based on the high-level features and the number of feature channels;

[0018] The first convolutional layer and the second convolutional layer are fused to obtain the third convolutional layer;

[0019] The third convolutional layer and the number of feature channels are processed according to the neural network activation function to obtain the decoder connection input corresponding to each layer.

[0020] According to the land parcel segmentation method provided by the present invention, the step of traversing all levels of decoders until the target image segmentation result output by the upper-level decoder is obtained includes:

[0021] The current level's decoder input and the next level's max pooling layer output are fed into the current level's decoder to obtain the current level's max pooling layer output.

[0022] Iterate through all levels of the decoder until you obtain the fourth, fifth, and sixth convolutional layers output by the upper-level decoder;

[0023] The fourth convolutional layer determines the land parcel boundary features of the target image, the fifth convolutional layer determines the land parcel range features, and the sixth convolutional layer determines the land parcel boundary distance features.

[0024] After segmenting the target image based on the parcel boundary features, parcel range features, and parcel boundary distance features, the target image segmentation result is obtained.

[0025] According to a land parcel segmentation method provided by the present invention, before downsampling the target image based on the upper-layer encoder, the method further includes:

[0026] Extract the land parcel boundary features from each sample image to obtain a first sample set, input the first sample set into the first segmentation model, and train a second segmentation model based on the output of the first segmentation model.

[0027] Extract the land parcel range features from each of the sample images to obtain a second sample set, and input the second sample set into the second segmentation model to train a third segmentation model based on the output of the second segmentation model;

[0028] Extract the land parcel boundary distance features from each of the sample images to obtain a third sample set, and input the third sample set into the third segmentation model to train a land parcel segmentation model based on the output of the third segmentation model;

[0029] The encoders and decoders at all levels are determined based on the land parcel segmentation model.

[0030] According to the land parcel segmentation method provided by the present invention, after training the land parcel segmentation model based on the output result of the third segmentation model and before downsampling the target image based on the upper encoder, the method further includes:

[0031] After obtaining the land parcel boundary features of the sample image, the first loss function is calculated;

[0032] After obtaining the land parcel range features of the sample image, the second loss function is calculated;

[0033] After obtaining the land parcel boundary distance features of the sample images, the third loss function is calculated;

[0034] The first loss function, the second loss function, and the third loss function are processed according to a weighted algorithm to determine the total loss function;

[0035] The land parcel segmentation model is processed according to the total loss function, and the target image is input into the processed land parcel segmentation model.

[0036] According to a land parcel segmentation method provided by the present invention, the first loss function, the second loss function, and the third loss function are processed according to a weighted algorithm to determine the total loss function, including:

[0037] The first weighting function is determined based on the first loss function and the first weighting value;

[0038] The second weighting function is determined based on the second loss function and the second weighting value;

[0039] The third weighting function is determined based on the third loss function and the third weighting value;

[0040] The total loss function is determined based on the first weighting function, the second weighting function, and the third weighting function;

[0041] The first weighted value is greater than the second weighted value;

[0042] The second weighted value is greater than or equal to the third weighted value.

[0043] In a second aspect, the present invention provides a land parcel partitioning device, comprising:

[0044] Acquisition Unit: Used to acquire the downsampled max pooling layer output and encoder output features based on the downsampled target image from the upper encoder, and to traverse all encoder levels until the corresponding max pooling layer output and encoder output features for each level are acquired.

[0045] Update unit: used to update the encoder output features corresponding to each level according to the high-dimensional space features, so as to obtain the decoder connection input corresponding to each level;

[0046] Processing unit: After inputting the input of the lower-level decoder and the upsampling result of the lower-level encoder to the lower-level decoder, the output result of the lower-level decoder is used as the upsampling result of the upper-level decoder. The process is repeated for all levels of decoders until the target image segmentation result output by the upper-level decoder is obtained.

[0047] The high-dimensional spatial features are determined based on the encoder output features, high-level features, and number of channels corresponding to each level;

[0048] The target image segmentation result is determined after segmenting the target image based on the land parcel boundary features, land parcel range features, and land parcel boundary distance features.

[0049] Thirdly, an electronic device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the land parcel partitioning method when executing the program.

[0050] Fourthly, a non-transitory computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the land parcel partitioning method.

[0051] This invention provides a method, apparatus, device, and medium for land parcel segmentation. It updates the encoder output features corresponding to each level based on high-dimensional spatial features to obtain the decoder connection input corresponding to each level. The input to the bottom-level decoder connection and the upsampling result of the bottom-level encoder are fed to the bottom-level decoder. The output result of the bottom-level decoder is used as the upsampling result of the next-level decoder. This process is repeated across all levels of decoders until the target image segmentation result output by the next-level decoder is obtained. This invention effectively solves the problems of gradient explosion and vanishing gradients, and fundamentally addresses the issue of salt-and-pepper generation. It simultaneously performs three tasks: determining the range, boundary, and distance to the parcel boundary of the planting area. A multi-task learning convolutional neural network is constructed. By using the distance to the parcel boundary parameter, the conversion from instance segmentation to semantic segmentation is achieved, improving the accuracy of land parcel segmentation. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0053] Figure 1 This is one of the flowcharts illustrating the land parcel division method provided by the present invention;

[0054] Figure 2 This is a schematic diagram of the process for obtaining the output of the maximum pooling layer and the encoder output features corresponding to each level, provided by the present invention.

[0055] Figure 3 This is a schematic diagram of the process for obtaining the decoder connection input corresponding to each level, provided by the present invention.

[0056] Figure 4 This is a schematic diagram of the process for obtaining the target image segmentation result output by the upper-layer decoder provided by the present invention;

[0057] Figure 5 This is the second flowchart of the land parcel division method provided by the present invention;

[0058] Figure 6 This is the third flowchart of the land parcel division method provided by the present invention;

[0059] Figure 7 This is a flowchart illustrating the process of determining the total loss function provided by the present invention;

[0060] Figure 8 This is one of the structural schematic diagrams of the land parcel division device provided by the present invention;

[0061] Figure 9 This is a schematic diagram of the structure of the residual block provided by the present invention;

[0062] Figure 10 This is a schematic diagram of the spatial attention module provided by the present invention;

[0063] Figure 11 This is the second schematic diagram of the land parcel division device provided by the present invention;

[0064] Figure 12 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0066] Existing remote sensing image crop classification and mapping methods are mainly implemented through two approaches: pixel-based and object-oriented. Object-oriented methods cluster remote sensing images based on spectral similarity, forming clusters of pixels with similar spectra, thus creating individual objects. With the rapid development of deep learning technology, this invention can perform visual plot segmentation by learning the contextual information of the plot scene. By aggregating the results with pixel-based methods, a plot-based crop classification result is formed. Specifically, plot-based crop distribution extraction can effectively avoid salt-and-pepper noise, better reflect real crop planting conditions, and facilitate the implementation and monitoring of plot-based measures such as crop replanting, making the crop classification result closer to real application scenarios.

[0067] Figure 1 This is one of the flowcharts illustrating the land parcel division method provided by the present invention. The present invention discloses a land parcel division method, comprising:

[0068] Based on the downsampled target image from the upper encoder, obtain the downsampled max pooling layer output and encoder output features. Iterate through all encoder levels until the corresponding max pooling layer output and encoder output features for each level are obtained.

[0069] Update the encoder output features corresponding to each level based on the high-dimensional space features to obtain the decoder connection input corresponding to each level;

[0070] After the input of the lower-level decoder is connected to the input and the upsampling result of the lower-level encoder, the output result of the lower-level decoder is used as the upsampling result of the upper-level decoder. The process is repeated for all levels of decoders until the target image segmentation result output by the upper-level decoder is obtained.

[0071] The high-dimensional spatial features are determined based on the encoder output features, high-level features, and number of channels corresponding to each level;

[0072] The target image segmentation result is determined after segmenting the target image based on the land parcel boundary features, land parcel range features, and land parcel boundary distance features.

[0073] In step 101, the upper-layer encoder is determined based on a deep learning-based land parcel segmentation network model, such as... Figure 8 As shown, Figure 8 This is one of the structural schematic diagrams of the land parcel division device provided by the present invention. Figure 8 The left side shows five encoders arranged from top to bottom. The encoder at the top is the upper encoder. The upper encoder receives the target image and then downsamples the target image. After obtaining the downsampled max pooling layer output and encoder output features, it continues to downsample the downsampled max pooling layer output according to the encoders arranged from top to bottom to obtain the corresponding max pooling layer output and encoder output features for each level.

[0074] Optionally, the target image includes at least the width, height, and number of channels of the target image. More specifically, the input remote sensing image block size is 256*256, including four channels: red, green, blue, and near-infrared.

[0075] In step 102, the encoder output features corresponding to each level are updated based on the high-dimensional spatial features to obtain the decoder connection input corresponding to each level. The high-dimensional spatial features are determined based on the encoder output features corresponding to each level, the high-level features, and the number of channels, such as... Figure 10 As shown, Figure 10 This is a schematic diagram of the spatial attention module provided by the present invention. The present invention adds four spatial attention blocks to the downsampling and upsampling connections in the classic encoder-decoder coupled segmentation model. In the spatial attention module, low-level features are updated by high-level features, which can be expressed by the following formula:

[0076]

[0077]

[0078] In equation (1), These represent low-level features and high-level features, respectively. l It is the output of the residual block of the encoding module, x h This is the output of the decoder residual block. The width and height w of the low-level features. l ,h l Typically, the width and height w of a high-level feature h ,h h 2 times, in equation (2), This represents a 1x1 convolutional layer, primarily used to adjust the number of channels in a feature. This represents a ReLU activation function layer with a 2x2 convolutional kernel, a 2x2 stride, and SamePadding, where the width and height of the low-level features are halved.

[0079] Furthermore, High-dimensional features after channel adjustment After the merger, adopt Determine g(x) l ,x h ), where F sigmoid This represents a convolutional layer with a 1x1 kernel, followed by an activation function layer with a sigmoid neural network activation function. Those skilled in the art will understand that g(x) l ,x h The output of ) is a feature layer of w×h×1. And This represents the point between low-level and high-level features. Ultimately, the low-level features will be multiplied by the high-dimensional feature g(x). l ,x h To implement the update.

[0080] In step 103, after inputting the input of the bottom-level decoder connection and the upsampling result of the bottom-level encoder to the bottom-level decoder, the output result of the bottom-level decoder is used as the upsampling result of the next-level decoder. The process is repeated for all levels of decoders until the target image segmentation result output by the next-level decoder is obtained. The input of the bottom-level decoder connection will be obtained in step 102.

[0081] Specifically, the encoder output features corresponding to the bottom layer are updated based on the high-dimensional spatial features to obtain the decoder connection input of the bottom layer. The upsampling result of the bottom layer encoder is the encoder output after upsampling of the bottom layer encoder. At this time, since the bottom layer encoder is connected to the bottom layer decoder, the input of the decoder connection input of the bottom layer and the upsampling result of the bottom layer encoder are input to the bottom layer decoder to obtain the upsampling result of the upper layer decoder output by the bottom layer decoder. In the second-to-last layer decoder, the encoder corresponding to the second-to-last layer decoder is determined, and the decoder connection input corresponding to the encoder corresponding to the second-to-last layer decoder is obtained. Then, the decoder connection input corresponding to the encoder corresponding to the second-to-last layer decoder and the upsampling result of the upper layer decoder output by the bottom layer decoder are input to the second-to-last layer decoder to determine the upsampling result of the upper layer decoder output by the second-to-last layer decoder. This process is repeated for all layers of decoders until the plot boundary features, plot range features, and plot boundary distance features output by the upper layer decoder are obtained. The target image segmentation result is determined based on the plot boundary features, plot range features, and plot boundary distance features.

[0082] This invention proposes a land parcel segmentation model based on a classical encoder-decoder architecture. This model incorporates residual blocks, a spatial attention mechanism, and a multi-task learning model that includes three tasks: boundary, extent, and distance from the center point. This achieves land parcel segmentation from meter-level resolution remote sensing imagery. Those skilled in the art will understand that a land parcel is the smallest unit for agricultural activities. Accurate land parcel identification helps improve the accuracy of remote sensing crop classification and yield estimation, and enhances the monitoring capabilities for agricultural management. This invention utilizes high-resolution remote sensing imagery to extract land parcel distribution information, supporting remote sensing agricultural monitoring, land resource surveys, and the verification of agricultural subsidies. The deep learning technology employed in this invention can identify ground features that are discernible to human vision, thus effectively distinguishing land parcel distribution on remote sensing images based on texture, color, and other information.

[0083] This invention provides a method, apparatus, device, and medium for land parcel segmentation. It updates the encoder output features corresponding to each level based on high-dimensional spatial features to obtain the decoder connection input corresponding to each level. The input to the bottom-level decoder connection and the upsampling result of the bottom-level encoder are fed into the bottom-level decoder. The output result of the bottom-level decoder is used as the upsampling result of the next-level decoder. This process is repeated across all levels of decoders until the target image segmentation result output by the next-level decoder is obtained. This invention effectively solves the problems of gradient explosion and vanishing gradients. It simultaneously performs three tasks: determining the range, boundary, and distance to the parcel boundary of the planting area. A multi-task learning convolutional neural network is constructed. By using the parameter of distance to the parcel boundary, the conversion from instance segmentation to semantic segmentation is achieved, improving the accuracy of land parcel segmentation.

[0084] Figure 2 This is a flowchart illustrating the process of obtaining the max pooling layer output and encoder output features corresponding to each level, as provided by the present invention. The step of traversing all encoder levels until the max pooling layer output and encoder output features corresponding to each level are obtained includes:

[0085] Based on the encoder output of the next level, downsample the output of the current level's max pooling layer to obtain the output of the next level's max pooling layer and the encoder output features of the next level;

[0086] Iterate through all encoder levels until you obtain the max pooling layer output and encoder output features corresponding to each level.

[0087] In step 1011, as Figure 8 As shown, Figure 8 Five encoders are shown on the left side of the image. The downsampling direction is from top to bottom. In each downsampling process, the maximum pooling layer output of the current level is used as the downsampling target. The maximum pooling layer output of the next level and the encoder output feature of the next level are output. The encoder output feature of the next level is used to obtain the corresponding decoder connection input in step 102, while the maximum pooling layer output of the next level is used for the next downsampling.

[0088] In step 1012, the present invention will combine Figure 9 The residual blocks of the encoder are used to traverse all levels of the encoder until the corresponding max-pooling layer output and encoder output features for each level are obtained, which further describes the process. Figure 9 This is a schematic diagram of the residual block structure provided by the present invention. To improve recognition performance, the present invention makes some modifications to the structure of the segmentation model based on the classic encoder-decoder architecture coupling. In the land parcel segmentation model, a residual block replaces the simple convolution operation, such as... Figure 9As shown, the residual block can solve the problems of gradient explosion and vanishing gradients, which occur as the depth of the neural network increases. The residual block does not directly parse the target output, but attempts to solve the difference between the output and the input, which is called the residual, and can be represented by equation (3).

[0089] y l =h(x l )+F(x l ,θ l (3)

[0090] x l It is the input of the entire residual block, y l It is the target output of the residual block, F(x) l ,θ l ) represents the residual. In equation (3), This represents a 1x1 convolutional layer, which can increase or decrease the number of channels in the feature layer, making it consistent with F(x). l ,θ l The output of F(x) is consistent with that of 0. l ,θ l The algorithm consists of two convolutional layers, followed by a batch normalization layer and a dropout layer. Each convolutional layer has a 3x3 kernel size, padded using same padding to ensure the width and height of the output features match the input. The number of kernels matches the number of channels in the input feature layer, while the dropout layer randomly assigns a value of 0 to each cell, which has been shown to effectively prevent overfitting.

[0091] In the land parcel partitioning model, such as Figure 8 As shown, there are a total of 9 residual blocks, 5 of which are used for encoding and 4 for decoding. The inputs and outputs of each module are shown in the table below:

[0092]

[0093]

[0094] Table 1

[0095] The encoder's residual block input is either the original image or the output of the previous residual block plus a max-pooling layer. The max-pooling layer aims to output the maximum value of a 2x2 region, so when the feature layer passes through the max-pooling layer, its width and height will be halved, but the number of channels will remain unchanged, and the input size of the residual block will change accordingly.

[0096] Figure 3This is a schematic diagram of the process for obtaining the decoder connection input corresponding to each level provided by the present invention. The step of updating the encoder output features corresponding to each level according to the high-dimensional space features to obtain the decoder connection input corresponding to each level includes:

[0097] The higher-level features are determined based on the width, height, and preset multiple of the encoder output features corresponding to each level;

[0098] The first convolutional layer is determined by processing the encoder output features and the number of feature channels based on the linear rectification function;

[0099] The second convolutional layer is determined based on the high-level features and the number of feature channels;

[0100] The first convolutional layer and the second convolutional layer are fused to obtain the third convolutional layer;

[0101] The third convolutional layer and the number of feature channels are processed according to the neural network activation function to obtain the decoder connection input corresponding to each layer.

[0102] In step 1021, higher-level features are determined based on the width, height, and a preset multiple of the encoder output features corresponding to each level. The preset multiple is 2. If the width and height of the lower-level features are w... l ,h l Then the width and height of the high-level feature are 2w. h 2h h .

[0103] In step 1022, the first convolutional layer is determined based on the encoder output features processed by the linear rectified function and the number of feature channels, in equation (1). These represent low-level features and high-level features, respectively. l It is the output of the residual block of the encoding module, x h This is the output of the decoder residual block. Based on the linear rectification function, the encoder output features and the number of feature channels can be processed... It means that, among them, This represents a linear rectified activation function layer with a 2x2 convolutional kernel, a 2x2 stride, and Same Padding, where the width and height of the low-level features are halved.

[0104] In step 1023, the second convolutional layer is determined based on the high-level features and the number of feature channels, which can be achieved through equation (2). This means that the second convolutional layer is the convolutional layer corresponding to the high-dimensional features after channel adjustment.

[0105] In step 1024, the first convolutional layer and the second convolutional layer are fused to obtain the third convolutional layer. This step can be achieved using equation (2). To express.

[0106] In step 1025, the third convolutional layer and the number of feature channels are processed according to the neural network activation function to obtain the decoder connection input corresponding to each layer, i.e., through... Let F represent this, where F sigmoid This represents a convolutional layer with a 1x1 kernel, followed by an activation function layer with a sigmoid neural network activation function. Those skilled in the art will understand that g(x) l ,x h The output of ) is a feature layer of w×h×1, while This represents the point between low-level and high-level features. Ultimately, the low-level features will be multiplied by the high-dimensional feature g(x). l ,x h To implement the update.

[0107] Optionally, the land parcel segmentation model of this invention contains four spatial attention modules. In one optional embodiment, as shown in Table 1, x l ∈R 256×256×32 x h ∈R 128×128×64 x l It is the output of the first residual block of the encoder, while x h This is the output of the 8th residual block of the decoder. Combining equations (1) and (2), and The size is 128×128×32; and The sum is then passed through a 1x1 convolutional layer and a filter to be converted to a size of 128×128×1. Next, the acquired features are upsampled to 256×256×1, and this value is finally used to update x. l The output size will be 256×256×32.

[0108] Figure 4 This is a flowchart illustrating the process of obtaining the target image segmentation result output by the upper-layer decoder provided by the present invention. The step of traversing all levels of decoders until the target image segmentation result output by the upper-layer decoder is obtained includes:

[0109] The current level's decoder input and the next level's max pooling layer output are fed into the current level's decoder to obtain the current level's max pooling layer output.

[0110] Iterate through all levels of the decoder until you obtain the fourth, fifth, and sixth convolutional layers output by the upper-level decoder;

[0111] The fourth convolutional layer determines the land parcel boundary features of the target image, the fifth convolutional layer determines the land parcel range features, and the sixth convolutional layer determines the land parcel boundary distance features.

[0112] After segmenting the target image based on the parcel boundary features, parcel range features, and parcel boundary distance features, the target image segmentation result is obtained.

[0113] In step 1031, the decoder connection input of the current level and the output of the max pooling layer of the next level are input to the decoder of the current level to obtain the max pooling layer output of the current level. The input of the decoder residual block is the result of connecting the upsampling of the previous residual block and the output of the corresponding position of the encoder. The upsampling is the inverse operation of the max pooling layer, which restores a value to a 2*2 region and fills empty values ​​with 0. As shown in Table 1, taking the 6th residual block as an example: the output size of the 5th residual block, that is, the last residual block of the encoding part, is 16×16×512. After upsampling, it will be expanded to 32×32×512. Then this feature value will be connected with the output value of the 4th residual block, that is, 32×32×256. Therefore, the input size of the 6th residual block is 32×32×768.

[0114] In step 1032, all levels of the decoder are traversed until the fourth, fifth, and sixth convolutional layers output by the upper-level decoder are obtained. Let X∈R w×l×c As input to the target image, the land parcel segmentation model will calculate the mask range T of the corresponding land parcel based on the input target image. mask ∈R w×h×1 Boundary diagram Y boundary ∈R w×h×1 And the distance Y from each pixel to the center point of the plot. dist ∈R w×h×1 Here, w is the width of the target image, h is the height of the target image, and c is the number of channels in the target image. The three outputs are probabilities between 0 and 1. In this embodiment, the input target image passes through an encoding module, which includes 5 residual blocks and a max-pooling layer. Optionally, the output of the encoding module is a 16*16*512 feature map, which will be used as the input of the decoding module. The decoding module includes 4 residual blocks and an upsampling layer, and finally outputs three plot boundaries, plot ranges, and distances to the plot center point, each 256*256*1 in size.

[0115] In step 1033, the land parcel boundary features of the target image are determined according to the fourth convolutional layer, the land parcel range features are determined according to the fifth convolutional layer, and the land parcel boundary distance features are determined according to the sixth convolutional layer. Those skilled in the art will understand that, besides the spatial attention module, distinguishing different land parcels is another key issue in land parcel identification. This invention aims to train the land parcel segmentation model through multi-task learning, including pixel-level segmentation, boundary prediction, and distance from pixels to land parcel boundaries. Since it is learned from shared representations, training the network using relevant multi-task losses can improve the accuracy of the initial task and the generalization of the model. For the land parcel segmentation model, the main goal is to force the model to learn the distance from pixels to land parcel boundaries in boundary detection and segmentation tasks. As shown in Table 1, the output size of the last residual block is 256×256×32. The output layer includes three convolutional layers with 1×1 kernels. In each convolutional layer, the size of the land parcel range, land parcel boundary, and distance to the land parcel boundary is reduced to 256×256×1.

[0116] In step 1034, after segmenting the target image based on the plot boundary features, plot range features, and plot boundary distance features, the target image segmentation result is obtained. Given the target image, the plot segmentation model outputs three features for segmenting the target image, thereby determining the target image segmentation result. This invention can simultaneously perform three tasks: determining the range, boundary, and distance to the plot boundary of the planting plot. It constructs a multi-task learning convolutional neural network and uses the distance to the plot boundary parameter to achieve the conversion from instance segmentation to semantic segmentation, improving the accuracy of plot segmentation.

[0117] Figure 5 This is a second flowchart illustrating the land parcel segmentation method provided by the present invention. Before downsampling the target image based on the upper encoder, it further includes:

[0118] Extract the land parcel boundary features from each sample image to obtain a first sample set, input the first sample set into the first segmentation model, and train a second segmentation model based on the output of the first segmentation model.

[0119] Extract the land parcel range features from each of the sample images to obtain a second sample set, and input the second sample set into the second segmentation model to train a third segmentation model based on the output of the second segmentation model;

[0120] Extract the land parcel boundary distance features from each of the sample images to obtain a third sample set, and input the third sample set into the third segmentation model to train a land parcel segmentation model based on the output of the third segmentation model;

[0121] The encoders and decoders at all levels are determined based on the land parcel segmentation model.

[0122] In step 201, the land parcel boundary features are extracted from each sample image to obtain a first sample set. The first sample set is then input into a first segmentation model to train a second segmentation model based on the output of the first segmentation model. In this embodiment, the second segmentation model is able to process the land parcel boundary features.

[0123] In step 202, the land parcel extent features are extracted from each of the sample images to obtain a second sample set. The second sample set is then input into the second segmentation model to train a third segmentation model based on the output of the second segmentation model. The third segmentation model is capable of processing land parcel boundary features and land parcel extent features.

[0124] In step 203, the land parcel boundary distance features are extracted from each of the sample images to obtain a third sample set. This third sample set is then input into the third segmentation model to train a land parcel segmentation model based on its output. This model is capable of processing land parcel boundary features, land parcel range features, and land parcel boundary distance features. The land parcel segmentation model includes encoders and decoders at all levels.

[0125] In other embodiments, the order in which land parcel boundary features, land parcel extent features, and land parcel boundary distance features are used as samples for training can be changed. Alternatively, all three features can be used as a sample set and simultaneously input into the model for training to determine the land parcel segmentation model. It is important to note that the processing capabilities of the land parcel segmentation model for these features will differ depending on the training order, resulting in variations in the output results.

[0126] Figure 6 This is the third flowchart of the land parcel segmentation method provided by the present invention. After training the land parcel segmentation model based on the output of the third segmentation model and before downsampling the target image based on the upper encoder, it also includes:

[0127] After obtaining the land parcel boundary features of the sample image, the first loss function is calculated;

[0128] After obtaining the land parcel range features of the sample image, the second loss function is calculated;

[0129] After obtaining the land parcel boundary distance features of the sample images, the third loss function is calculated;

[0130] The first loss function, the second loss function, and the third loss function are processed according to a weighted algorithm to determine the total loss function;

[0131] The land parcel segmentation model is processed according to the total loss function, and the target image is input into the processed land parcel segmentation model.

[0132] In steps 301 and 302, after acquiring the land parcel boundary features of the sample image, a first loss function is calculated; after acquiring the land parcel extent features of the sample image, a second loss function is calculated. Optionally, in the land parcel segmentation boundary detection task, the goal of the land parcel segmentation model is to identify two categories: land parcels and background. The set similarity metric function measures the overlap and similarity between two sets; it is an adjusted average of precision and recall. The corresponding loss function (Dice loss) is a commonly used loss function that enables the land parcel segmentation model to perform better in terms of precision and recall. Therefore, the Dice Loss is used as the loss for land parcel boundaries and extent. The Dice Loss can be defined as follows:

[0133]

[0134] In equation (4), A and B are two sets;

[0135] Dice loss(L Dice ) can be defined as follows:

[0136]

[0137] Where, smoothness is the smoothing factor, set to 1, i.e., the first loss function is expressed as L. Dice-seg The second loss function is expressed as L Dice-boundary .

[0138] In step 303, for the task of determining the distance from a pixel to the plot boundary, this invention uses the mean absolute error as the loss function, which is defined as:

[0139]

[0140]

[0141] In equation (6), y i This is the true value of the pixel-to-boundary distance. is the estimated distance from the pixel to the boundary, MAE is the mean absolute error of the distance to the parcel boundary, and m is the number of pixels.

[0142] In equation (7), MAE is the mean absolute error, and L MAE Let be the mean absolute error as a function of time.

[0143] Those skilled in the art will understand that the land parcel segmentation model is a deep learning model for multi-task learning, which can simultaneously predict three tasks: the extent of the land parcel, its boundaries, and the distance to the land parcel boundaries.

[0144] In step 304, the first loss function, the second loss function, and the third loss function are processed according to a weighted algorithm to determine the total loss function. The present invention will use a weighted algorithm to determine the total loss function, which will be further described in the specific embodiments described later.

[0145] In step 305, the land parcel segmentation model is processed according to the total loss function to input the target image into the processed land parcel segmentation model. The present invention improves the accuracy and recall of the land parcel segmentation model by processing the land parcel segmentation model through the total loss function.

[0146] Figure 7 This is a flowchart illustrating the process of determining the total loss function provided by the present invention. The process involves processing the first loss function, the second loss function, and the third loss function using a weighted algorithm to determine the total loss function, including:

[0147] The first weighting function is determined based on the first loss function and the first weighting value;

[0148] The second weighting function is determined based on the second loss function and the second weighting value;

[0149] The third weighting function is determined based on the third loss function and the third weighting value;

[0150] The total loss function is determined based on the first weighting function, the second weighting function, and the third weighting function;

[0151] The first weighted value is greater than the second weighted value;

[0152] The second weighted value is greater than or equal to the third weighted value.

[0153] In step 3041, the first weighting function is determined based on the first loss function and the first weighting value. Let the first loss function be L. Dice-seg If the first weighting value is λ1, then the first weighting function is λ1L. Dice-seg .

[0154] In step 3042, the second weighting function is determined based on the second loss function and the second weighting value. Let the second loss function be L. Dice-boundar If the second weighting value is λ2, then the second weighting function is λ2L. Dice-boundary .

[0155] In step 3043, the third weighting function is determined based on the third loss function and the third weighting value. Let the third loss function be L. MAE-dist If the third weighting value is λ3, then the third weighting function is λ3L. MAE-dist .

[0156] In step 3044, the total loss function is determined based on the first weighting function, the second weighting function, and the third weighting function, and can be determined by the following formula:

[0157] L total =λ1L Dice-seg +λ2L Dice-boundar +λ3L MAE-dist (8)

[0158] In equation (8), L total Let λ1L be the total loss function. Dice-seg Let λ2L be the first weighting function. Dice-boundar For the second weighting function, λ3L MAE-dist This is the third weighting function.

[0159] Optionally, since the distance to the land parcel boundary is needed to distinguish different parcels, this invention sets the distance to the land parcel boundary as the highest weight. Therefore, λ3 is set to 0.5, the first weighting value λ1 for the range of the land parcel is 0.25, and the second weighting value λ2 related to the boundary of the land parcel is 0.25. Dice-seg L is the loss function for dividing the land parcel area. Dice-boundar L is the loss function for parcel boundary identification. cross-dist This is the mean absolute error loss function for the distance to the pixel boundary.

[0160] This invention can extract land parcel boundaries from remote sensing images of any scale. It reads the labeled image file format image using an open-source package, divides it into 256*256 tiles, and sets an overlap of 128 pixels to ensure that each land parcel occupies a certain proportion in the image, avoiding the problem of land parcels being too small to be easily identified. Then, each tile is detected and stitched together.

[0161] In practical applications, this invention uses the Northwest Plain as the first plot. For example, using 4m resolution remote sensing data from multispectral satellites, 680 samples were manually labeled through visual interpretation. Each sample is 256*256 pixels in size. Based on the designed plot segmentation model, the plot segmentation model of this invention outputs a map of the planting plots. The intersection-over-union ratio of this result is 0.75, and the accuracy of the binary classification model reaches 0.88. The model output can be aggregated with the pixel-oriented crop classification results to form plot-scale crop classification results for the Hetao Plain.

[0162] To test the spatial transferability of the model and facilitate crop classification research at the plot scale, the plot segmentation model was tested on a second plot. Since the second plot was more separable than the first, imagery was used for plot segmentation experiments. First, cloud-free imagery of the second plot was synthesized on a relevant platform. Near-infrared, red, blue, and green bands were selected as inputs, and plot segmentation was successfully performed on the second plot with an accuracy of 87% and an intersection-over-union ratio of 0.71. The results show that the plot segmentation model has high spatial transferability. When plots are visually separable, this model can efficiently generate plot distribution maps and has the potential for large-scale plot boundary mapping.

[0163] Figure 11 This is the second schematic diagram of the land parcel segmentation device provided by the present invention. The present invention discloses a land parcel segmentation device, including an acquisition unit 1: used to acquire the downsampled max pooling layer output and encoder output features based on the target image downsampled by the upper encoder, traverse all levels of encoders until the max pooling layer output and encoder output features corresponding to each level are acquired. The working principle of the acquisition unit 1 can be referred to the aforementioned step 101, and will not be repeated here.

[0164] The land parcel segmentation device also includes an update unit 2: used to update the encoder output features corresponding to each level according to the high-dimensional spatial features, so as to obtain the decoder connection input corresponding to each level. The working principle of the update unit 2 can be referred to the aforementioned step 102, and will not be repeated here.

[0165] The land parcel segmentation device further includes a processing unit 3: for inputting the input of the bottom layer decoder connection and the upsampling result of the bottom layer encoder to the bottom layer decoder, and using the output result of the bottom layer decoder as the upsampling result of the upper layer decoder, traversing all layers of decoders until the target image segmentation result output by the upper layer decoder is obtained. The working principle of the processing unit 3 can be referred to the aforementioned step 103, and will not be repeated here.

[0166] The high-dimensional spatial features are determined based on the encoder output features, high-level features, and number of channels corresponding to each level;

[0167] The target image segmentation result is determined after segmenting the target image based on the land parcel boundary features, land parcel range features, and land parcel boundary distance features.

[0168] This invention provides a method, apparatus, device, and medium for land parcel segmentation. It updates the encoder output features corresponding to each level based on high-dimensional spatial features to obtain the decoder connection input corresponding to each level. The input to the bottom-level decoder connection and the upsampling result of the bottom-level encoder are fed into the bottom-level decoder. The output result of the bottom-level decoder is used as the upsampling result of the next-level decoder. This process is repeated across all levels of decoders until the target image segmentation result output by the next-level decoder is obtained. This invention effectively solves the problems of gradient explosion and vanishing gradients. It simultaneously performs three tasks: determining the range, boundary, and distance to the parcel boundary of the planting area. A multi-task learning convolutional neural network is constructed. By using the parameter of distance to the parcel boundary, the conversion from instance segmentation to semantic segmentation is achieved, improving the accuracy of land parcel segmentation.

[0169] Figure 12 This is a schematic diagram of the structure of the electronic device provided by the present invention. Figure 12 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 12 As shown, the electronic device may include: a processor 110, a communication interface 120, a memory 130, and a communication bus 140, wherein the processor 110, the communication interface 120, and the memory 130 communicate with each other through the communication bus 140. The processor 110 can call logical instructions in the memory 130 to execute a land parcel segmentation method. This method includes: downsampling the target image based on the upper-level encoder, obtaining the downsampled max-pooling layer output and encoder output features; traversing all encoder levels until the max-pooling layer output and encoder output features corresponding to each level are obtained; updating the encoder output features corresponding to each level based on high-dimensional spatial features to obtain the decoder connection input corresponding to each level; inputting the lower-level decoder connection input and the upsampling result of the lower-level encoder to the lower-level decoder, using the output result of the lower-level decoder as the upsampling result of the upper-level decoder; traversing all decoders until the target image segmentation result output by the upper-level decoder is obtained; the high-dimensional spatial features are determined based on the encoder output features, high-level features, and number of channels corresponding to each level; the target image segmentation result is determined after segmenting the target image based on land parcel boundary features, land parcel range features, and land parcel boundary distance features.

[0170] Furthermore, the logical instructions in the aforementioned memory 130 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0171] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute a land parcel segmentation method provided by the above methods. The method includes: downsampling the target image based on the upper-level encoder, obtaining the downsampled max-pooling layer output and encoder output features, traversing all levels of encoders until obtaining the max-pooling layer output and encoder output features corresponding to each level; updating the encoder output features corresponding to each level based on high-dimensional spatial features to obtain the decoder connection input corresponding to each level; inputting the lower-level decoder connection input and the upsampling result of the lower-level encoder to the lower-level decoder, and using the output result of the lower-level decoder as the upsampling result of the upper-level decoder, traversing all levels of decoders until obtaining the target image segmentation result output by the upper-level decoder; the high-dimensional spatial features are determined based on the encoder output features, high-level features, and number of channels corresponding to each level; the target image segmentation result is determined after segmenting the target image based on land parcel boundary features, land parcel range features, and land parcel boundary distance features.

[0172] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the land parcel segmentation method provided by the above methods. The method includes: downsampling the target image based on the upper-level encoder, obtaining the downsampled max-pooling layer output and encoder output features, traversing all levels of encoders until obtaining the max-pooling layer output and encoder output features corresponding to each level; updating the encoder output features corresponding to each level based on high-dimensional spatial features to obtain the decoder connection input corresponding to each level; inputting the lower-level decoder connection input and the upsampling result of the lower-level encoder to the lower-level decoder, using the output result of the lower-level decoder as the upsampling result of the upper-level decoder, traversing all levels of decoders until obtaining the target image segmentation result output by the upper-level decoder; the high-dimensional spatial features are determined based on the encoder output features, high-level features, and number of channels corresponding to each level; the target image segmentation result is determined after segmenting the target image based on land parcel boundary features, land parcel range features, and land parcel boundary distance features.

[0173] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0174] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0175] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method of parcel segmentation, characterized by, The method comprises the following steps: down-sampling the target image according to the upper-layer encoder to obtain the down-sampled maximum pooling layer output and the encoder output feature, and traversing all levels of the encoder until the maximum pooling layer output and the encoder output feature corresponding to each level are obtained; updating the encoder output feature corresponding to each level according to the high-dimensional space feature to obtain the decoder connection input corresponding to each level; inputting the decoder connection input of the bottom layer and the up-sampling result of the bottom-layer encoder into the bottom-layer decoder, taking the output result of the bottom-layer decoder as the up-sampling result of the decoder of the upper layer, and traversing all levels of the decoder until the target image segmentation result output by the upper-layer decoder is obtained; the high-dimensional space feature is determined according to the encoder output feature corresponding to each level, the high-level feature and the number of channels; the target image segmentation result is determined by segmenting the target image according to the parcel boundary feature, the parcel range feature and the parcel boundary distance feature; the method comprises the following steps: determining the high-level feature according to the width, height and preset multiple of the encoder output feature corresponding to each level; processing the encoder output feature and the number of feature channels according to a linear rectifier function to determine a first convolutional layer; determining a second convolutional layer according to the high-level feature and the number of feature channels; fusing the first convolutional layer and the second convolutional layer to obtain a third convolutional layer; processing the third convolutional layer and the number of feature channels according to a neural network activation function to obtain the decoder connection input corresponding to each level; the method comprises the following steps: inputting the decoder connection input of the current level and the maximum pooling layer output of the next level into the decoder of the current level to obtain the maximum pooling layer output of the current level; traversing all levels of the decoder until the fourth convolutional layer, the fifth convolutional layer and the sixth convolutional layer output by the upper-layer decoder are obtained; determining the parcel boundary feature of the target image according to the fourth convolutional layer, determining the parcel range feature according to the fifth convolutional layer, and determining the parcel boundary distance feature according to the sixth convolutional layer; segmenting the target image according to the parcel boundary feature, the parcel range feature and the parcel boundary distance feature to obtain the target image segmentation result.

2. The land division method according to claim 1, characterized by, the method comprises the following steps: down-sampling the maximum pooling layer output of the current level according to the encoder of the next level to obtain the maximum pooling layer output of the next level and the encoder output feature of the next level; traversing all levels of the encoder until the maximum pooling layer output and the encoder output feature corresponding to each level are obtained.

3. The land division method according to claim 1, characterized by, Before down-sampling the target image according to the upper-layer encoder, the method further comprises the following steps: extracting the parcel boundary feature in each sample image to obtain a first sample set, inputting the first sample set into a first segmentation model, and training a second segmentation model according to the output result of the first segmentation model; extracting a parcel range feature in each of the sample images, obtaining a second sample set, inputting the second sample set into the second segmentation model, and training a third segmentation model according to an output result of the second segmentation model; extracting a parcel boundary distance feature in each of the sample images, obtaining a third sample set, inputting the third sample set into the third segmentation model, and training a parcel segmentation model according to an output result of the third segmentation model; the all-level encoders and the all-level decoders are determined according to the parcel segmentation model.

4. The land division method according to claim 3, characterized by, After the parcel segmentation model is trained according to the output result of the third segmentation model, before the target image is down-sampled by the upper-level encoder, the method further includes: calculating a first loss function after the parcel boundary feature of the sample image is obtained; calculating a second loss function after the parcel range feature of the sample image is obtained; calculating a third loss function after the parcel boundary distance feature of the sample image is obtained; processing the first loss function, the second loss function and the third loss function according to a weighting algorithm to determine a total loss function; processing the parcel segmentation model according to the total loss function to input the target image into the processed parcel segmentation model.

5. The land division method according to claim 4, characterized in that, Processing the first loss function, the second loss function and the third loss function according to a weighting algorithm to determine a total loss function includes: determining a first weighted function according to the first loss function and a first weighting value; determining a second weighted function according to the second loss function and a second weighting value; determining a third weighted function according to the third loss function and a third weighting value; determining a total loss function according to the first weighted function, the second weighted function and the third weighted function; the first weighting value is greater than the second weighting value; the second weighting value is greater than or equal to the third weighting value.

6. A land division device characterized by comprising: includes: an obtaining unit configured to obtain a maximum pooling layer output and an encoder output feature after the target image is down-sampled by an upper-level encoder, and traverse all-level encoders until the maximum pooling layer output and the encoder output feature corresponding to each level are obtained; an updating unit configured to update the encoder output feature corresponding to each level according to a high-dimensional space feature to obtain a decoder connection input corresponding to each level; a processing unit configured to input the decoder connection input of a bottom-level and an up-sampling result of a bottom-level encoder into a bottom-level decoder, take an output result of the bottom-level decoder as an up-sampling result of a decoder of a next level, and traverse all-level decoders until a target image segmentation result output by an upper-level decoder is obtained; the high-dimensional space feature is determined according to the encoder output feature corresponding to each level, a high-level feature and a channel number; the target image segmentation result is determined after the target image is segmented according to the parcel boundary feature, the parcel range feature and the parcel boundary distance feature; the updating of the encoder output feature corresponding to each level according to the high-dimensional space feature to obtain the decoder connection input corresponding to each level includes: determining a high-level feature according to a width, a height and a preset multiple of the encoder output feature corresponding to each level; determining a first convolutional layer according to a linear rectifier function processing an encoder output feature and a feature channel number; determining a second convolutional layer according to a high-level feature and a feature channel number; fusing the first convolutional layer and the second convolutional layer to obtain a third convolutional layer; processing the third convolutional layer and a feature channel number according to a neural network activation function to obtain a decoder connection input corresponding to each level; the traversing of all levels of the decoder until a target image segmentation result output by an upper-level decoder is obtained, comprising: inputting the decoder connection input of a current level and a maximum pooling layer output of a next level into a decoder of the current level to obtain a maximum pooling layer output of the current level; traversing all levels of the decoder until a fourth convolutional layer, a fifth convolutional layer and a sixth convolutional layer output by an upper-level decoder are obtained; determining a plot boundary feature of a target image according to the fourth convolutional layer, determining a plot range feature according to the fifth convolutional layer, and determining a plot boundary distance feature according to the sixth convolutional layer; after segmenting the target image according to the plot boundary feature, the plot range feature and the plot boundary distance feature, obtaining a target image segmentation result.

7. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the plot segmentation method of any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the plot segmentation method of any one of claims 1 to 5.

Images