CN115937704A - Remote sensing image road segmentation method based on topology perception neural network - Google Patents

Abstract

The invention belongs to the technical field of remote sensing image semantic segmentation methods, in particular to a remote sensing image road segmentation method based on topology perception neural network. It includes the following steps: 1) divide the high-resolution remote sensing data set into training set and test set, and preprocess the remote sensing images in the training set and test set respectively; 2) build the image semantic segmentation network model; 3) input the training set The semantic segmentation network first initializes the semantic segmentation network, then updates the parameters in the model, and optimizes the loss function until convergence; 4) Input the test set data into the trained generator module to obtain high-precision semantic segmentation results. The invention solves the problem of insufficient feature expression ability, inability to accurately perceive the topological structure of the road, limited context information to a single training sample, large loss of edge and detail information in the process of semantic segmentation, and large loss of edge and detail information in the existing remote sensing image road automatic extraction technology The large sample size required and so on.

Description

Road segmentation method for remote sensing images based on topology-aware neural network

Technical Field

The invention belongs to the technical field of remote sensing image semantic segmentation methods, and specifically is a remote sensing image road segmentation method based on a topology-aware neural network.

Background Art

With the continuous launch of high-resolution remote sensing satellites, the increasing application of drone technology, and the continuous improvement of the spatial resolution of remote sensing images, earth observation technology has entered the era of high-resolution remote sensing big data. In high-resolution remote sensing big data, road network information not only directly reflects the economic and social development level of a country or region, but also has important application value in traffic management, urban planning, automatic navigation and other fields.

Traditional high-resolution remote sensing image road segmentation mainly relies on manual annotation, which is not only time-consuming and laborious, but also highly subjective. In order to improve detection efficiency, various object-based semantic segmentation methods have been continuously proposed and have achieved certain application results. However, these technologies rely heavily on low-level semantic features designed manually based on texture, geometry, and edges, and cannot be well applied to complex scenes such as high-resolution remote sensing images. Therefore, they have poor generalization performance and are easily affected by various noise interferences.

The fully convolutional neural network based on deep learning can automatically obtain high-level semantic features from remote sensing images through a multi-layer network structure and nonlinear transformation. In addition, the end-to-end, pixel-to-pixel implementation of the fully convolutional neural network can provide pixel-level road recognition and positioning, and is currently the most promising method for extracting roads from remote sensing images. However, the fully convolutional neural network currently used for road extraction from remote sensing images mainly focuses on pixel classification accuracy, cannot accurately perceive the topological structure of the road, and the edge and detail information is lost in the process of semantic segmentation, and is easily affected by occlusion or shadows, which seriously affects the accuracy and integrity of the results. In addition, the fully convolutional model often relies on a large number of high-precision training samples or pre-trained models trained from a large number of samples of daily scenes. However, pixel-level annotation of remote sensing images is often time-consuming and laborious. At the same time, due to the large difference between daily scene images and remote sensing images, the segmentation accuracy of the pre-trained model trained from daily scene images is often unsatisfactory. In addition, the huge number of parameters in these complex models places high demands on storage and computing devices, and training and applying models are also very time-consuming. These defects greatly limit the practical application of these methods.

Summary of the invention

The present invention aims to comprehensively solve the problems that the existing high-resolution remote sensing image road segmentation method cannot accurately perceive the topological structure, the edge and detail information are lost greatly during the semantic segmentation process, it is easily disturbed by occlusion and shadow, the model is inefficient and difficult to train. A remote sensing image road segmentation method based on a topology-aware neural network is provided.

The present invention adopts the following technical scheme: a high-resolution remote sensing image semantic segmentation network, including a downsampling path, a spatial context module and an upsampling path, the downsampling path includes a deformable convolution layer and 5 continuous downsampling units, the input data first passes through the deformable convolution layer to obtain a semantic feature map, then passes through 5 continuous downsampling units for feature extraction and downsampling, and finally outputs a feature map; the upsampling path is composed of 5 continuous upsampling units, an A aggregation operation, a convolution layer and a Softmax layer in sequence, and finally generates a segmentation prediction map; the 5 downsampling units and the 5 upsampling units correspond one to one, and are connected by a jump connection adjusted by the attention of a channel attention module; the spatial context module performs segmentation and fusion on the feature map output by the downsampling path and outputs it to the upsampling path.

In some embodiments, the deformable convolution layer is a deformable convolution with a convolution kernel size of 3×3 and a stride of 1, which adds a learnable offset variable and weight coefficient to the position of each sampling point in the 3×3 convolution kernel.

In some embodiments, the downsampling unit includes a primary aggregation module, an aggregation operation, and a down-conversion module, wherein the primary aggregation module is responsible for extracting features, and then the aggregation operation aggregates the input and output of the primary aggregation module, and then transmits the resulting feature map to the down-conversion module and the channel attention module respectively, wherein the output of the channel attention module is multiplied by the input by corresponding elements to obtain the jump connection features adjusted by the channel attention and input to the corresponding part of the upsampling path.

In some embodiments, the upsampling unit consists of a group of up-conversion modules, B aggregation operations and one-time aggregation modules, wherein the up-conversion module is responsible for restoring the spatial resolution of the feature map through upsampling, the B aggregation operation is responsible for aggregating the jump connection feature map adjusted by the channel attention module and the feature map obtained by the up-conversion module, and the one-time aggregation module is responsible for extracting features from the results of the B aggregation operation.

输入至一次聚合模块后，首先经过

个卷积模块，得到

个新特征图

，其中前两个卷积模块为可变形卷积模块；然后将所得

个特征图

进行通道堆叠操作，得到通道堆叠后的特征图

；所述卷积模块，由批归一化层、ReLU激活函数层、3

3卷积层和随机失活层依次组成；所述可变形卷积模块，由批归一化层、ReLU激活函数层、3

3可变形卷积层和随机失活层依次组成。In some embodiments, the structure of the primary aggregation module is as follows:

After being input into the primary aggregation module, it first passes through

convolution modules, we get

New feature map

, where the first two convolution modules are deformable convolution modules; then the obtained

Feature Map

Perform channel stacking operation to obtain the feature map after channel stacking

; The convolution module consists of a batch normalization layer, a ReLU activation function layer, 3

3 convolutional layers and random dropout layers in sequence; the deformable convolution module consists of a batch normalization layer, a ReLU activation function layer, 3

3 Deformable convolutional layers and random dropout layers are composed in sequence.

的输入特征分别经过全局最大池化和全局平均池化得到两个

的特征图；接着将它们分别送入一个共享多层感知器:第一层神经元个数为

，r为减少率，r=16，激活函数为Relu，第二层神经元个数为

；输出的两个特征进行相加，再经过Sigmoid激活操作，生成最终的通道注意力；通道注意力和输入特征做对应元素乘法操作。In some embodiments, the structure of the channel attention module is:

The input features of are respectively subjected to global maximum pooling and global average pooling to obtain two

feature map; then they are sent to a shared multilayer perceptron: the number of neurons in the first layer is

, r is the reduction rate, r=16, the activation function is Relu, and the number of neurons in the second layer is

; The two output features are added, and then the Sigmoid activation operation is performed to generate the final channel attention; the channel attention and the input features perform corresponding element multiplication operations.

In some embodiments, the spatial context module includes two paths: "height direction" and "width direction";

，其中C代表通道数目，H代表特征图高度，W代表特征图宽度，在垂直方向被分为H个切片：

，对每个切片

，

，使用大小为

卷积操作进行线性投影得到

，然后将

输入第一个线性层

，得到注意力

,再对注意力

依次在特征图尺寸和通道维度分别使用Softmax和L1正则化，随后将正则化后的注意力输入第二个线性层

，得到新的切片

，最后将所有新切片沿“高度方向”聚合为大小为

的特征图；The "height direction" path comes from the feature map of the downsampling path, and its size is

, where C represents the number of channels, H represents the feature map height, and W represents the feature map width, and is divided into H slices in the vertical direction:

, for each slice

,

, using a size of

The convolution operation is linearly projected to obtain

, then

Input the first linear layer

, get attention

, and then pay attention

Softmax and L1 regularization are used in the feature map size and channel dimension respectively, and then the regularized attention is input into the second linear layer

, get a new slice

, and finally aggregate all new slices along the "height direction" into a size of

The feature map of

In the “width direction” path, the feature map from the downsampling path is divided into W slices along the “width” direction, and the slices are updated and aggregated into feature maps in the same way;

的特征图通过加法运算进行融合。From the "height direction" and "width direction" paths, the size is

The feature maps are fused by addition operation.

A remote sensing image road segmentation method based on topology-aware neural network includes the following steps:

S100: Divide the high-resolution remote sensing dataset into a training set and a test set, and preprocess the images in the training set and the test set respectively;

S200: Building a high-resolution remote sensing image semantic segmentation network;

S300: input the preprocessed training set images in S100 into the high-resolution remote sensing image semantic segmentation network in S200 for training. First, the high-resolution remote sensing image segmentation network is initialized using the He Uniform method, and then the parameters in the model are updated to optimize the loss function until convergence.

S400: Input the preprocessed test set remote sensing images into the segmentation network trained in S200, and output the semantic segmentation results of the high-resolution remote sensing images.

In step S100, the preprocessing includes manual image annotation, image cropping and data enhancement;

The manual image annotation specifically includes: manually annotating the roads in the high-resolution image at the pixel level in ArcGIS software to obtain a labeled surface crack image;

The image cropping is specifically as follows: randomly cropping the labeled high-resolution remote sensing image into sub-images of 512 pixels×512 pixels;

The data enhancement includes: performing random scale transformation on sub-images, random angle image rotation, and vertical and horizontal image flipping to obtain a high-resolution remote sensing image.

In step S300, the loss function adopts Dice Loss, which is specifically:

代表X和Y的交集，

和

代表X和Y中元素的个数。Where X represents the set of predicted values corresponding to all pixels in the prediction image, and Y represents the set of values corresponding to all pixels in the label image.

represents the intersection of X and Y,

and

Represents the number of elements in X and Y.

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

1) The present invention adopts a spatial context module, which converts the traditional layer-by-layer convolution connection form into a slice-by-slice convolution form in the feature map, so that information can be transmitted between pixel rows and columns in the map. Therefore, even when the road is blocked or other appearance clues are poor, the spatial context module is still suitable for detecting crack targets with long-distance continuous shapes due to its powerful topological morphological perception ability.

2) The spatial context module in the present invention uses two linear layers with memory function to replace the Key vector and Value vector in the traditional self-attention mechanism, which can not only obtain global context information from all samples in the entire training set, but also reduce the computational complexity.

3) The present invention uses wavelet transform and inverse wavelet transform in the down-conversion module and up-conversion module in the network to implement down-sampling and up-sampling to further avoid the loss of details and edge information in the model, thereby improving the model performance.

4) In the downsampling and upsampling paths of the present invention, a one-time aggregation module is used to extract road features. This design allows the model to reuse features more efficiently and reduces redundant calculations, making the model easier to train, eliminating the need for pre-training models and reducing the number of samples required for model training. At the same time, the one-time aggregation module is combined with deformable convolution, which is beneficial for obtaining accurate road geometry features and enhancing topology perception capabilities.

5) The jump connection of the present invention adopts a channel attention module, which is conducive to selecting appropriate channel information for fusion and avoiding interference, so as to better improve the semantic segmentation accuracy of the model.

Through the above beneficial effects, the present invention comprehensively solves the problems existing in the existing high-resolution remote sensing image road extraction methods, such as the inability to accurately perceive the topological structure of the road, the acquisition of global context information is limited to a single sample, the loss of edge and detail information in the semantic segmentation process is large, and the number of samples required for training is large.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic diagram of the overall structure of a high-resolution remote sensing image semantic segmentation network constructed in the method of the present invention;

FIG2 is a schematic diagram of the structural composition of a primary aggregation module combined with deformable convolution in a high-resolution remote sensing image semantic segmentation network constructed in the method of the present invention;

FIG3 is a schematic diagram of the composition structure of the channel attention module in the high-resolution remote sensing image semantic segmentation network constructed in the method of the present invention;

FIG4 is a schematic diagram of the composition structure of the spatial context module in the high-resolution remote sensing image semantic segmentation network constructed in the method of the present invention.

DETAILED DESCRIPTION

The specific implementation modes of the present invention are described below in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention.

The present invention relates to a surface crack detection method based on a multi-scale topological perception neural network, comprising the following steps:

Step 1: Divide the high-resolution remote sensing dataset into a training set and a test set, and preprocess the images in the training set and the test set respectively.

The preprocessing includes manual image annotation, image cropping and data enhancement; the manual image annotation specifically includes: manually performing pixel-level semantic annotation on roads in high-resolution images in ArcGIS software to obtain labeled surface crack images.

The image cropping is specifically as follows: randomly cropping the labeled high-resolution remote sensing image into sub-images of 512 pixels×512 pixels.

The data enhancement includes: performing random scale transformation on sub-images, random angle image rotation, and vertical and horizontal image flipping to obtain a high-resolution remote sensing image.

Step 2: Build a high-resolution remote sensing image semantic segmentation network. As shown in Figure 1, the high-resolution remote sensing image semantic segmentation network consists of a downsampling path, a spatial context module and an upsampling path in sequence. The downsampling path and the upsampling path are connected by a jump connection adjusted by channel attention.

的取值为48的语义特征图，随后通过5个连续的下采样单元进行特征提取和下采样。Downsampling path: Starting from the input data, first pass through a deformable convolution layer with a convolution kernel size of 3×3 and a stride of 1 to get the number of channels

The semantic feature map with a value of 48 is then extracted and downsampled through 5 consecutive downsampling units.

Deformable convolution adds a learnable offset variable and weight coefficient to the position of each sampling point in the conventional convolution kernel; through these variables, the convolution kernel can sample arbitrarily near the current position and eliminate the interference of useless context information, and is no longer limited to the regular grid of conventional convolution; deformable convolution is conducive to the accurate extraction of complex geometric features of the road.

The downsampling unit is composed of a single aggregation module, an aggregation operation, and a down-conversion module in sequence. The single aggregation module is responsible for extracting features. The aggregation operation then aggregates the input and output of the single aggregation module, and then transmits the resulting feature map to the down-conversion module and the channel attention module respectively. The output of the channel attention module (channel attention) is multiplied by the input by corresponding elements to obtain the jump connection features adjusted by the channel attention and input into the corresponding part of the upsampling path.

输入至一次聚合模块后，首先经过n个卷积模块，得到n个新特征图

(其中前两个卷积模块为可变形卷积模块)；然后将所得

个特征图

进行通道堆叠操作，得到通道堆叠后的特征图

；一次性聚合模块将特征图一次性聚合，即重复利用了网络所提取的特征又避免了冗余计算，因而可以有效降低模型训练的难度并减少所需样本数量。As shown in Figure 2, the structure of the primary aggregation module is as follows:

After being input into the primary aggregation module, it first passes through n convolution modules to obtain n new feature maps

(The first two convolution modules are deformable convolution modules); then

Feature Map

Perform channel stacking operation to obtain the feature map after channel stacking

The one-time aggregation module aggregates the feature maps at one time, which reuses the features extracted by the network and avoids redundant calculations, thus effectively reducing the difficulty of model training and the number of samples required.

的取值分别为4、5、7、10、12，对应输出的特征图通道数目m的取值分别为112、192、304、464、656；每个上采样单元中一次聚合模块包含的常规卷积和可变形卷积模块总个数

的取值分别为12、10、7、5、4，输出到下一个上转换模块的特征图通道数

的取值为192、160、112、80、64。The total number of regular convolution and deformable convolution modules contained in each primary aggregation module in the downsampling path

The values of are 4, 5, 7, 10, and 12, respectively, and the corresponding values of the number of feature map channels m of the output are 112, 192, 304, 464, and 656, respectively; the total number of conventional convolution and deformable convolution modules contained in the aggregation module in each upsampling unit is

The values of are 12, 10, 7, 5, and 4, respectively, and the number of feature map channels output to the next up-conversion module

The values are 192, 160, 112, 80, and 64.

3卷积层或可变形卷积层(通道数m=16)和随机失活层（失活概率p=0.2）依次组成。The convolution module or deformable convolution module in the primary aggregation module consists of a batch normalization layer, a ReLU activation function layer, 3

It consists of 3 convolutional layers or deformable convolutional layers (channel number m=16) and random inactivation layers (inactivation probability p=0.2).

的输入特征分别经过全局最大池化和全局平均池化得到两个

的特征图，接着将它们分别送入一个共享多层感知器:第一层神经元个数为

（r为减少率，r=16），激活函数为Relu，第二层神经元个数为

；输出的两个特征进行相加，再经过Sigmoid激活操作，生成最终的通道注意力；通道注意力和输入特征图做对应元素乘法操作。As shown in Figure 3, the channel attention module structure is as follows:

The input features of are respectively subjected to global maximum pooling and global average pooling to obtain two

The feature maps are then sent to a shared multilayer perceptron: the number of neurons in the first layer is

(r is the reduction rate, r=16), the activation function is Relu, and the number of neurons in the second layer is

; The two output features are added, and then the Sigmoid activation operation is performed to generate the final channel attention; the channel attention and the input feature map perform corresponding element multiplication operations.

、

、

和

，其中下标中字母的前后顺序分别代表水平和垂直方向，G代表高频信息、而D代表低频信息，例如GG代表：水平和垂直方向均为高频信息下标；将

子图像输出，而其余子图像传输至上采样路径对应的模块。The down-conversion module uses Daubechies wavelet as the wavelet basis. After a wavelet transform, each channel of the image is decomposed into four sub-images with half the original height and width:

,

,

and

, where the order of the letters in the subscripts represents the horizontal and vertical directions respectively, G represents high-frequency information, and D represents low-frequency information. For example, GG represents: both the horizontal and vertical directions are high-frequency information subscripts;

The sub-image is output, while the remaining sub-images are transmitted to the modules corresponding to the upsampling path.

，其中C代表通道数目，H代表特征图高度，W代表特征图宽度）在垂直方向被分为H个切片：

，首先对每个切片

（

）使用大小为

卷积操作进行线性投影得到

，然后将

输入第一个线性层

（

,其中S=64，C=656），得到注意力

,再对注意力

依次在特征图尺寸和通道维度分别使用Softmax和L1正则化，随后将正则化后的注意力输入第二个线性层

（

,其中C=656,S=64），得到新的切片

，最后将所有新切片沿“高度方向”聚合为大小为

的特征图；在“宽度方向”路径，来自下采样路径的特征图沿“宽度”方向被分为W个切片，切片按照前两个阶段的方式更新并聚合；来自“高度方向”和“宽度方向”两条路径、大小均为

的特征图通过加法运算进行融合；空间上下文模块将传统的卷积层接层(Layer-by-layer)的连接形式的转为特征图中片连片（Slice-by-slice）线性复杂度自注意力的形式，使得图中像素行和列之间能够传递信息，同时采用具有记忆功能的线性层替代传统自注意力中的Key矢量和Value矢量，因此即使是遥感图像中的道路被遮挡或被阴影覆盖，空间上下文模块仍然适用于检测长距离连续形状的道路目标，并且能够提取分布于整个训练集的全局上下文信息，同时降低计算复杂度。As shown in FIG4 , the spatial context module consists of two paths, “height direction” and “width direction”; in the “height direction” path, the feature map from the downsampling path (assuming the size is

, where C represents the number of channels, H represents the feature map height, and W represents the feature map width) is divided into H slices in the vertical direction:

, first for each slice

（

) Use size

The convolution operation is linearly projected to obtain

, then

Input the first linear layer

（

, where S=64, C=656), get attention

, and then pay attention

Softmax and L1 regularization are used in the feature map size and channel dimension respectively, and then the regularized attention is input into the second linear layer

（

, where C=656, S=64), get the new slice

, and finally aggregate all new slices along the "height direction" into a size of

In the “width direction” path, the feature map from the downsampling path is divided into W slices along the “width direction”, and the slices are updated and aggregated in the same way as in the first two stages; the feature maps from the “height direction” and “width direction” paths are both

The feature maps are fused by addition operations; the spatial context module converts the traditional convolutional layer-by-layer connection form into the slice-by-slice linear complexity self-attention form in the feature map, so that information can be transmitted between pixel rows and columns in the map. At the same time, a linear layer with memory function is used to replace the Key vector and Value vector in the traditional self-attention. Therefore, even if the road in the remote sensing image is blocked or covered by shadows, the spatial context module is still suitable for detecting road targets with long-distance continuous shapes, and can extract global context information distributed in the entire training set while reducing computational complexity.

The upsampling path is composed of 5 consecutive upsampling units, a convolution layer with a convolution kernel size of 1×1 and a step size of 1, and a Softmax layer, and finally generates a segmentation prediction map.

The upsampling path is composed of 5 consecutive upsampling units, an aggregation operation (denoted as A aggregation operation), a convolution layer with a convolution kernel size of 1×1 and a step size of 1, and a Softmax layer, and finally generates a segmentation prediction map.

The upsampling unit consists of a group of up-conversion modules, aggregation operations (denoted as B aggregation operations) and a one-time aggregation module, wherein the up-conversion module is responsible for restoring the spatial resolution of the feature map through upsampling, the B aggregation operation is responsible for aggregating the skip connection feature map adjusted by the channel attention module and the feature map obtained by the up-conversion module, and the one-time aggregation module is responsible for extracting features from the results of the B aggregation operation.

、

和

）在通过1×1卷积通道数提升后与空间上下文模块或上一个一次聚合模块输出的新特征组合，进行Daubechies小波为小波基的小波逆变换，从而实现上采样。The up-conversion module converts the high-frequency information feature map (

,

and

) After the number of channels is increased through 1×1 convolution, it is combined with the new features output by the spatial context module or the previous one-time aggregation module, and an inverse wavelet transform with Daubechies wavelet as the wavelet basis is performed to achieve upsampling.

The A aggregation operation in the upsampling path aggregates the result of the B aggregation operation in the last upsampling unit with the output of the primary aggregation module again.

The convolution layer with a convolution kernel size of 1×1 and a stride of 1 converts the number of channels to 2.

Step 3: Input the preprocessed training set images in step 1 into the high-resolution remote sensing image semantic segmentation network in step 2 for training. First, use the He Uniform method to initialize the high-resolution remote sensing image segmentation network, then update the parameters in the model, and optimize the loss function until convergence.

代表X和Y的交集，

和

代表X和Y中元素的个数。Where X represents the set of predicted values corresponding to all pixels in the prediction image, and Y represents the set of values corresponding to all pixels in the label image.

represents the intersection of X and Y,

and

Represents the number of elements in X and Y.

Step 4: Input the preprocessed test set remote sensing images into the segmentation network trained in step 3, and output the semantic segmentation results of the high-resolution remote sensing images.

What needs to be explained about the specific structure of the present invention is that the connection relationship between the various component modules adopted in the present invention is definite and feasible. Except for the special instructions in the embodiments, the specific connection relationship can bring about the corresponding technical effects and solve the technical problems raised by the present invention without relying on the execution of the corresponding software program. Except for the specific instructions, the models and connection methods of the components, modules and specific components appearing in the present invention belong to the existing technologies such as published patents, published journal articles or common knowledge that can be obtained by technical personnel in this field before the application date, and there is no need to elaborate, so that the technical solution provided in this case is clear, complete and feasible, and the corresponding physical products can be reproduced or obtained according to the technical means.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims (10)

1. A high-resolution remote sensing image semantic segmentation network is characterized in that: comprises a down-sampling path, a spatial context module and an up-sampling path,

the down-sampling path comprises a deformable convolution layer and 5 continuous down-sampling units, input data firstly passes through the deformable convolution layer to obtain a semantic feature map, then feature extraction and down-sampling are carried out through the 5 continuous down-sampling units, and finally a feature map is output;

the up-sampling path sequentially consists of 5 continuous up-sampling units, an aggregation operation A, a convolution layer and a Softmax layer, and finally a segmentation prediction graph is generated;

the 5 down sampling units and the 5 up sampling units are in one-to-one correspondence and are connected by adopting a jump connection which is adjusted by the attention of the channel attention module;

and the spatial context module is used for segmenting and fusing the feature map output by the down-sampling path and outputting the feature map to the up-sampling path.

2. The high resolution remote sensing image semantic segmentation network according to claim 1, characterized in that: the deformable convolution layer is a deformable convolution with a convolution kernel size of 3 x 3 and a step size of 1, and this operation adds a learnable offset variable and weight coefficient to the position of each sample point in the 3 x 3 convolution kernel.

3. The high resolution remote sensing image semantic segmentation network according to claim 2, characterized in that: the down-sampling unit comprises a primary aggregation module, an aggregation operation module and a down-conversion module, wherein the primary aggregation module is responsible for extracting features, then the aggregation operation module aggregates the input and the output of the primary aggregation module, and then the result feature graph is respectively transmitted to the down-conversion module and the channel attention module, wherein the output and the input of the channel attention module are multiplied by corresponding elements to obtain jump connection features subjected to channel attention adjustment and input the jump connection features to the corresponding part of the upper sampling path.

4. The high resolution remote sensing image semantic segmentation network according to claim 1, characterized in that: the up-sampling unit comprises a group of up-conversion modules, a B aggregation operation and a primary aggregation module, wherein the up-conversion modules are responsible for recovering the spatial resolution of the feature map through up-sampling, the B aggregation operation is responsible for aggregating the jump connection feature map adjusted by the channel attention module and the feature map obtained by the up-conversion modules, and the primary aggregation module is responsible for extracting features from the result of the B aggregation operation.

5. The high resolution remote sensing image semantic segmentation network according to claim 3 or 4, characterized in that: the structure of the primary polymerization module is as follows: characteristic diagram

After being input into a primary polymerization module, the data is firstly processed through->

A convolution module to get >>

Individual new characteristic map>

Wherein the first two convolution modules are deformable convolution modules; then combining the result>

Characteristic diagram

Performing channel stacking operation to obtain a characteristic diagram after channel stacking>

；

The convolution module consists of a batch normalization layer, a ReLU activation function layer and 3

3, sequentially forming a convolution layer and a random inactivation layer; the deformable convolution module consists of a batch normalization layer, a ReLU activation function layer and a 3 ^ er>

3 a deformable convolution layer and a random deactivation layer.

6. The high resolution remote sensing image semantic segmentation network according to claim 5, characterized in that: the structure of the channel attention module is as follows: is of the size of

Are subject to global maximum pooling and global mean pooling, respectively, to obtain two->

A characteristic diagram of (1);

then respectively sending the data to a shared multilayer perceptron, wherein the number of first layer neurons is

R is decreasing rate, r =16, activation function is Relu, number of second layer neurons is +>

；/>

Adding the two output characteristics, and generating final channel attention through Sigmoid activation operation; and carrying out corresponding element multiplication operation on the channel attention and the input features.

7. The high resolution remote sensing image semantic segmentation network according to claim 6, characterized in that: the space context module comprises two paths of a height direction and a width direction;

wherein the "height direction" path is from a feature map of the down-sampled path having a size of

Where C represents the number of channels, H represents the feature height, W represents the feature width, divided into H slices in the vertical direction:

For each slice->

，

Is used with a size of->

Convolution operation performs linear projection to get->

Then will->

Input the first linear level->

Get attention->

Then attention is called>

Regularization using Softmax and L1, respectively, in the feature size and channel dimensions in turn, and then inputting the regularized attention into the second linear layer ≦>

Obtaining new sections>

Finally, all new sections are combined in the "height direction" to a size ^ 4>

A characteristic diagram of (1);

in the 'width direction' path, the feature map from the downsampling path is divided into W slices along the 'width' direction, and the slices are updated and aggregated into the feature map in the same way;

the two paths from the height direction and the width direction are both

The feature maps of (a) are fused by addition.

8. A remote sensing image road segmentation method based on a topology-aware neural network is characterized by comprising the following steps: comprises the following steps of (a) preparing a solution,

s100: dividing the high-resolution remote sensing data set into a training set and a testing set, and respectively preprocessing images in the training set and the testing set;

s200: building a high-resolution remote sensing image semantic segmentation network;

s300: inputting the preprocessed training set image in the S100 into a high-resolution remote sensing image semantic segmentation network in the S200 for training, initializing the high-resolution remote sensing image semantic segmentation network by using a He Uniform method, updating parameters in a model, and optimizing a loss function until convergence;

s400: and inputting the preprocessed test set remote sensing image into the trained segmentation network in S200, and outputting a semantic segmentation result of the high-resolution remote sensing image.

9. The remote sensing image road segmentation method based on the topology aware neural network as claimed in claim 8, characterized in that: in the step S100, the preprocessing includes image manual labeling, image cropping, and data enhancement;

the image manual labeling specifically comprises the following steps: manually carrying out pixel-level semantic annotation on a road in the high-resolution image in ArcGIS software to obtain a labeled surface crack image;

the image cutting specifically comprises the following steps: randomly cutting the high-resolution remote sensing image with the label into a sub-image of 512 pixels multiplied by 512 pixels;

the data enhancement comprises: and carrying out scale random transformation, random angle image rotation and image vertical and horizontal overturning on the subimages to obtain a high-resolution remote sensing image.

10. The remote sensing image road segmentation method based on the topology aware neural network as claimed in claim 8, characterized in that: in step S300, the Loss function adopts Dice Loss, which specifically includes:

wherein X represents the set of predicted values corresponding to all pixels in the prediction map, Y represents the set of values corresponding to all pixels in the label image,

represents the intersection of X and Y>

And &>

Represents the number of elements in X and Y. />

Images