A road weak information extraction method and system based on ASU-Net

Abstract

The application discloses a kind of based on ASU-Net's road weak information extraction method and system, it is related to image feature identification technical field, method includes: using encoder-decoder structure to construct road extraction network model ASU-Net, wherein, using ResNet-50 as encoder, and introducing hollow convolution module in encoder, introduce zonal convolution module in decoder;Remote sensing satellite road image is input into ASU-Net model, and the encoder is sampled to the image, while the receptive field is expanded by the parallel cascade branch of the preset hollow rate in hollow convolution module;The features learned by the encoder are up-sampled by transposed convolution, while the zonal convolution module is used to extract road linear features using four different directions of zonal convolution.The technical scheme of the present application improves the performance of road extraction in complex environment, achieves a better balance between precision and recall, and improves the accuracy of road extraction.

Description

Technical Field

[0001] This invention relates to the field of image feature recognition technology, and in particular to a road weak information extraction method and a road weak information extraction system based on ASU-Net. Background Technology

[0002] Road extraction can be viewed as a semantic segmentation task, requiring the classification of each pixel to determine whether it belongs to the road or background category. However, due to the excessive interference from the background category, road extraction from remote sensing imagery has always faced several significant challenges:

[0003] (1) In remote sensing road images, roads usually occupy only a relatively small proportion, with the remaining large part being the background. This leads to a serious imbalance between the foreground and background, posing challenges when training road extraction models. Because models generally tend to learn the background category that occupies a larger proportion, while learning the road category is relatively insufficient, this will lead to a decrease in model performance and affect accuracy.

[0004] (2) In terms of geometric shape and texture features, roads are very similar to background objects such as rivers, ditches, and farmland, so misclassification is very likely to occur

[33] . This makes it difficult for road extraction algorithms to accurately distinguish the categories of these objects, which may lead to the incorrect labeling of objects such as rivers, ditches, and farmland as roads. Moreover, this situation will not only affect the accuracy of road extraction results, but may also affect subsequent geographic information analysis and applications.

[0005] (3) Urban roads usually have complex traffic environments, and vehicles, vegetation, buildings and other objects can also obstruct the roads. Based on the above factors, the results of road extraction often contain fragmented and discontinuous roads

[34] , because the model may interrupt or misidentify the roads, making it very difficult to extract continuous roads. In addition, the complex traffic environment, such as the variety of road shapes or directions, makes it impossible to accurately capture the roads.

[0006] (4) Since small roads are weakly represented in remote sensing images, it is easy to lose information about small roads during feature extraction, resulting in the omission of small roads

[31] . Therefore, when extracting roads, special attention should be paid to the identification of small roads, and corresponding algorithms should be designed to solve this problem in order to improve the completeness of road extraction results.

[0007] In summary, road extraction from remote sensing imagery is a highly challenging task, and designing an effective method to address these difficulties is also extremely difficult. Therefore, many high-performing methods are still needed to further improve the connectivity and accuracy of road extraction in remote sensing imagery, providing more comprehensive support for the geographic information field. Summary of the Invention

[0008] To address the aforementioned problems, this invention provides a road weak information extraction method and system based on ASU-Net. By constructing a road extraction network model, ASU-Net, that integrates strip convolution and dilated convolution, the method employs strip convolution modules to capture image features, reducing interference from irrelevant regions in feature learning. Dilated convolution modules expand the receptive field, and skip connections fuse multi-scale features to better capture global contextual information while preserving lower-level features. This improves the performance of road extraction in complex environments, achieving a better balance between precision and recall, and enhancing the accuracy of road extraction.

[0009] To achieve the above objectives, this invention provides a method for extracting weak road information based on ASU-Net, comprising:

[0010] A road extraction network model ASU-Net is constructed using an encoder-decoder structure, wherein ResNet-50 is used as the encoder, and a dilated convolution module is introduced into the encoder, and a strip convolution module is introduced into the decoder.

[0011] The remote sensing satellite road image is input into the ASU-Net model. The encoder performs downsampling on the image and expands the receptive field through parallel cascaded branches with a preset hole rate in the dilated convolution module.

[0012] The features learned by the encoder are upsampled by transposed convolution, and the linear features of the road are extracted by using strip convolution in four different directions through the strip convolution module.

[0013] In the above technical solution, preferably, the encoder and the decoder perform multi-scale feature information fusion through a skip connection.

[0014] In the above technical solution, preferably, the ResNet-50 includes a preset number of convolutional layers and uses a bottleneck module as a residual module, and the ResNet-50 is pre-trained on ImageNet.

[0015] In the above technical solution, preferably, the output feature of the residual module is the sum of the residual mapping of the input features after passing through the convolutional layer and the input features. The residual module uses a 1×1 convolutional layer for dimensionality reduction, a 3×3 convolutional layer, and a 1×1 convolutional layer for dimensionality restoration.

[0016] In the above technical solution, preferably, the dilated convolution module adopts four parallel branches, wherein the first three branches fuse feature maps of different receptive field sizes in a cascade manner, and the last branch adds image-level features through global average pooling, and then reduces the number of channels to the number of input channels through convolution operation.

[0017] In the above technical solution, preferably, the encoder reduces the image size to a preset ratio through five downsampling operations, and increases the receptive field accordingly based on the porosity of the cascaded branches in the dilated convolution module.

[0018] In the above technical solution, preferably, the strip convolution module includes four strip convolutions: horizontal, vertical, left diagonal, and right diagonal. The input tensor is fed into the four parallel strip convolution branches after a 1×1 convolution operation. Then, the four output tensors are connected and a 1×1 convolution operation is performed to capture long-range context information in the feature map from four different directions, so that each pixel in the output feature map can be associated with pixels in the four directions of the input feature map.

[0019] In the above technical solution, preferably, the calculation formula for adjusting the receptive field of pixels in the output feature map by adjusting the dilatation rate of the dilated convolution module is as follows:

[0020] S=(r-1)×(k-1)+k

[0021] Where S is the receptive field, r is the hole ratio, and k is the kernel size.

[0022] In the above technical solution, preferably, the formula for calculating the receptive field size of the cascaded dilated convolutions in the dilated convolution module is:

[0023] S' = S1 + S2 - 1

[0024] Where S' is the receptive field size after cascading, and S1 and S2 are the receptive field sizes of the two dilated convolutions in the cascade, respectively.

[0025] This invention also proposes a road weak information extraction system based on ASU-Net, which applies the road weak information extraction method based on ASU-Net disclosed in any of the above technical solutions, including:

[0026] The extraction model building module is used to construct a road extraction network model ASU-Net using an encoder-decoder structure, wherein ResNet-50 is used as the encoder, and a dilated convolution module is introduced into the encoder, and a strip convolution module is introduced into the decoder.

[0027] The image sampling and processing module is used to input remote sensing satellite road images into the ASU-Net model. The encoder performs downsampling operations on the images and expands the receptive field through parallel cascaded branches with a preset hole rate in the dilated convolution module.

[0028] The road feature extraction module is used to upsample the features learned by the encoder through transposed convolution, and at the same time, to extract linear road features through the strip convolution module using strip convolution in four different directions.

[0029] Compared with existing technologies, the beneficial effects of this invention are as follows: by constructing a road extraction network model ASU-Net that integrates strip convolution and dilated convolution, the strip convolution module is used to capture image features, reducing interference from irrelevant regions in feature learning. The dilated convolution module expands the receptive field, and multi-scale features are integrated through skip connections to better capture global contextual information and retain lower-level features. This improves the performance of road extraction in complex environments, achieves a better balance between precision and recall, and improves the accuracy of road extraction. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the network structure of the ASU-Net model disclosed in one embodiment of the present invention;

[0031] Figure 2 This is a schematic diagram of the residual module disclosed in one embodiment of the present invention;

[0032] Figure 3 This is a schematic diagram of the bottleneck module disclosed in one embodiment of the present invention;

[0033] Figure 4 This is a schematic diagram illustrating the principle of a dilated convolution module disclosed in one embodiment of the present invention;

[0034] Figure 5 This is a schematic diagram of the dilated convolution process disclosed in one embodiment of the present invention;

[0035] Figure 6 This is a schematic diagram illustrating the process of increasing the receptive field of dilated convolution according to an embodiment of the present invention.

[0036] Figure 7 This is a schematic diagram illustrating the principle of a strip convolution module disclosed in one embodiment of the present invention;

[0037] Figure 8 This is a schematic diagram of road images from a Zhouqu road dataset disclosed in one embodiment of the present invention;

[0038] Figure 9 This is a schematic diagram illustrating the experimental results comparing different algorithms on the Zhouqu road dataset disclosed in one embodiment of the present invention;

[0039] Figure 10 This is a schematic diagram of ablation experiment results for different models on the Zhouqu road dataset disclosed in one embodiment of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] The present invention will now be described in further detail with reference to the accompanying drawings:

[0042] like Figure 1 As shown, a road weak information extraction method based on ASU-Net provided by the present invention includes:

[0043] A road extraction network model ASU-Net is constructed using an encoder-decoder structure, in which ResNet-50 is used as the encoder, and a dilated convolution module is introduced into the encoder, and a strip convolution module is introduced into the decoder.

[0044] The remote sensing satellite road image is input into the ASU-Net model. The encoder performs downsampling operation on the image and expands the receptive field through parallel cascaded branches with a preset hole rate in the dilated convolution module.

[0045] The features learned by the encoder are upsampled by transposed convolution, and linear road features are extracted by using a strip convolution module with four different directions.

[0046] In this embodiment, addressing the shortcomings of the U-Net model in road extraction, this invention constructs an ASU-Net road extraction network model that integrates strip convolution and dilated convolution. Strip convolution modules are used to capture image features, reducing interference from irrelevant regions in feature learning. Dilated convolution modules expand the receptive field, while skip connections fuse multi-scale features to better capture global contextual information and retain lower-level features. This improves the performance of road extraction in complex environments, achieving a better balance between precision and recall, and enhancing the accuracy of road extraction.

[0047] Specifically, the ASU-Net model mainly includes a dilated convolutional module to expand the receptive field and a strip convolutional module to align with the shape of the road and extract linear features of the road. In the encoder part, because ResNet performs excellently in feature learning, this invention uses ResNet-50 as the encoder.

[0048] In the decoder section, this invention upsamples the learned features to the size of the input image through transposed convolution, and extracts the linear features of the road through a strip convolution module.

[0049] Dilated convolution, also known as dilated convolution, is a convolutional operation in convolutional neural networks. It was initially proposed to solve image segmentation problems. Common image segmentation algorithms typically use pooling layers to increase the receptive field, but this also reduces the size of the feature maps. Upsampling is usually used to restore the image size, but this process of shrinking and then enlarging the feature maps results in a loss of detail. In contrast, dilated convolution can increase the model's receptive field without reducing image resolution. Dilated convolution introduces one or more gaps into the convolution kernel to increase the receptive field, allowing the network to capture a wider range of contextual information. Moreover, compared to ordinary convolution operations, dilated convolution maintains the same number of parameters and does not increase the model's parameter count.

[0050] In the above embodiments, preferably, the encoder and decoder perform multi-scale feature information fusion through skip connections.

[0051] If the decoder only uses the final output feature map of the encoder, it can easily miss pixels and miss small-sized paths. Therefore, in accordance with the U-Net model, skip connections are added to enable low-level information sharing between the encoder and decoder, which preserves low-level features and achieves multi-scale feature fusion.

[0052] By using the first four layers of the encoder feature map and the output of the dilated convolution module, the decoder can extract multi-scale contextual information from high-level feature maps and recover edge details from low-level feature maps, thus refining the decoder's output.

[0053] like Figure 2 and Figure 3 As shown, in the above embodiment, preferably, ResNet-50 includes a preset number of convolutional layers and uses a bottleneck module as a residual module. ResNet-50 is pre-trained on ImageNet.

[0054] To improve feature reuse, this invention uses residual connections to add the original input tensor to the output tensor, thus preserving feature information.

[0055] In the above embodiments, preferably, the output feature of the residual module is the sum of the residual mapping of the input feature after passing through the convolutional layer and the input feature. The residual module uses a 1×1 convolutional layer for dimensionality reduction, a 3×3 convolutional layer, and a 1×1 convolutional layer for dimensionality restoration.

[0056] Specifically, in the ResNet model, the residual block contains a short-circuit connection that spans multiple neural layers, allowing input data to be added to the output data. This operation enables the network to easily learn the feature representation of the identity mapping, ensuring that the learned features can be directly passed to subsequent layers without introducing excessive parameters or computational overhead. Through the design of the residual block, a network model can maintain its performance during training even with the addition of more neural layers.

[0057] like Figure 2 As shown, the ResNet residual module contains two mappings: the identity mapping and the residual mapping. The identity mapping directly passes the input features to the next layer, while the residual mapping performs some non-linear transformations on the input features before passing them, used to learn the difference between the input and output. The calculation process of the residual module is shown in the following formula: output = F(x) + x. Where F(x) represents the residual mapping, which is usually composed of a series of convolutional layers.

[0058] The bottleneck module consists of stacked convolutional layers and skip connections. That is, after the input features undergo nonlinear transformation, the output data is added to the input data, and finally passed through an activation function. This makes up a complete residual module.

[0059] like Figure 3 As shown, the bottleneck module uses a 1×1 convolutional layer for dimensionality reduction, followed by a 3×3 convolutional layer, and finally a 1×1 convolutional layer for dimensionality restoration, thus reducing computational cost while maintaining computational accuracy. Especially in deep ResNet networks, using the bottleneck module as a residual module can significantly reduce the number of model parameters and computational cost, thereby lowering training costs.

[0060] like Figure 4 As shown, in the above implementation, preferably, the dilated convolution module uses four parallel branches. The first three branches fuse feature maps of different receptive field sizes in a cascade manner, and the last branch adds image-level features through global average pooling and then reduces the number of channels to the number of input channels through convolution operations.

[0061] Specifically, the encoder output possesses rich high-level features. To improve the utilization of the encoder output, this invention introduces a dilated convolution module at the bottom of the encoder to expand the receptive field. Four parallel dilated convolution branches provide four different receptive field sizes. The size of the feature map's receptive field can be modified by adjusting the dilation rate. The grid effect exists in dilated convolutions, leading to the loss of local information and a decrease in the performance of small-sized target detection. Concatenating dilated convolutions with different dilation rates helps to extract information from a wider range of pixels and avoids the grid effect.

[0062] In the above implementation, preferably, the encoder reduces the image size to a preset ratio through five downsampling operations, for example, reducing the image size from 512×512 to 16×16. The receptive field is increased accordingly based on the dilatancy of the cascaded branches in the dilated convolution module. The dilatancy rates of the cascaded convolutions in the dilated convolution module are set to 1, 2, and 4, and the receptive fields of the parallel branches are set to 3, 7, and 15, respectively. This allows for approximate coverage of the entire feature map. Furthermore, an average pooling branch is added to obtain global information about the image.

[0063] like Figure 5 As shown, in the above embodiment, preferably, the dilated convolution module adjusts the receptive field of pixels in the output feature map by adjusting the dilation rate using the following formula:

[0064] S=(r-1)×(k-1)+k

[0065] Where S is the receptive field, r is the hole ratio, and k is the kernel size.

[0066] Specifically, the receptive field of a 3×3 convolution is 3×3, and the receptive field of a 3×3 convolution with a dilatation rate of 2 is 5×5.

[0067] In the above embodiments, preferably, the formula for calculating the receptive field size of the cascaded dilated convolutions in the dilated convolution module is:

[0068] S' = S1 + S2 - 1

[0069] Where S' is the receptive field size after cascading, and S1 and S2 are the receptive field sizes of the two dilated convolutions in the cascade, respectively.

[0070] like Figure 6 The diagram shows the process of increasing the receptive field of the dilated convolution module.

[0071] like Figure 7 As shown, in the above embodiment, preferably, the strip convolution module includes four strip convolutions: horizontal, vertical, left diagonal, and right diagonal. The input tensor is fed into the four parallel strip convolution branches after a 1×1 convolution operation. Then, the four output tensors are connected and a 1×1 convolution operation is performed to capture long-range context information in the feature map from four different directions, so that each pixel in the output feature map can be associated with pixels in the four directions of the input feature map.

[0072] Specifically, most existing convolutional neural networks typically use square kernels for convolution, learning feature maps within a square window. This type of convolution is well-suited for most natural objects with blocky shapes. However, the roads studied in this invention are characterized by their narrowness, large span, and continuous distribution. If square convolutions were used to capture road features, a very large square might be needed to cover an entire road, inevitably merging a lot of irrelevant information. Therefore, the strip convolution of this invention uses elongated kernels, which better conform to the shape of roads and can better capture features of road regions. Strip convolutions capture local contextual information along a specified spatial direction and prevent irrelevant regions from interfering with feature learning.

[0073] The strip convolution module uses four strip convolutions—horizontal, vertical, left diagonal, and right diagonal—to capture long-range contextual information from four different directions, enabling each pixel in the output feature map to be associated with multiple pixels in the input feature map in four directions. Let K∈R 2r+1 The kernel is a 2r+1 banded convolution, D = (D h D w ) represents the direction of the strip convolution kernel K, Y D ∈R C×H×W This represents the result of a strip convolution operation. The formula for strip convolution is as follows:

[0074]

[0075] Where X*K represents the convolution operation, and D is the direction vector of the strip convolution. For horizontal, vertical, left diagonal, and right diagonal convolutions, D is (0,1), (1,0), (1,1), and (-1,1), respectively. For the convolution kernel K, this invention sets r to 4, so that the strip convolution has 9 parameters, the same number of parameters as the 3×3 square convolution operation.

[0076] In the strip convolution module, the input tensor is fed into four parallel branches after a 1×1 convolution operation, each branch performing a strip convolution in one direction. The four output tensors are then concatenated and subjected to a 1×1 convolution operation.

[0077] This invention also proposes a road weak information extraction system based on ASU-Net, which applies the road weak information extraction method based on ASU-Net disclosed in any of the above embodiments, including:

[0078] The model building module is extracted to construct the road extraction network model ASU-Net using an encoder-decoder structure. ResNet-50 is used as the encoder, and a dilated convolution module is introduced into the encoder, while a strip convolution module is introduced into the decoder.

[0079] The image sampling and processing module is used to input remote sensing satellite road images into the ASU-Net model. The encoder performs downsampling operations on the images and expands the receptive field through parallel cascaded branches with preset dilatancy rates in the dilated convolution module.

[0080] The road feature extraction module is used to upsample the features learned by the encoder through transposed convolution, and at the same time, it uses a strip convolution module to extract linear road features using strip convolution in four different directions.

[0081] The road weak information extraction system based on ASU-Net disclosed in the above embodiments has the same functions as the steps of the road weak information extraction method based on ASU-Net disclosed in the above embodiments. In the implementation process, the operation is carried out with reference to the above embodiments, and will not be repeated here.

[0082] The method and system for extracting weak road information based on ASU-Net disclosed in the above embodiments are used to illustrate the process and effect of the method and system in the following embodiments.

[0083] Example 1:

[0084] The embodiments use the Zhouqu road dataset to evaluate the performance of the proposed method.

[0085] like Figure 8 As shown, the Zhouqu Road Dataset is aerial imagery data of Zhouqu County, Gansu Province. This dataset covers road images in various scenes, including the county seat, suburbs, rural areas, and mountainous areas. This dataset can, to some extent, represent the road characteristics of county towns in the mountainous and hilly regions of central and western China.

[0086] This dataset has a resolution of 5 meters and an image size of 1024×1024. A total of 1305 images are labeled in this dataset. The dataset is divided into three parts according to an 8:1:1 ratio: 910 images for training, 195 for validation, and 200 for testing. The training dataset is enhanced through scaling, rotation, and cropping.

[0087] To validate the model proposed in this paper, a comparative analysis was conducted with the results extracted by BiSeNet and LinkNet on the Zhouqu dataset. Table 1 lists the experimental results of the model evaluation. All data in the table are rounded to four decimal places.

[0088] Table 1 Comparison of experimental results for different models on the Zhouqu road dataset.

[0089]

[0090] It can be seen that the ASUNET network structure proposed in this invention outperforms the BiSeNet and LinkNet models in all metrics.

[0091] Figure 9 The comparison of different algorithms on the Zhouqu road dataset is shown, where (a) is remote sensing imagery; (b) is road labeling; (c) is BiSeNet; (d) is LinkNet; and (e) is ASU-Net. The similarity between dried-up riverbeds and roads is high. Compared to other algorithms, the ASU-Net algorithm can avoid the influence of the riverbed to some extent (I). Even in situations where the background and road information are highly similar (II) road information is highly similar to village houses; III) road information is highly similar to surrounding farmland, the algorithm can still extract complete and clear roads relatively well. Furthermore, in suburban scenes (IV), the algorithm can even identify unlabeled road segments with high continuity and completeness. In more complex urban scenes (V) city edges, VI city centers, the proposed roads better match people's visual perception of roads.

[0092] To verify the effectiveness of the dilated convolution module and the striped convolution module proposed in this invention, an ablation experiment was conducted using the U-Net model as the base model. The results of the ablation experiment are shown in Table 2 below.

[0093] Table 2 Ablation Results

[0094]

[0095] With the addition of ResNet-50, dilated convolution, and strip convolution, the accuracy has been continuously improved, especially after adding dilated convolution. However, after adding strip convolution, the accuracy improvement is not significant. This may be because in this dataset, there are landforms similar to dry riverbeds or roads, where strip convolution has limited effect.

[0096] like Figure 10As shown, (a) is a remote sensing image; (b) is a road label; (c) is U-Net+ResNet-50; (d) is U-Net+ResNet-50+hole; and (e) is U-Net+ResNet-50+hole+band (ASU-Net).

[0097] As can be seen from the Zhouqu road dataset, with the addition of ResNet, dilated convolution, and strip convolution, the road information becomes clearer and more complete in farmland scenarios (I and II), while effectively removing the impact of bare soil and exposed riverbeds on road information extraction; in the urban fringe, the extraction of road information changes significantly, while the extraction effect of roads in the city center is better.

[0098] In summary, the method proposed in this invention outperforms several mainstream semantic segmentation networks, such as BiSeNet, LinkNet, and U-Net, in terms of accuracy, recall, F1 score, crossover ratio (CR), and mean crossover ratio (MCR) on the Zhouqu road dataset, indicating that the ASU-Net model has good performance.

[0099] The ASU-Net model achieves a better balance between precision and recall by introducing dilated and striped convolutional modules, thereby improving the F1 score and mIoU. Results are also improved when handling roads that are occluded by vegetation or have a similar color to the background.

[0100] To verify the effectiveness of the dilated convolution module and the strip convolution module, experiments were conducted using different models: ResU-Net, ResU-Net + dilated convolution module, and ResU-Net + dilated convolution module + strip convolution module. Experimental results show that the introduction of the dilated convolution module and the strip convolution module achieves a better balance between precision and recall, thereby improving the F1 score and mean intersection-over-union (MCI) performance, demonstrating that these two modules can perform more accurate segmentation for road image extraction.

[0101] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for extracting weak road information based on ASU-Net, characterized in that, include: A road extraction network model, ASU-Net, is constructed using an encoder-decoder structure. ResNet-50 is used as the encoder, comprising a predetermined number of convolutional layers and a bottleneck module as the residual module. The ResNet-50 is pre-trained on ImageNet. The output feature of the residual module is the sum of the residual mapping of the input features after convolutional layers and the input features. The residual module uses a 1×1 convolutional layer for dimensionality reduction, a 3×3 convolutional layer, and a 1×1 convolutional layer for dimensionality restoration. A dilated convolutional module is introduced in the encoder, and a strip convolutional module is introduced in the decoder. The remote sensing satellite road image is input into the ASU-Net model. The encoder performs downsampling operation on the image. The encoder reduces the image size to a preset ratio through five downsampling operations. The first four layers of the encoder feature map and the output of the dilated convolution module are passed to the decoder through skip connections to perform multi-scale feature information fusion. Simultaneously, the receptive field is expanded through parallel cascaded branches with preset dilation rates in the dilated convolution module. The dilated convolution module employs four parallel branches. The first three branches fuse feature maps of different receptive field sizes in a cascaded manner, with dilation rates of 1, 2, and 4 respectively, corresponding to receptive fields of 3, 7, and 15. The last branch adds image-level features through global average pooling and then reduces the number of channels to the input channel number through convolution operations. The encoder increases the receptive field accordingly based on the dilation rates of the cascaded branches in the dilated convolution module. The formula for calculating the pixel receptive field in the output feature map by adjusting the dilation rate of the dilated convolution module is as follows: S =( r -1)×( k -1)+ k wherein, S is a receptive field, r is a void ratio, k is a kernel size; In the dilated convolution module, the formula for calculating the receptive field size of the cascaded dilated convolutions is: S ＇= S 1+ S 2-1 wherein, S ' is the size of the receptive field after concatenation, S 1 and S 2 are the size of the receptive field of each of the two dilated convolutions, respectively; The features learned by the encoder are upsampled by transposed convolution. Simultaneously, the linear road features are extracted by the strip convolution module using four strip convolutions in four different directions. The strip convolution module includes four strip convolutions: horizontal, vertical, left diagonal, and right diagonal. The input tensor is fed into the four parallel strip convolution branches after a 1×1 convolution operation. Then, the four output tensors are concatenated and a 1×1 convolution operation is performed. This captures long-range contextual information in the feature map from four different directions, so that each pixel in the output feature map can be associated with pixels in the input feature map in the four directions.

2. An ASU-Net-based road weak information extraction system, characterized in that, The method for extracting weak road information based on ASU-Net as described in claim 1 includes: The extraction model building module is used to construct a road extraction network model ASU-Net using an encoder-decoder structure, wherein ResNet-50 is used as the encoder, and a dilated convolution module is introduced into the encoder, and a strip convolution module is introduced into the decoder. The image sampling and processing module is used to input remote sensing satellite road images into the ASU-Net model. The encoder performs downsampling operations on the images and expands the receptive field through parallel cascaded branches with a preset hole rate in the dilated convolution module. The road feature extraction module is used to upsample the features learned by the encoder through transposed convolution, and at the same time, to extract linear road features through the strip convolution module using strip convolution in four different directions.

Images