Remote sensing image road extraction method based on whole-course feature information aggregation

Abstract

The invention discloses a remote sensing image road extraction method based on whole-course feature information aggregation, and the method comprises the steps: carrying out the road information extraction of a processed remote sensing image through building a remote sensing image road extraction model of a strip convolution attention module based on space attention, and obtaining a binary image of the remote sensing image containing the road information; therefore, the road information in the remote sensing image to be subjected to road extraction is extracted. Through a strip convolution attention-based remote sensing image road extraction model with a space attention mechanism, the problem that a strip with holes is extracted after a road is shielded can be solved, and meanwhile, on the basis of a multi-scale information aggregation module, when dense trees or tall buildings shield the road, the strip with holes is extracted. And clear road information can be extracted, so that the extracted road information is complete, and the precision of extracting the road information is improved.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing image segmentation and deep learning technology, and in particular to a remote sensing image road extraction method based on full-process feature information aggregation. Background Art

[0002] Remote sensing technology plays a key role in application areas such as urban planning, vehicle navigation, autonomous driving, and geographic information updating. Among them, extracting paths from high-resolution remote sensing images is an important downstream task in remote sensing image technology analysis. High-resolution remote sensing images provide unprecedented detailed ground information, making the extraction of road networks more accurate and comprehensive, and can meet the needs of various industries for large-scale regional road data. Therefore, extracting road networks from high-resolution remote sensing images has always been a research hotspot in road network extraction.

[0003] Traditional road extraction methods often rely on manually designed features. Not only is the feature design process cumbersome and costly, but the features also have limited expressiveness and generalization capabilities. In some scenarios, they are sensitive to noise and complex backgrounds in the image, resulting in low accuracy of the extraction results. Therefore, traditional methods have difficulty coping with complex road morphology and diverse environmental conditions in high-resolution images.

[0004] With the development of deep learning technology, deep neural networks have shown great potential in processing high-resolution remote sensing images. Compared with traditional methods, the architecture of convolutional neural networks can better capture complex features and patterns in images, thereby improving the accuracy and robustness of extraction. Some research tasks have applied deep network models with encoder-decoder structures to road extraction tasks and achieved satisfactory results. However, it is challenging to extract roads from high-resolution images. Images usually cover a large area of ​​the surface and contain various complex natural and man-made environments. Especially the occlusion problem of trees and buildings. This will lead to fragmented results of road extraction. From a geometric point of view, roads in remote sensing images are discretely distributed strips. When the road is occluded, the extracted feature result is a strip with holes. Roads in remote sensing images are usually identified by their linear morphology, texture and color features. When dense trees or tall buildings block the road, the originally clear road information becomes invisible in these occluded areas. The final extracted result is fragmented and the occluded road information is lost. Summary of the invention

[0005] The invention discloses a remote sensing image road extraction method based on full-course feature information aggregation to overcome the above technical problems.

[0006] In order to achieve the above object, the technical solution of the present invention is:

[0007] A method for extracting roads from remote sensing images based on full-course feature information aggregation includes the following steps:

[0008] S1: Based on remote sensing technology, obtain remote sensing images for road extraction;

[0009] S2: Preprocessing the remote sensing image to be subjected to road extraction;

[0010] S3: Establish a remote sensing image road extraction model based on a strip convolution attention module of spatial attention and a multi-scale information aggregation module; obtain a binary image of the remote sensing image containing road information based on the preprocessed remote sensing image to complete the extraction of road information in the remote sensing image to be subjected to road extraction.

[0011] Furthermore, the remote sensing image road extraction model includes an encoder module based on spatial attention, a decoder module, a first multi-scale information aggregation module MSIA1, a second multi-scale information aggregation module MSIA2, and a third multi-scale information aggregation module MSIA3;

[0012] The encoder module includes a first encoder layer E1, a second encoder layer E2, a third encoder layer E3, a fourth encoder layer E4, a first strip convolutional attention module SCAM1 based on spatial attention, a second strip convolutional attention module SCAM2 based on spatial attention, and a third strip convolutional attention module SCAM3 based on spatial attention;

[0013] The decoder module includes a first decoder layer D1, a second decoder layer D2, a third decoder layer D3, and a fourth decoder layer D4;

[0014] The first encoder layer E1 is used to obtain road edge features according to the preprocessed remote sensing image;

[0015] The input end of the first strip convolutional attention module SCAM1 based on spatial attention is connected to the output end of the first encoder layer E1, and is used to obtain spatially enhanced road edge features according to the road edge features;

[0016] The input end of the second encoder layer E2 is connected to the output end of the first strip convolutional attention module SCAM1 based on spatial attention, and is used to obtain road shape features according to the spatially enhanced road edge features;

[0017] The input end of the second strip convolutional attention module SCAM2 based on spatial attention is connected to the output end of the second encoder layer E2, and is used to obtain the local context feature of the road area according to the road shape feature;

[0018] The input end of the third encoder layer E3 is connected to the output end of the second strip convolutional attention module SCAM2 based on spatial attention, and is used to obtain complete road and road background feature information according to the local context features of the road area;

[0019] The input end of the third strip convolutional attention module SCAM3 based on spatial attention is connected to the output end of the third encoder layer E3, and is used to obtain globally enhanced road features according to the complete road and road background feature information;

[0020] The input end of the fourth encoder layer E4 is connected to the output end of the third strip convolutional attention module SCAM3 based on spatial attention, and is used to obtain the final feature information of the global semantic expression according to the globally enhanced road features;

[0021] The input end of the first multi-scale information aggregation module MSIA1 is respectively connected to the output end of the first strip convolutional attention module SCAM1 based on spatial attention, the output end of the second strip convolutional attention module SCAM2 based on spatial attention, the output end of the third strip convolutional attention module SCAM3 based on spatial attention, and the output end of the fourth encoder layer E4, so as to obtain the fused high-level road features according to the spatially enhanced road edge features, the local context features of the road area, the globally enhanced road features, and the final feature information of the global semantic expression;

[0022] The input end of the second multi-scale information aggregation module MSIA2 is respectively connected to the output end of the first strip convolutional attention module SCAM1 based on spatial attention, the output end of the second strip convolutional attention module SCAM2 based on spatial attention, the output end of the third strip convolutional attention module SCAM3 based on spatial attention, and the output end of the fourth encoder layer E4, and is used to obtain the fused secondary road features according to the spatially enhanced road edge features, the local context features of the road area, the globally enhanced road features, and the final feature information of the global semantic expression;

[0023] The input end of the third multi-scale information aggregation module MSIA3 is respectively connected to the output end of the first strip convolutional attention module SCAM1 based on spatial attention, the output end of the second strip convolutional attention module SCAM2 based on spatial attention, the output end of the third strip convolutional attention module SCAM3 based on spatial attention and the output end of the fourth encoder layer E4, and is used to obtain the fused preliminary road features according to the spatially enhanced road edge features, the local context features of the road area, the globally enhanced road features and the final feature information of the global semantic expression;

[0024] The input end of the fourth decoder layer D4 is connected to the output end of the fourth encoder layer E4, and is used to obtain a global decoding feature according to the final feature information of the global semantic expression;

[0025] The output end of the fourth decoder layer D4 and the output end of the first multi-scale information aggregation module MSIA1 are both connected to a first adding module, which is used to add the global decoding features and the fused high-level road features element by element to obtain the fused high-level road decoding features;

[0026] The input end of the third decoder layer D3 is connected to the output end of the first adding module, and is used to obtain the second stage decoding features according to the fused high-level road decoding features;

[0027] The output end of the third decoder layer D3 and the output end of the second multi-scale information aggregation module MSIA2 are both connected to the second addition module; used to add the second stage decoding features and the fused secondary road features element by element to obtain the fused secondary road decoding features;

[0028] The input end of the second decoder layer D2 is connected to the output end of the second adding module, and is used to obtain the decoding features of the third stage according to the fused secondary road decoding features;

[0029] The output end of the second decoder layer D2 and the output end of the third multi-scale information aggregation module MSIA3 are both connected to a third adding module, which is used to add the decoded features of the third stage and the fused preliminary road features element by element to obtain the fused preliminary road decoding features;

[0030] The input end of the first decoder layer D1 is connected to the output end of the third adding module, and is used to obtain a binary image of the remote sensing image containing road information based on the fused preliminary road decoding features to complete the extraction of the remote sensing image of road information.

[0031] Furthermore, the relationship between the fused high-level road features, the fused secondary road features, and the fused preliminary road features is:

[0032] L1=2L2=4L3; FB1=0.5FB2=0.25FB3

[0033] Among them, L1 represents the number of channels of the fused high-level road features; L2 represents the number of channels of the fused secondary road features; L3 represents the number of channels of the fused preliminary road features; FB1 represents the resolution of the fused high-level road features; FB2 represents the resolution of the fused secondary road features; FB3 represents the resolution of the fused preliminary road features.

[0034] Furthermore, the relationship between the global decoding feature, the second stage decoding feature and the third stage decoding feature is:

[0035] Lj1=2Lj2=4Lj3; FBj1=0.5FBj2=0.25FBj3

[0036] Among them, Lj1 represents the number of channels of the global decoding feature; Lj2 represents the number of channels of the decoding feature of the second stage; Lj3 represents the number of channels of the decoding feature of the third stage; FBj1 represents the resolution of the global decoding feature; FBj2 represents the resolution of the decoding feature of the second stage; FBj3 represents the resolution of the decoding feature of the third stage.

[0037] Furthermore, the structures of the first stripe convolutional attention module SCAM1 based on spatial attention, the second stripe convolutional attention module SCAM2 based on spatial attention, and the third stripe convolutional attention module SCAM3 based on spatial attention are the same;

[0038] The first stripe convolution attention module SCAM1 based on spatial attention includes a channel attention module and a spatial attention module;

[0039] The input end of the channel attention module is connected to the output end of the first encoder layer E1, and is used to obtain spatial enhancement features according to the output of the first encoder layer E1.

[0040] The output end of the channel attention module and the output end of the first encoder layer E1 are both connected to a first multiplication module, which is used to obtain enhanced road edge features according to the spatial enhancement features and the road edge features;

[0041] The input end of the spatial attention module is connected to the first multiplication module, and is used to obtain a spatial context enhancement feature according to the enhanced road edge feature;

[0042] The output end of the first multiplication module and the output end of the spatial attention module are both connected to the second multiplication module, for obtaining comprehensive enhanced features based on the spatial context enhancement features and enhanced road edge features; the output end of the second multiplication module is connected to the input end of the second encoder layer E2, for obtaining spatially enhanced road edge features based on the comprehensive enhancement features.

[0043] Further, the spatial attention module includes a second horizontal strip convolution layer, a vertical strip convolution layer, a second main diagonal strip convolution layer, a second sub-diagonal strip convolution layer, a third horizontal strip convolution layer, a third vertical strip convolution layer, a third main diagonal strip convolution layer, a third sub-diagonal strip convolution layer, and a fourth horizontal strip convolution layer, a fourth vertical strip convolution layer, a fourth main diagonal strip convolution layer, and a fourth sub-diagonal strip convolution layer;

[0044] The input ends of the second horizontal strip convolution layer, the second main diagonal strip convolution layer, the third horizontal strip convolution layer, the third main diagonal strip convolution layer, the fourth horizontal strip convolution layer, and the fourth main diagonal strip convolution layer are all connected to the output end of the first multiplication module; and are used to obtain local horizontal edge features, local main diagonal edge features, fine horizontal edge features, fine main diagonal edge features, global horizontal edge features, and global main diagonal edge features according to the enhanced road edge features output by the first multiplication module;

[0045] The input end of the second vertical strip convolution layer is connected to the output end of the second horizontal strip convolution layer to obtain local horizontal and vertical edge features according to the local horizontal edge features;

[0046] The input end of the second sub-diagonal strip convolution layer is connected to the output end of the second main diagonal strip convolution layer to obtain local main and sub-diagonal edge features according to the local main diagonal edge features;

[0047] The input end of the third vertical strip convolution layer is connected to the output end of the second horizontal strip convolution layer to obtain fine horizontal and vertical edge features according to the fine horizontal edge features;

[0048] The input end of the third sub-diagonal strip convolution layer is connected to the output end of the second main diagonal strip convolution layer to obtain fine main and sub-diagonal edge features according to the fine main diagonal edge features;

[0049] The input end of the fourth vertical strip convolution layer is connected to the output end of the second horizontal strip convolution layer to obtain global horizontal and vertical edge features according to the global horizontal edge features;

[0050] The input end of the fourth sub-diagonal strip convolution layer is connected to the output end of the second main diagonal strip convolution layer to obtain the global main and sub-diagonal edge features according to the global main diagonal edge features;

[0051] The output end of the second vertical strip convolution layer, the output end of the second sub-diagonal strip convolution layer, the output end of the third vertical strip convolution layer, the output end of the third sub-diagonal strip convolution layer, the output end of the fourth vertical strip convolution layer, the output end of the fourth sub-diagonal strip convolution layer, the output end and the output end of the first multiplication module are all connected to the fourth multiplication module, and are used to obtain spatial context enhancement features according to the enhanced road edge features, local horizontal and vertical edge features, local main and secondary diagonal edge features, fine horizontal and vertical edge features, fine main and secondary diagonal edge features, global horizontal and vertical edge features, and global main and secondary diagonal edge features output by the first multiplication module.

[0052] Further, the first multi-scale information aggregation module MSIA1, the second multi-scale information aggregation module MSIA2, and the third multi-scale information aggregation module MSIA3 have the same structure;

[0053] The first multi-scale information aggregation module MSIA1 includes: a horizontal strip convolution layer, a vertical strip convolution layer, a first diagonal strip convolution layer, a second diagonal strip convolution layer and a square convolution layer;

[0054] The input ends of the horizontal strip convolution, the vertical strip convolution, the first diagonal strip convolution, the second diagonal strip convolution and the square convolution are all connected to the output end of the first strip convolution attention module SCAM1 based on spatial attention, and are used to obtain horizontal strip convolution features, vertical strip convolution features, first diagonal strip convolution features, second diagonal strip convolution features, and square convolution features according to the spatially enhanced road edge features, respectively.

[0055] The output end of the first strip convolution attention module SCAM1 based on spatial attention and the output segment of the horizontal strip convolution are both connected to the input end of the first splicing module C1, and are respectively used to obtain the spliced ​​vertical enhancement features according to the contour features of the horizontal road and the horizontal strip convolution features obtained after the first average pooling of the spatially enhanced road edge features;

[0056] The output end of the first strip convolution attention module SCAM1 based on spatial attention and the output segment of the square convolution are both connected to the input end of the third splicing module C3, and are respectively used to obtain the spliced ​​horizontal enhancement features according to the contour features and square convolution features of the longitudinal road obtained after performing the second average pooling on the spatially enhanced road edge features;

[0057] The output ends of the first splicing module C1, the vertical strip convolution, the first diagonal strip convolution, the second diagonal strip convolution and the third splicing module C3 are all connected to the second splicing module C2, and are respectively used to obtain the spliced ​​multi-scale fusion features according to the spliced ​​vertical enhancement features, the vertical strip convolution features, the first diagonal strip convolution features, the second diagonal strip convolution features, and the spliced ​​horizontal enhancement features;

[0058] The output end of the second splicing module C2 is connected to the fifth adding module, and is used to obtain the output of the first multi-scale information aggregation module MSIA1, that is, the fused preliminary road features, according to the spliced ​​multi-scale fusion features.

[0059] Beneficial effects: The present invention provides a remote sensing image road extraction method based on full-process feature information aggregation. By establishing a remote sensing image road extraction model based on a strip convolution attention module with spatial attention, the road information of the processed remote sensing image is extracted, and a binary image of the remote sensing image containing the road information is obtained to complete the extraction of road information in the remote sensing image to be subjected to road extraction. The remote sensing image road extraction model based on strip convolution attention with a spatial attention mechanism can solve the problem that when the road is blocked, the extracted strip is a strip with holes. At the same time, based on the multi-scale information aggregation module, when dense trees or tall buildings block the road, clear road information can be extracted, so that the extracted road information is complete, and the accuracy of road information extraction is improved.

[0060] Figures illustrate clear road information

[0061] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0062] Figure 1 It is a flow chart of the method for extracting roads from remote sensing images based on full-course feature information aggregation of the present invention;

[0063] Figure 2 is an overall structural diagram of a remote sensing image road extraction model in an embodiment of the present invention;

[0064] Figure 3 FIG. 1 is an overall structural diagram of a SCAM in an embodiment of the present invention;

[0065] Figure 4 4 is a structural diagram of a channel attention module in a SCAM in an embodiment of the present invention;

[0066] Figure 5 4 is a structural diagram of a spatial attention module in SCAM in an embodiment of the present invention;

[0067] Figure 6 A schematic diagram of an optimized jump connection in an embodiment of the present invention;

[0068] Figure 7 Schematic diagram of the MSIA module structure in an embodiment of the present invention;

[0069] Figure 8 Schematic diagram of comparison of road extraction results in an embodiment of the present invention. DETAILED DESCRIPTION

[0070] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0071] This embodiment introduces a remote sensing image road extraction method based on full-range feature information aggregation, which extracts road information based on the geometric texture features of the remote sensing image. Figure 1 As shown,

[0072] S1: Based on remote sensing technology, obtain remote sensing images for road extraction;

[0073] S2: Preprocessing the remote sensing image to be subjected to road extraction;

[0074] Specifically, the acquired high-resolution remote sensing image with a resolution greater than or equal to 1024×1024 is cropped according to the cache capacity of the processing terminal. In this embodiment, the obtained initial image is cropped to a size of 512×512 to prevent the server processing terminal from being unable to smoothly extract roads due to the image being too large.

[0075] S3: Establish a remote sensing image road extraction model based on a strip convolution attention module of spatial attention; obtain a binary image of the remote sensing image containing road information according to the preprocessed remote sensing image to complete the extraction of road information in the remote sensing image to be subjected to road extraction.

[0076] Preferably, the remote sensing image road extraction model includes an encoder module based on spatial attention, a decoder module, a first multi-scale information aggregation module MSIA1, a second multi-scale information aggregation module MSIA2, and a third multi-scale information aggregation module MSIA3;

[0077] The encoder module includes a first encoder layer E1, a second encoder layer E2, a third encoder layer E3, a fourth encoder layer E4, a first strip convolutional attention module SCAM1 based on spatial attention, a second strip convolutional attention module SCAM2 based on spatial attention, and a third strip convolutional attention module SCAM3 based on spatial attention;

[0078] The decoder module includes a first decoder layer D1, a second decoder layer D2, a third decoder layer D3, and a fourth decoder layer D4;

[0079] The first encoder layer E1 is used to obtain road edge features according to the preprocessed remote sensing image;

[0080] The input end of the first strip convolutional attention module SCAM1 based on spatial attention is connected to the output end of the first encoder layer E1, and is used to obtain spatially enhanced road edge features according to the road edge features;

[0081] The input end of the second encoder layer E2 is connected to the output end of the first strip convolutional attention module SCAM1 based on spatial attention, and is used to obtain road shape features according to the spatially enhanced road edge features;

[0082] The input end of the second strip convolutional attention module SCAM2 based on spatial attention is connected to the output end of the second encoder layer E2, and is used to obtain the local context feature of the road area according to the road shape feature;

[0083] The input end of the third encoder layer E3 is connected to the output end of the second strip convolutional attention module SCAM2 based on spatial attention, and is used to obtain complete road and road background feature information according to the local context features of the road area;

[0084] The input end of the third strip convolutional attention module SCAM3 based on spatial attention is connected to the output end of the third encoder layer E3, and is used to obtain globally enhanced road features according to the complete road and road background feature information;

[0085] The input end of the fourth encoder layer E4 is connected to the output end of the third strip convolutional attention module SCAM3 based on spatial attention, and is used to obtain the final feature information of the global semantic expression according to the globally enhanced road features;

[0086] The input end of the first multi-scale information aggregation module MSIA1 is respectively connected to the output end of the first strip convolutional attention module SCAM1 based on spatial attention, the output end of the second strip convolutional attention module SCAM2 based on spatial attention, the output end of the third strip convolutional attention module SCAM3 based on spatial attention, and the output end of the fourth encoder layer E4, for obtaining fused high-level road features according to the spatially enhanced road edges, local context features of the road area, globally enhanced road features, and final feature information of global semantic expression;

[0087] The input end of the second multi-scale information aggregation module MSIA2 is respectively connected to the output end of the first strip convolutional attention module SCAM1 based on spatial attention, the output end of the second strip convolutional attention module SCAM2 based on spatial attention, the output end of the third strip convolutional attention module SCAM3 based on spatial attention, and the output end of the fourth encoder layer E4, and is used to obtain the fused secondary road features according to the spatially enhanced road edge features, the local context features of the road area, the globally enhanced road features, and the final feature information of the global semantic expression;

[0088] The input end of the third multi-scale information aggregation module MSIA3 is respectively connected to the output end of the first strip convolutional attention module SCAM1 based on spatial attention, the output end of the second strip convolutional attention module SCAM2 based on spatial attention, the output end of the third strip convolutional attention module SCAM3 based on spatial attention and the output end of the fourth encoder layer E4, and is used to obtain the fused preliminary road features according to the spatially enhanced road edge features, the local context features of the road area, the globally enhanced road features and the final feature information of the global semantic expression;

[0089] Among them, the relationship between the fused high-level road features, the fused secondary road features, and the fused preliminary road features is;

[0090] L1=2L2=4L3; FB1=0.5FB2=0.25FB3

[0091] Among them, L1 represents the number of channels of the fused high-level road features; L2 represents the number of channels of the fused secondary road features; L3 represents the number of channels of the fused preliminary road features; FB1 represents the resolution of the fused high-level road features; FB2 represents the resolution of the fused secondary road features; FB3 represents the resolution of the fused preliminary road features;

[0092] In the embodiment, the number of channels of the fused high-level road features is 384, and the resolution is 32×32; the number of channels of the fused secondary road features is 192, and the resolution is 64×64; the number of channels of the fused preliminary road features is 96, and the resolution is 128×128.

[0093] The input end of the fourth decoder layer D4 is connected to the output end of the fourth encoder layer E4, and is used to obtain a global decoding feature according to the final feature information of the global semantic expression;

[0094] The output end of the fourth decoder layer D4 and the output end of the first multi-scale information aggregation module MSIA1 are both connected to a first adding module, which is used to add the global decoding features and the fused high-level road features element by element to obtain the fused high-level road decoding features;

[0095] The input end of the third decoder layer D3 is connected to the output end of the first adding module, and is used to obtain the second stage decoding features according to the fused high-level road decoding features;

[0096] The output end of the third decoder layer D3 and the output end of the second multi-scale information aggregation module MSIA2 are both connected to the second addition module; used to add the second stage decoding features and the fused secondary road features element by element to obtain the fused secondary road decoding features;

[0097] The input end of the second decoder layer D2 is connected to the output end of the second adding module, and is used to obtain the decoding features of the third stage according to the fused secondary road decoding features;

[0098] The output end of the second decoder layer D2 and the output end of the third multi-scale information aggregation module MSIA3 are both connected to a third adding module, which is used to add the decoded features of the third stage and the fused preliminary road features element by element to obtain the fused preliminary road decoding features;

[0099] The input end of the first decoder layer D1 is connected to the output end of the third adding module, and is used to obtain a binary image of the remote sensing image containing road information based on the fused preliminary road decoding features to complete the extraction of the remote sensing image of road information.

[0100] Preferably, the relationship between the global decoding feature, the second stage decoding feature and the third stage decoding feature is:

[0101] Lj1=2Lj2=4Lj3; FBj1=0.5FBj2=0.25FBj3

[0102] Among them, Lj1 represents the number of channels of the global decoding feature; Lj2 represents the number of channels of the decoding feature of the second stage; Lj3 represents the number of channels of the decoding feature of the third stage; FBj1 represents the resolution of the global decoding feature; FBj2 represents the resolution of the decoding feature of the second stage; FBj3 represents the resolution of the decoding feature of the third stage;

[0103] The relationship between the fused preliminary road decoding features, the fused secondary road decoding features and the fused advanced road decoding features is consistent with the relationship between the global decoding features, the second stage decoding features and the third stage decoding features.

[0104] In this embodiment, the number of channels of the global decoding feature is 384, and the resolution is 32×32; the number of channels of the second-stage decoding feature is 192, and the resolution is 64×64; the number of channels of the third-stage decoding feature is 96, and the resolution is 128×128.

[0105] Preferably, the structures of the first stripe convolutional attention module SCAM1 based on spatial attention, the second stripe convolutional attention module SCAM2 based on spatial attention, and the third stripe convolutional attention module SCAM3 based on spatial attention are the same;

[0106] The first stripe convolution attention module SCAM1 based on spatial attention includes a channel attention module and a spatial attention module;

[0107] Specifically, this embodiment takes the first strip convolutional attention module SCAM1 based on spatial attention as an example:

[0108] The input end of the channel attention module is connected to the output end of the first encoder layer E1, and is used to obtain spatial enhancement features according to the output of the first encoder layer E1;

[0109] The output end of the channel attention module and the output end of the first encoder layer E1 are both connected to a first multiplication module, which is used to obtain enhanced road edge features according to the spatial enhancement features and the road edge features;

[0110] The input end of the spatial attention module is connected to the first multiplication module, and is used to obtain a spatial context enhancement feature according to the enhanced road edge feature;

[0111] The output end of the first multiplication module and the output end of the spatial attention module are both connected to the second multiplication module, and are used to obtain comprehensive enhanced features according to the spatial context enhancement features and the enhanced road edge features; the output end of the second multiplication module is connected to the input end of the second encoder layer E2, and is used to obtain spatially enhanced road edge features according to the comprehensive enhancement features;

[0112] Preferably, the spatial attention module includes a second horizontal strip convolution layer, a vertical strip convolution layer, a second main diagonal strip convolution layer, a second sub-diagonal strip convolution layer, a third horizontal strip convolution layer, a third vertical strip convolution layer, a third main diagonal strip convolution layer, a third sub-diagonal strip convolution layer, and a fourth horizontal strip convolution layer, a fourth vertical strip convolution layer, a fourth main diagonal strip convolution layer, and a fourth sub-diagonal strip convolution layer;

[0113] The input ends of the second horizontal strip convolution layer, the second main diagonal strip convolution layer, the third horizontal strip convolution layer, the third main diagonal strip convolution layer, the fourth horizontal strip convolution layer, and the fourth main diagonal strip convolution layer are all connected to the output end of the first multiplication module; and are used to obtain local horizontal edge features, local main diagonal edge features, fine horizontal edge features, fine main diagonal edge features, global horizontal edge features, and global main diagonal edge features according to the enhanced road edge features output by the first multiplication module;

[0114] The input end of the second vertical strip convolution layer is connected to the output end of the second horizontal strip convolution layer to obtain local horizontal and vertical edge features according to the local horizontal edge features;

[0115] The input end of the second sub-diagonal strip convolution layer is connected to the output end of the second main diagonal strip convolution layer to obtain local main and sub-diagonal edge features according to the local main diagonal edge features;

[0116] The input end of the third vertical strip convolution layer is connected to the output end of the second horizontal strip convolution layer to obtain fine horizontal and vertical edge features according to the fine horizontal edge features;

[0117] The input end of the third sub-diagonal strip convolution layer is connected to the output end of the second main diagonal strip convolution layer to obtain fine main and sub-diagonal edge features according to the fine main diagonal edge features;

[0118] The input end of the fourth vertical strip convolution layer is connected to the output end of the second horizontal strip convolution layer to obtain global horizontal and vertical edge features according to the global horizontal edge features;

[0119] The input end of the fourth sub-diagonal strip convolution layer is connected to the output end of the second main diagonal strip convolution layer to obtain the global main and sub-diagonal edge features according to the global main diagonal edge features;

[0120] The output end of the second vertical strip convolution layer, the output end of the second sub-diagonal strip convolution layer, the output end of the third vertical strip convolution layer, the output end of the third sub-diagonal strip convolution layer, the output end of the fourth vertical strip convolution layer, the output end of the fourth sub-diagonal strip convolution layer, the output end and the output end of the first multiplication module are all connected to the fourth multiplication module, and are used to obtain spatial context enhancement features according to the enhanced road edge features, local horizontal and vertical edge features, local main and sub-diagonal edge features, fine horizontal and vertical edge features, fine main and sub-diagonal edge features, global horizontal and vertical edge features, and global main and sub-diagonal edge features output by the first multiplication module;

[0121] Specifically, in this embodiment, the size of the second horizontal strip convolution layer is 1×7, the size of the second vertical strip convolution layer is 7×1, and the size of the second main diagonal strip convolution layer and the second sub-diagonal strip convolution layer are both 7×7; the size of the third horizontal strip convolution layer is 1×11, the size of the third vertical strip convolution layer is 11×1, and the size of the third main diagonal strip convolution layer and the third sub-diagonal strip convolution layer are both 11×11; the size of the fourth horizontal strip convolution layer is 1×21, the size of the fourth vertical strip convolution layer is 21×1, and the size of the fourth main diagonal strip convolution layer and the fourth sub-diagonal strip convolution layer are both 21×21. Through the spatial attention module based on strip convolution in this embodiment, the long-distance dependencies in the image can be captured.

[0122] Preferably, the first multi-scale information aggregation module MSIA1, the second multi-scale information aggregation module MSIA2, and the third multi-scale information aggregation module MSIA3 have the same structure;

[0123] Taking the first multi-scale information aggregation module MSIA1 as an example: the first multi-scale information aggregation module MSIA1 includes: a horizontal strip convolution layer, a vertical strip convolution layer, a first diagonal strip convolution layer, a second diagonal strip convolution layer and a square convolution layer; the input ends of the horizontal strip convolution, the vertical strip convolution, the first diagonal strip convolution, the second diagonal strip convolution and the square convolution are all connected to the output end of the first strip convolution attention module SCAM1 based on spatial attention, and are respectively used to obtain horizontal strip convolution features, vertical strip convolution features, first diagonal strip convolution features, second diagonal strip convolution features, and square convolution features according to the spatially enhanced road edge features.

[0124] The output end of the first strip convolution attention module SCAM1 based on spatial attention and the output segment of the horizontal strip convolution are both connected to the input end of the first splicing module C1, and are respectively used to obtain the spliced ​​vertical enhancement features according to the contour features of the horizontal road and the horizontal strip convolution features obtained after the first average pooling of the spatially enhanced road edge features;

[0125] The output end of the first strip convolution attention module SCAM1 based on spatial attention and the output segment of the square convolution are both connected to the input end of the third splicing module C3, and are respectively used to obtain the spliced ​​horizontal enhancement features according to the contour features and square convolution features of the longitudinal road obtained after performing the second average pooling on the spatially enhanced road edge features;

[0126] The output ends of the first splicing module C1, the vertical strip convolution layer, the first diagonal strip convolution layer, the second diagonal strip convolution layer and the third splicing module C3 are all connected to the second splicing module C2, and are respectively used to obtain the spliced ​​multi-scale fusion features according to the spliced ​​vertical enhancement features, the vertical strip convolution features, the first diagonal strip convolution features, the second diagonal strip convolution features, and the spliced ​​horizontal enhancement features;

[0127] The output end of the second splicing module C2 is connected to the fifth adding module, and is used to obtain the output of the first multi-scale information aggregation module MSIA1, that is, the fused preliminary road features, according to the spliced ​​multi-scale fusion features.

[0128] Specifically, in this embodiment, the size of the horizontal strip convolution layer is 1×3, the size of the vertical strip convolution layer is 3×1, the size of the first diagonal strip convolution layer is 3×3, the size of the second diagonal strip convolution layer is 3×3, the size of the square convolution layer is 3×3, and the expansion rate of each of the above strip convolution layers is 1. Through the setting of the strip convolution layer of this embodiment, the remote sensing road extraction model implemented in this embodiment can expand the receptive field during the extraction process and reduce the hole problem caused by occlusion.

[0129] Specifically, Figure 2 The overall structure of the remote sensing image road extraction model of this embodiment is shown, which is composed of an encoder-decoder network, and a strip convolutional attention module SCAM (Strip Convolutional Attention Module) is embedded on the encoder to assist in extracting features, where the MSIA (Multi-scale Information Aggregation Modules) module is designed to align with the texture features of the road area with occlusion and correspond to road features of different scales. This embodiment can aggregate different levels of feature information from high dimension to low dimension through improved skip connections.

[0130] Figure 3The overall structure of SCAM of this embodiment is shown, which includes a channel attention module and a spatial attention module based on deep strip convolution. The channel attention module is used to enhance the importance of specific feature channels while suppressing less relevant factors. By adaptively adjusting the weight of each channel, the model can better capture key features. The introduction of the spatial attention module enables the model to better capture local dependencies in specific directions, thereby improving the overall feature expression capability.

[0131] Figure 4 The channel attention module in SCAM of this embodiment is shown. Given an intermediate feature map F∈R C×H×W As input, it undergoes two global pooling operations, namely adaptive average pooling and adaptive maximum pooling, where F represents the road edge features output by E1; R represents the range of the feature map; C represents the number of channels; H represents the height; and W represents the width. These pooling operations aggregate the spatial information of each channel into a single scalar, thereby generating a channel-level description vector V∈R C×1×1 ,. Next, the two channel-level description vectors are transformed nonlinearly through a shared fully connected network, namely a multi-layer perceptron (MLP), to generate a weight vector M∈R used to represent the importance of each channel. C×1×1 , where M represents the weight vector. The weight vectors are activated by the sigmoid function and then added together to fuse the weight information under the two pooling methods. Finally, the weight vector is reshaped into a refined feature F with channel attention. c ∈R C×H×W , where the channel attention calculation process can be summarized as:

[0132]

[0133] MC(F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

[0134] Among them, F c represents the refined features; MC(F) represents the features extracted by channel attention; represents element-by-element multiplication, σ represents the sigmoid function; MLP(·) represents nonlinear transformation; AvgPool(F) represents average pooling; MaxPool(F) represents maximum pooling;

[0135] Figure 5It is the spatial attention module in SCAM, which uses a multi-scale deep strip convolution structure to enhance the network's ability to extract road features. Different convolution kernel sizes and shapes make the network sensitive to road results of different sizes and directions. The feature map after the multi-scale convolution operation is added to the original feature map to obtain the fused road feature information. Then, a 1×1 convolution layer is used to generate a spatial attention map, which is used to emphasize the parts of the feature map with important spatial position information. The calculation process of spatial attention can be described as follows:

[0136]

[0137] Among them, F S Represents the features extracted by SCAM; DWConv represents deep convolution; Branch i ,i∈{1,2,3...,6} represents the i-th branch in the spatial attention module; i represents the sequence number of the branch in the spatial attention module; MS(F C ) represents the features extracted by spatial attention; Conv 1×1 Represents 1×1 convolution;

[0138] Figure 6 The schematic diagram of the optimized jump connection of this embodiment is shown, taking the E2 branch as an example, where the output of E1 is transmitted through convolution, and the outputs of E3 and E4 are restored to the same size as E2 through linear interpolation method before being transmitted.

[0139] Figure 7 The MSIA module of this embodiment is demonstrated, using 3×3 square convolution, 3×1 horizontal strip convolution and 1×3 vertical strip convolution, 5 branches composed of two diagonal strip convolutions, and the expansion rates are 1, 2, 4, and 8. Horizontal pooling branches and vertical pooling branches are introduced to aggregate the global information of the road in the horizontal and vertical directions. On the five convolution branches, element-by-element addition is used to fuse the feature maps in different receptive fields. The feature maps finally output by different branches are then connected and a 1×1 convolution operation is performed. Finally, the original feature map is combined with the fused feature map to obtain rich road semantic information.

[0140] Specifically, this embodiment adds SCAM on the basis of the encoder to enhance the global information capture capability. After that, the encoder not only outputs the semantic feature map to the decoder, but also the encoder of each layer outputs the spatial information extracted by itself to the decoders at each level through the optimized jump connection. The output of each layer is then aggregated as the input of the MSIA module.

[0141] Table 1-Quantitative comparison of SCIFNet and the latest methods on the DeepGlobe dataset

[0142]

[0143] Table 2 - Quantitative comparison of SCIFNet with state-of-the-art methods on the Massachusetts road dataset

[0144]

[0145] Specifically, this embodiment compares the performance of our proposed method SCIFNet (i.e., the skip connection and context information aggregation network of this embodiment) with several state-of-the-art road extraction methods on DeepGlobe and Massachusetts datasets, including: LinkNet50, Global Aware Road Detection Network with Multi-Scale Residual Learning (GAMSNet), Swin Transformer (i.e., the backbone network for feature extraction), Remote Sensing Road Extraction Network based on Road Background and Full-length Features (RCFSNet), and Occlusion Aware Extraction Network (OARENet).

[0146] From Table 1, we can see the quantitative results of different methods for road extraction on the DeepGlobe dataset. Our proposed SCIFNet achieves the best IoU (68.78%), precision (82.29%) and F1 score (81.50%).

[0147] From Table 2, our proposed method still achieves the best IoU (59.00%), precision (79.01%), and F1 score (74.21%) on the Massachusetts dataset.

[0148] from Figure 8 In the figure, the visualization results of roads extracted by different methods on the DeepGlobe dataset are shown. The selected images all have road areas blocked by trees or buildings. In these cases, the extraction results of LinkNet50, GAMSNet, SwinTransformer, RCFSNet, OARENet, etc. show that the blocked roads are fragmented or ignored. However, the extraction results using the SCIFNet model we proposed are not affected by the occlusion of trees or buildings, and the extracted roads are connected. Among them, Figure 8In the figure, (c) is the extraction result of LinkNet50, (d) is the extraction result of GAMSNet, (e) is the extraction result of RCFSNet, (f) is the extraction result of OARENet, (g) is the extraction result of Swin Transformer, (h) is the extraction result of SCIFNet, (a) is the original remote sensing image, and (b) is the real road mask in the remote sensing image.

[0149] This embodiment has the following beneficial effects:

[0150] The third strip convolution attention module SCAM based on spatial attention proposed in this embodiment can serially aggregate the attention mechanism and the strip convolution. While retaining the channel attention weight, it effectively extracts spatial features by using a multi-scale strip convolution module, making the road features obtained by the encoder network richer.

[0151] This embodiment proposes a multi-scale information aggregation module MSIA for the special geometric shapes of densely occluded roads. It can adopt different expansion rates in feature maps, use strip convolutions in multiple directions and scales to capture long-range context information, and aggregate global information of road features from the horizontal and vertical directions through two branches.

[0152] In this embodiment, an optimized skip connection method is used to construct multiple skip connections that fuse features from different stages, and combine high-resolution detail information with low-resolution semantic information and output them to the decoder branch to improve the accuracy of decoder feature recovery.

[0153] 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 with 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

1. A method for extracting roads from remote sensing images based on full-range feature information aggregation, characterized in that: The steps include: S1: Based on remote sensing technology, obtain remote sensing images for road extraction; S2: Preprocessing the remote sensing image to be subjected to road extraction; S3: Establish a remote sensing image road extraction model based on a strip convolution attention module of spatial attention and a multi-scale information aggregation module; obtain a binary image of the remote sensing image containing road information based on the preprocessed remote sensing image to complete the extraction of road information in the remote sensing image to be subjected to road extraction.

2. The method for extracting roads from remote sensing images based on full-range feature information aggregation according to claim 1, characterized in that: The remote sensing image road extraction model includes an encoder module based on spatial attention, a decoder module, a first multi-scale information aggregation module MSIA1, a second multi-scale information aggregation module MSIA2, and a third multi-scale information aggregation module MSIA3; The encoder module includes a first encoder layer E1, a second encoder layer E2, a third encoder layer E3, a fourth encoder layer E4, a first strip convolutional attention module SCAM1 based on spatial attention, a second strip convolutional attention module SCAM2 based on spatial attention, and a third strip convolutional attention module SCAM3 based on spatial attention; The decoder module includes a first decoder layer D1, a second decoder layer D2, a third decoder layer D3, and a fourth decoder layer D4; The first encoder layer E1 is used to obtain road edge features according to the preprocessed remote sensing image; The input end of the first strip convolutional attention module SCAM1 based on spatial attention is connected to the output end of the first encoder layer E1, and is used to obtain spatially enhanced road edge features according to the road edge features; The input end of the second encoder layer E2 is connected to the output end of the first strip convolutional attention module SCAM1 based on spatial attention, and is used to obtain road shape features according to the spatially enhanced road edge features; The input end of the second strip convolutional attention module SCAM2 based on spatial attention is connected to the output end of the second encoder layer E2, and is used to obtain the local context feature of the road area according to the road shape feature; The input end of the third encoder layer E3 is connected to the output end of the second strip convolutional attention module SCAM2 based on spatial attention, and is used to obtain complete road and road background feature information according to the local context features of the road area; The input end of the third strip convolutional attention module SCAM3 based on spatial attention is connected to the output end of the third encoder layer E3, and is used to obtain globally enhanced road features according to the complete road and road background feature information; The input end of the fourth encoder layer E4 is connected to the output end of the third strip convolutional attention module SCAM3 based on spatial attention, and is used to obtain the final feature information of the global semantic expression according to the globally enhanced road features; The input end of the first multi-scale information aggregation module MSIA1 is respectively connected to the output end of the first strip convolutional attention module SCAM1 based on spatial attention, the output end of the second strip convolutional attention module SCAM2 based on spatial attention, the output end of the third strip convolutional attention module SCAM3 based on spatial attention, and the output end of the fourth encoder layer E4, so as to obtain the fused high-level road features according to the spatially enhanced road edge features, the local context features of the road area, the globally enhanced road features, and the final feature information of the global semantic expression; The input end of the second multi-scale information aggregation module MSIA2 is respectively connected to the output end of the first strip convolutional attention module SCAM1 based on spatial attention, the output end of the second strip convolutional attention module SCAM2 based on spatial attention, the output end of the third strip convolutional attention module SCAM3 based on spatial attention, and the output end of the fourth encoder layer E4, and is used to obtain the fused secondary road features according to the spatially enhanced road edge features, the local context features of the road area, the globally enhanced road features, and the final feature information of the global semantic expression; The input end of the third multi-scale information aggregation module MSIA3 is respectively connected to the output end of the first strip convolutional attention module SCAM1 based on spatial attention, the output end of the second strip convolutional attention module SCAM2 based on spatial attention, the output end of the third strip convolutional attention module SCAM3 based on spatial attention and the output end of the fourth encoder layer E4, and is used to obtain the fused preliminary road features according to the spatially enhanced road edge features, the local context features of the road area, the globally enhanced road features and the final feature information of the global semantic expression; The input end of the fourth decoder layer D4 is connected to the output end of the fourth encoder layer E4, and is used to obtain a global decoding feature according to the final feature information of the global semantic expression; The output end of the fourth decoder layer D4 and the output end of the first multi-scale information aggregation module MSIA1 are both connected to a first adding module, which is used to add the global decoding features and the fused high-level road features element by element to obtain the fused high-level road decoding features; The input end of the third decoder layer D3 is connected to the output end of the first adding module, and is used to obtain the second stage decoding features according to the fused high-level road decoding features; The output end of the third decoder layer D3 and the output end of the second multi-scale information aggregation module MSIA2 are both connected to the second addition module; used to add the second stage decoding features and the fused secondary road features element by element to obtain the fused secondary road decoding features; The input end of the second decoder layer D2 is connected to the output end of the second adding module, and is used to obtain the decoding features of the third stage according to the fused secondary road decoding features; The output end of the second decoder layer D2 and the output end of the third multi-scale information aggregation module MSIA3 are both connected to a third adding module, which is used to add the decoded features of the third stage and the fused preliminary road features element by element to obtain the fused preliminary road decoding features; The input end of the first decoder layer D1 is connected to the output end of the third adding module, and is used to obtain a binary image of the remote sensing image containing road information based on the fused preliminary road decoding features to complete the extraction of the remote sensing image of road information.

3. The method for extracting roads from remote sensing images based on full-range feature information aggregation according to claim 2, characterized in that: The relationship between the fused high-level road features, the fused secondary road features, and the fused preliminary road features is: L1=2L2=4L3; FB1=0.5FB2=0.25FB3 Among them, L1 represents the number of channels of the fused high-level road features; L2 represents the number of channels of the fused secondary road features; L3 represents the number of channels of the fused preliminary road features; FB1 represents the resolution of the fused high-level road features; FB2 represents the resolution of the fused secondary road features; FB3 represents the resolution of the fused preliminary road features.

4. The method for extracting roads from remote sensing images based on full-range feature information aggregation according to claim 2, characterized in that: The relationship between the global decoding features, the second stage decoding features and the third stage decoding features is: Lj1=2Lj2=4Lj3; FBj1=0.5FBj2=0.25FBj3 Among them, Lj1 represents the number of channels of the global decoding feature; Lj2 represents the number of channels of the decoding feature of the second stage; Lj3 represents the number of channels of the decoding feature of the third stage; FBj1 represents the resolution of the global decoding feature; FBj2 represents the resolution of the decoding feature of the second stage; FBj3 represents the resolution of the decoding feature of the third stage.

5. The method for extracting roads from remote sensing images based on full-range feature information aggregation according to claim 2, characterized in that: The structures of the first stripe convolutional attention module SCAM1 based on spatial attention, the second stripe convolutional attention module SCAM2 based on spatial attention, and the third stripe convolutional attention module SCAM3 based on spatial attention are the same; The first stripe convolutional attention module SCAM1 based on spatial attention includes a channel attention module and a spatial attention module; The input end of the channel attention module is connected to the output end of the first encoder layer E1, and is used to obtain spatial enhancement features according to the output of the first encoder layer E1. The output end of the channel attention module and the output end of the first encoder layer E1 are both connected to a first multiplication module, which is used to obtain enhanced road edge features according to the spatial enhancement features and the road edge features; The input end of the spatial attention module is connected to the first multiplication module, and is used to obtain a spatial context enhancement feature according to the enhanced road edge feature; The output end of the first multiplication module and the output end of the spatial attention module are both connected to the second multiplication module, for obtaining comprehensive enhanced features based on the spatial context enhancement features and enhanced road edge features; the output end of the second multiplication module is connected to the input end of the second encoder layer E2, for obtaining spatially enhanced road edge features based on the comprehensive enhancement features.

6. The method for extracting roads from remote sensing images based on full-range feature information aggregation according to claim 2, characterized in that: The spatial attention module includes a second horizontal strip convolution layer, a vertical strip convolution layer, a second main diagonal strip convolution layer, a second sub-diagonal strip convolution layer, a third horizontal strip convolution layer, a third vertical strip convolution layer, a third main diagonal strip convolution layer, a third sub-diagonal strip convolution layer, and a fourth horizontal strip convolution layer, a fourth vertical strip convolution layer, a fourth main diagonal strip convolution layer, and a fourth sub-diagonal strip convolution layer; The input ends of the second horizontal strip convolution layer, the second main diagonal strip convolution layer, the third horizontal strip convolution layer, the third main diagonal strip convolution layer, the fourth horizontal strip convolution layer, and the fourth main diagonal strip convolution layer are all connected to the output end of the first multiplication module; and are used to obtain local horizontal edge features, local main diagonal edge features, fine horizontal edge features, fine main diagonal edge features, global horizontal edge features, and global main diagonal edge features according to the enhanced road edge features output by the first multiplication module; The input end of the second vertical strip convolution layer is connected to the output end of the second horizontal strip convolution layer to obtain local horizontal and vertical edge features according to the local horizontal edge features; The input end of the second sub-diagonal strip convolution layer is connected to the output end of the second main diagonal strip convolution layer to obtain local main and sub-diagonal edge features according to the local main diagonal edge features; The input end of the third vertical strip convolution layer is connected to the output end of the second horizontal strip convolution layer to obtain fine horizontal and vertical edge features according to the fine horizontal edge features; The input end of the third sub-diagonal strip convolution layer is connected to the output end of the second main diagonal strip convolution layer to obtain fine main and sub-diagonal edge features according to the fine main diagonal edge features; The input end of the fourth vertical strip convolution layer is connected to the output end of the second horizontal strip convolution layer to obtain global horizontal and vertical edge features according to the global horizontal edge features; The input end of the fourth sub-diagonal strip convolution layer is connected to the output end of the second main diagonal strip convolution layer to obtain the global main and sub-diagonal edge features according to the global main diagonal edge features; The output end of the second vertical strip convolution layer, the output end of the second sub-diagonal strip convolution layer, the output end of the third vertical strip convolution layer, the output end of the third sub-diagonal strip convolution layer, the output end of the fourth vertical strip convolution layer, the output end of the fourth sub-diagonal strip convolution layer, the output end and the output end of the first multiplication module are all connected to the fourth multiplication module, and are used to obtain spatial context enhancement features according to the enhanced road edge features, local horizontal and vertical edge features, local main and secondary diagonal edge features, fine horizontal and vertical edge features, fine main and secondary diagonal edge features, global horizontal and vertical edge features, and global main and secondary diagonal edge features output by the first multiplication module.

7. The method for extracting roads from remote sensing images based on full-range feature information aggregation according to claim 2, characterized in that: The structures of the first multi-scale information aggregation module MSIA1, the second multi-scale information aggregation module MSIA2, and the third multi-scale information aggregation module MSIA3 are the same; The first multi-scale information aggregation module MSIA1 includes: a horizontal strip convolution layer, a vertical strip convolution layer, a first diagonal strip convolution layer, a second diagonal strip convolution layer and a square convolution layer; The input ends of the horizontal strip convolution, the vertical strip convolution, the first diagonal strip convolution, the second diagonal strip convolution and the square convolution are all connected to the output end of the first strip convolution attention module SCAM1 based on spatial attention, and are used to obtain horizontal strip convolution features, vertical strip convolution features, first diagonal strip convolution features, second diagonal strip convolution features, and square convolution features according to the spatially enhanced road edge features, respectively. The output end of the first strip convolution attention module SCAM1 based on spatial attention and the output segment of the horizontal strip convolution are both connected to the input end of the first splicing module C1, and are respectively used to obtain the spliced ​​vertical enhancement features according to the contour features of the horizontal road and the horizontal strip convolution features obtained after the first average pooling of the spatially enhanced road edge features; The output end of the first strip convolution attention module SCAM1 based on spatial attention and the output segment of the square convolution are both connected to the input end of the third splicing module C3, and are respectively used to obtain the spliced ​​horizontal enhancement features according to the contour features and square convolution features of the longitudinal road obtained after performing the second average pooling on the spatially enhanced road edge features; The output ends of the first splicing module C1, the vertical strip convolution, the first diagonal strip convolution, the second diagonal strip convolution and the third splicing module C3 are all connected to the second splicing module C2, and are respectively used to obtain the spliced ​​multi-scale fusion features according to the spliced ​​vertical enhancement features, the vertical strip convolution features, the first diagonal strip convolution features, the second diagonal strip convolution features, and the spliced ​​horizontal enhancement features; The output end of the second splicing module C2 is connected to the fifth adding module, and is used to obtain the output of the first multi-scale information aggregation module MSIA1, that is, the fused preliminary road features, according to the spliced ​​multi-scale fusion features.

Images