A high-resolution remote sensing road extraction method based on space-frequency domain features

Abstract

The application provides a high-resolution remote sensing road extraction method based on space-frequency domain features, comprising: acquiring a high-resolution remote sensing image, dividing the remote sensing image into preset pixel blocks through block embedding, and obtaining encoded input features after linear mapping; a spatial information feature processing branch adds a residual error of the encoded input features after performing convolution processing on the encoded input features to the original encoded input features, to obtain spatial domain detail enhancement features; the spatial domain detail enhancement features and enhanced frequency domain features are spliced in the channel dimension, and after performing channel fusion through a convolution layer, frequency domain enhancement fusion features are obtained; the spatial domain global feature map and the frequency domain enhancement fusion features are spliced in the channel dimension, to obtain single-level fusion features in the encoding stage.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a high-resolution remote sensing road extraction method based on spatial frequency domain features. Background Technology

[0002] In the field of remote sensing image analysis, road extraction technology has irreplaceable value for key applications such as urban planning, traffic management, and disaster emergency response. This technology helps us quickly acquire road network information from a high-altitude perspective, providing important support for social development and public safety. However, despite some progress in this field, existing technologies still have many shortcomings when facing the practical needs of complex environments, and breakthroughs are urgently needed.

[0003] Currently, many road extraction methods struggle to adapt to varying road morphology in different scenarios when processing high-resolution remote sensing imagery, particularly in urban, rural, and mountainous environments. These methods typically fail to fully capture the overall road layout and subtle features, resulting in inaccurate extraction results when faced with complex background interference, often exhibiting problems such as road breaks or blurred boundaries. This limitation makes it difficult for these techniques to meet the high-precision requirements in practical applications.

[0004] A deeper technical challenge lies in effectively integrating the overall structural information of a road with its local edge details. As a continuous network, the overall layout of a road relies on long-distance spatial relationships for understanding—for example, how a road extends from the city center to the suburbs. Local details, on the other hand, are reflected in the clarity of the road edges, such as whether there are obstructions along the roadside or whether the boundaries with the surrounding environment are clear. If both aspects of information cannot be considered simultaneously, the extracted road structure may be incomplete or have blurred boundaries. For instance, in urban environments, roads may be obscured by tall buildings or vehicles, making their overall direction difficult to determine; while in mountainous scenes, the winding shape of roads and the undulating terrain can make edge details difficult to discern. This contradiction between the overall and the local has become a major obstacle to technological advancement.

[0005] Therefore, how to capture both the overall spatial relationship and local edge details of roads in complex and ever-changing scenarios, and ensure that the extraction results can reflect the whole picture of the road network and clearly present the boundary features, has become a key problem that this study urgently needs to solve. Summary of the Invention

[0006] This invention provides a high-resolution remote sensing road extraction method based on spatial frequency domain features, mainly comprising: High-resolution remote sensing images are acquired, divided into preset pixel blocks through block embedding, and linearly mapped to obtain coded input features; The encoded input features are synchronously input into the global spatial feature extraction module and the Fourier feature module, and the spatial domain feature extraction and frequency domain feature enhancement processing are performed in parallel. The global spatial feature extraction module is built based on the Swing Transformer and obtains a spatial global feature map by capturing long-distance road dependencies and global context information. The Fourier feature module contains frequency domain and spatial information feature processing branches that are parallel and share encoded input features. The frequency domain information feature processing branch performs a two-dimensional fast Fourier transform on the encoded input features, divides the frequency domain signal into multiple groups according to the channel, extracts and enhances the low-frequency main body and high-frequency edge information, and obtains the enhanced frequency domain features through inverse Fourier transform; The spatial information feature processing branch convolves the encoded input features and adds them to the original feature residual to obtain spatial domain detail enhancement features; The spatial domain detail enhancement features and the enhanced frequency domain feature channels are concatenated and then fused through convolutional layer channels to obtain the frequency domain enhanced fused features; The spatial domain global feature map is concatenated with the frequency domain enhanced fusion feature channel to obtain a single-level fusion feature in the coding stage. In the encoding stage, corresponding modules are set up layer by layer to complete multi-level feature extraction. The encoded outputs of each level are transmitted to the decoding stage through skip connections to complete multi-level feature fusion. In the decoding stage, the feature map resolution is restored level by level. The road probability prediction map is output through the segmentation layer and obtained by binarization.

[0007] Furthermore, the acquisition of high-resolution remote sensing images, followed by block embedding, division into preset pixel blocks, and linear mapping to obtain coded input features, includes: Acquire high-resolution remote sensing images, and divide the high-resolution remote sensing images into image blocks of preset pixels as input; The image block is divided into pixel blocks of a preset pixel size by block embedding, and linear mapping is completed. After mapping the pixel dimension of the pixel block to the preset embedding dimension, the encoded input feature is obtained. The encoded input features are synchronously input into the global spatial feature extraction module and the Fourier feature module. The spatial domain global feature map output by the global spatial feature extraction module provides a spatial prior for the overall road structure for subsequent frequency domain feature processing.

[0008] Furthermore, the frequency domain information feature processing branch performs a two-dimensional fast Fourier transform on the encoded input features, divides the frequency domain signal into multiple groups according to channels, extracts and enhances the low-frequency main body and high-frequency edge information, and obtains enhanced frequency domain features through inverse Fourier transform, including: Perform a two-dimensional fast Fourier transform on the encoded input features, and divide the resulting complex frequency domain features into a preset number of signal groups in the channel dimension; The signals of a predetermined number of groups are divided into a low-frequency processing group and a high-frequency processing group. The two groups of grouped signals, which are divided based on the complete frequency domain features of the same two-dimensional fast Fourier transform output, are processed synchronously and in parallel. The low-frequency processing group contains half of the signals of the predetermined number of groups, and each group is processed by simulating a low-pass filter through an average pooling layer with a different window size to extract the low-frequency main information. The high-frequency processing group contains the remaining half of the signals of the predetermined number of groups, and each group is processed by simulating a high-pass filter through a convolutional layer with a different kernel size to extract the high-frequency edge information. The signals of a preset number of groups corresponding to the low-frequency main information and the high-frequency edge information are spliced ​​together in the channel dimension to obtain a fused frequency domain signal after integrating the low-frequency fusion feature and the high-frequency fusion feature. An inverse Fourier transform is performed on the fused frequency domain signal to obtain enhanced frequency domain features.

[0009] Furthermore, the step of concatenating the spatial domain detail enhancement features with the enhanced frequency domain feature channels and fusing them through convolutional layer channels to obtain the frequency domain enhanced fusion features includes: After aligning the number of channels and the size of the spatial domain detail enhancement features and the enhanced frequency domain features, the channel dimensions are then stitched together. After the splicing is completed, channel fusion is performed through a convolutional layer to obtain frequency domain enhanced fusion features.

[0010] Furthermore, the encoding stage progressively sets up corresponding modules to complete multi-level feature extraction, including: The encoding stage is set up with a four-layer encoding structure arranged sequentially from input to output. The first two layers are configured with parallel global spatial feature extraction modules and Fourier feature modules to output single-layer fused features. The last two layers are configured with only global spatial feature extraction modules to output single-layer spatial features. The two types of features are unified as encoded output features. The output of the previous level coding structure is used as the input of the next level coding structure. The single-level coding structure completes feature extraction and fusion to obtain single-level fused features. Then, the features are downsampled by a convolutional layer with a preset stride, and the downsampled features are input into the next level coding structure. The features passed to the decoding stage through skip connections are the output features of each level in the encoding stage, which correspond one-to-one with the upsampling level resolution in the decoding stage.

[0011] Furthermore, the decoding stage progressively restores the feature map resolution, outputs a road probability prediction map through the segmentation layer, and obtains the road segmentation result after binarization, including: The decoding stage adopts a four-layer decoding structure with the same number of layers as the encoding stage and progressively corresponding resolutions. Each layer of the decoding structure performs upsampling on the input features through transposed convolution. After upsampling is completed, the encoded output features of the corresponding level in the encoding stage are obtained through skip connections. The encoded output features are then concatenated with the upsampled features of the current level in the channel dimension to obtain the decoded fusion features. The decoded and fused features are input into a preset aggregated feature integration module. The aggregated feature integration module performs multi-level semantic information integration on the decoded and fused features to complete multi-level feature fusion. The integrated features are output and input into the next level of decoding structure. After the decoding stage completes upsampling and feature integration at all levels, the road probability prediction map is output through the segmentation layer, and then binarized post-processing is performed to obtain the binarized road segmentation result.

[0012] Furthermore, the step of dividing the image block into pixel blocks of a preset pixel size through block embedding, completing a linear mapping, and mapping the pixel dimension of the pixel block to a preset embedding dimension, yields the encoded input features, including: The image block is divided using a pixel size smaller than the preset pixel to generate multiple non-overlapping pixel blocks; Perform a linear mapping on each pixel block to map the pixel dimension of the pixel block to a preset embedding dimension, thereby obtaining the encoded input features.

[0013] Furthermore, the step of concatenating the six sets of signals corresponding to the low-frequency main information and the high-frequency edge information along the channel dimension to obtain the fused frequency domain signal after integrating the low-frequency fusion features and the high-frequency fusion features includes: The signals output from the average pooling layer of half of the preset number of low-frequency processing groups are spliced ​​together along the channel dimension to obtain low-frequency fusion features. The signals output from the remaining half of the pre-set number of convolutional layers in the high-frequency processing group are concatenated along the channel dimension to obtain high-frequency fusion features. The low-frequency fusion feature and the high-frequency fusion feature are concatenated along the channel dimension to obtain a fused frequency domain signal.

[0014] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses a high-resolution remote sensing road extraction method based on spatial and frequency domain features. Addressing the complexities of road extraction from high-resolution remote sensing images, particularly the challenges of diverse road structures, complex edge details, and missing global semantic information in urban, rural, and mountainous environments, this invention proposes a solution that integrates spatial and frequency domain features. By constructing a global spatial feature extraction module and a Fourier feature module, this invention captures long-distance dependencies and multi-scale frequency domain information. Simultaneously, by combining parallel processing and skip connection mechanisms in the encoding and decoding stages, it achieves accurate segmentation of the main road structure and edge details, significantly improving segmentation accuracy and robustness, especially in complex environments. Ultimately, this invention achieves efficient extraction of the overall road network layout and local details, significantly improving adaptability to multiple scenarios and providing reliable support for applications such as urban planning, traffic monitoring, and disaster emergency response, demonstrating excellent technical performance. Attached Figure Description

[0015] Figure 1 This is a flowchart of a high-resolution remote sensing road extraction method based on spatial frequency domain features according to the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] like Figure 1 This embodiment of a high-resolution remote sensing road extraction method based on spatial frequency domain features may specifically include: This invention provides a high-resolution remote sensing road extraction method based on spatial and frequency domain features, aiming to achieve accurate segmentation of road structures in high-resolution remote sensing images by combining feature information from both the spatial and frequency domains. The technical solution of this invention is described in detail below with reference to specific embodiments to make the objectives, technical solutions, and advantages of this invention clearer.

[0018] The method provided by this invention employs an encoder-decoder structure. The encoding stage consists of a backbone convolutional neural network, such as the Swin Transformer, responsible for extracting multi-scale, highly semantic spatial features layer by layer from the original input image. The final output of this encoding stage is the encoded input feature referred to in this invention, with dimensions of B×576×H×W, where B is the batch size, and H and W are the height and width of the feature map, serving as the common input source for subsequent dual-branch processing.

[0019] After the encoding phase, the network introduces two complementary feature enhancement branches in parallel: The global spatial feature extraction module directly operates on the above-mentioned encoded input features, and captures long-distance spatial contextual dependencies through spatial attention, dilated convolution or non-local operations; The Fourier feature module also starts with this encoded input feature, first transforming it to the frequency domain, and then enhancing the frequency domain representation capability through a grouped frequency domain modulation mechanism.

[0020] The encoding stage is the foundational stage for feature extraction in the entire network, and the encoded input features are the final output of this stage, serving as the common input for both the global spatial feature extraction module and the Fourier feature module. These two modules run in parallel after the encoding stage, and their outputs are integrated in the subsequent fusion stage before being fed into the decoder to complete the final segmentation.

[0021] In one embodiment, the method provided by the present invention is mainly aimed at complex scenarios of road extraction in high-resolution remote sensing images. By capturing long-range dependencies and global context information, and simultaneously fusing frequency domain features and spatial domain features, the accuracy and robustness of road segmentation are improved.

[0022] First, high-resolution remote sensing imagery is acquired as input and divided into predefined pixel blocks through block embedding. These blocks are then fed into a global spatial feature extraction module to capture long-range dependencies and global contextual information, extracting a spatial domain global feature map. This map is used to locate the overall road structure and provide spatial priors. Specifically, high-resolution remote sensing imagery typically has a large spatial coverage and rich detail, such as road networks in urban areas and narrow paths in rural areas. Dividing the imagery into predefined pixel blocks effectively reduces computational complexity while preserving local details, laying the foundation for subsequent feature extraction. The global spatial feature extraction module analyzes the long-range dependencies of roads in the imagery through a specific mechanism, such as road continuity and intersection distribution, thereby generating a spatial domain global feature map that characterizes the overall road structure. This feature map not only helps locate the main road direction but also provides important spatial prior information for subsequent frequency domain processing.

[0023] In one possible implementation, the image segmentation process can be adjusted based on image resolution and computational resources. For example, for urban area remote sensing imagery with a resolution of 0.5 meters per pixel, it can be divided into 512×512 pixel image blocks to ensure that each block contains sufficient road structure information while avoiding excessively large blocks that would overload the computation. During segmentation, an overlapping approach can be used, where adjacent image blocks overlap by a certain percentage, such as 10%, to prevent road structures from being fragmented at block boundaries. This method ensures the continuity of road information in subsequent processing. The segmented image blocks will then serve as input units for subsequent feature extraction.

[0024] Specifically, the block embedding process involves further subdividing each 512×512 pixel image block into non-overlapping 4×4 pixel blocks. Each block is flattened into a vector, and then a fully connected layer performs a linear mapping to convert the pixel dimension to a preset embedding dimension of 96, resulting in encoded input features. These features are then fed into the global spatial feature extraction module for processing. It should be noted that the size of the block embedding can be adjusted according to the specific application scenario. For example, in rural imagery, where road structures are relatively simple, the pixel block size can be appropriately increased to 32×32 to reduce computational load.

[0025] In another embodiment, image blocks are divided using a preset pixel size to generate smaller pixel blocks. These smaller pixel blocks are then input into the global spatial feature extraction module for processing. The selection of the preset pixel size needs to comprehensively consider the image resolution and the scale of road features. For example, when processing high-resolution images of urban areas, the road width may only be a few meters, corresponding to a small pixel range. Therefore, a pixel size of 8×8 or 16×16 can be selected for division to capture subtle road edge information. When processing suburban or rural images, the road width is larger and the structure is simpler, so a pixel size of 32×32 can be selected to cover a wider spatial range. During the division process, a sliding window can be used to extract the image block by block, ensuring that each small pixel block completely covers the image content. Specifically, these small pixel blocks undergo feature embedding processing before being input into the module. This involves mapping the pixel values ​​to a 96-dimensional high-dimensional feature vector through a linear transformation and adding a learnable absolute position code to preserve spatial location information. The global spatial feature extraction module is built based on the Swin Transformer, using an 8×8 window size, 3 attention heads, and 2 network layers. The input is an embedded feature vector sequence with dimensions H / 4×W / 4×96, where H and W are the height and width of the original image. First, the input is divided into 8×8 non-overlapping windows. Within each window, multi-head self-attention is computed: for each head, the input is projected as a query Q, key K, and value V matrix, each with dimensions 96 / 3=32; attention scores are then calculated. , where sqrt(32) is a scaling factor of approximately 5.657; then the output is a weighted sum of A multiplied by V. The two-layer network repeats this process to capture long-distance dependencies between pixel blocks, such as road intersections. The output is an enhanced feature map with dimensions H / 4×W / 4×96, supporting road structure localization. All feature vectors refer to the flattened sequence of all feature vectors in the enhanced feature map. Mapping of feature vector sequence to two-dimensional feature map: The enhanced feature map is essentially a three-dimensional tensor with dimensions H / 4×W / 4×96, and its spatial dimension H / 4 × W / 4 directly corresponds to the spatial position after downsampling of the original image. This three-dimensional tensor itself is a global feature map in the spatial domain. Recombining all feature vectors into a two-dimensional feature map actually means maintaining its original two-dimensional spatial structure. The global feature map in the spatial domain, H / 4×W / 4×96, will be directly used as the output of the global spatial feature extraction module. In the encoding stage, it will be combined with the frequency domain enhancement and fusion features output by the Fourier feature module, and after channel alignment, they will be spliced ​​together. It will also be passed to the decoding stage through skip connections and fused with the upsampled features.

[0026] In one possible implementation, after processing the relationships between small pixel blocks, the global spatial feature extraction module recombines all feature vectors into a two-dimensional feature map, i.e., a spatial domain global feature map. This feature map retains the overall structural information of roads in the image, such as the direction of main roads and the distribution of side roads, while also containing global contextual information, such as the spatial relationship between roads and surrounding buildings or green spaces. This method of generating feature maps can provide important spatial references for subsequent processing.

[0027] Next, the encoded input features are directly input into the frequency domain information feature processing branch and the spatial information feature processing branch of the Fourier feature module for Fast Fourier Transform (FFT). Specifically, before entering frequency domain processing, the encoded input features undergo channel dimension adjustment to ensure that their feature dimensions are consistent with the input requirements of subsequent processing modules. In one approach, after aligning the number of channels and the size of the spatial domain detail enhancement features and the enhanced frequency domain features, channel dimension concatenation is performed. Before concatenation, alignment is performed on the two features. Channel alignment uses a 1×1 convolutional layer to unify the two features to 96 channels; size alignment uses bilinear interpolation to adjust the feature map size to be consistent. A two-dimensional Fast Fourier Transform is performed on the feature map without windowing. The real and imaginary parts of the complex frequency domain features are processed simultaneously to transform the feature map from the spatial domain to the frequency domain in order to extract the frequency information of the road structure, such as the main road structure corresponding to low-frequency information and the road edge details corresponding to high-frequency information. This transformation process lays the foundation for subsequent multi-scale frequency domain feature fusion. Synchronous processing of the real and imaginary parts of complex frequency domain features can include: the encoded input features input to the Fourier feature module are transformed by Fast Fourier Transform (FFT) to obtain a complex frequency domain feature tensor. The original encoded input feature size is H / 4×W / 4×96, for example, 64×64×96. Since FFT does not change the spatial size and number of channels, but each channel becomes complex, it is represented as: the real part tensor Re is 64×64×96, and the imaginary part tensor Im is 64×64×96. First, the channels are grouped, such as dividing the signal into six groups. The 96 channels are divided into 6 groups, and each group contains 16 channels, 96÷6=16. After grouping, the real and imaginary parts are each split into 6 sub-tensors: Re1, Re2, ..., Re6, each with a size of 64×64×16; Im1, Im2, ..., Im6, each with a size of 64×64×16. If the channels are first expanded to 576 using a 1×1 convolution, each group will have 96 channels (576 ÷ 6 = 96), with the logic remaining identical. Then, synchronous processing is performed, using the same weight-shared convolutional layer to process both the real part Reg and the imaginary part Img for each group. The selected network module can be a standard 3×3 convolutional layer, with the output channel count remaining at 16 (maintaining dimensionality consistency). The convolutional kernel weights W, with a size of 3×3×16×16, are used to process both Reg and Img. The real and imaginary parts share the same set of convolutional kernel parameters, ensuring they are consistently transformed in the same feature space, maintaining the mathematical consistency of the complex structure. Finally, the complex features are reassembled, and the processed real and imaginary parts are reassembled to restore the complete complex frequency domain features. The output is still a pair of complex tensors.

[0028] In one implementation, the global spatial feature extraction module serves as branch 1: encoding input features to the SwinTransformer for processing into a spatial domain global feature map; the Fourier feature module serves as branch 2: encoding input features into a frequency domain / spatial information branch into a frequency domain enhanced fusion feature; subsequently, the outputs of the two branches, namely the spatial domain global feature map and the frequency domain enhanced fusion feature, are concatenated along the channel dimension to form a single-level fusion feature.

[0029] In one implementation, the spatial domain global feature map is generated as follows: encoded input features are used as input; processed by a multi-head self-attention mechanism of the Swin Transformer; long-distance spatial dependencies and global contextual information are captured; the output remains a three-dimensional tensor of H / 4×W / 4×96. This is a high-level semantic feature enhanced by the Transformer deep network, where the 96-dimensional vector at each location has fused information from the surrounding region and even the global context. The spatial domain global feature map is concatenated with the frequency domain enhanced fusion features output by the Fourier module to form the final fusion feature for subsequent decoding. In one embodiment, the attention mechanism determines which regions are spatially strongly correlated by calculating the correlation matrix between small pixel blocks. For example, in urban imagery, a main road may be connected to another main road in the distance via an intersection; this long-distance dependency can be captured by the attention mechanism. Simultaneously, the module analyzes the relationship between the road and its surrounding environment, such as whether the road passes through a park area or industrial area, thereby generating a feature map containing global contextual information. This feature map helps subsequent modules to more accurately locate the road.

[0030] The overall road structure information and spatial context information contained in the spatial domain global feature map can serve as prior knowledge, guiding the Fourier feature module to better separate the main road information and edge details in the frequency domain. For example, in frequency domain processing, spatial priors can help the module more accurately identify which frequency components correspond to the broad main part of the road and which frequency components correspond to the fine edges. The introduction of this prior information can significantly improve the targeting of frequency domain feature extraction. Furthermore, the inter-domain global feature map Fglobal is used as prior knowledge to dynamically modulate the frequency domain features in the Fourier feature module to guide it to better separate the main road information and edge details. Specifically, this guidance process is implemented through a channel attention guidance mechanism: first, global average pooling is performed on the inter-domain global feature map to generate a channel description vector. Where C is the number of feature channels. This vector z encodes the overall structure and contextual information of the road in the spatial domain. Next, vector z is input into a bottleneck structure consisting of two fully connected layers, with ReLU activation function in the middle layer and Sigmoid activation function in the output layer, thereby generating a normalized channel weight vector. The weight vector s and the complex frequency domain feature F freq The multiplication operation is performed on the real part Re(F) respectively. freq ) and the imaginary part Im(F freq The specific calculation formula is: the new real part Re(F) freq ') = s⊗ Re(F freq The new imaginary part Im(F) freq ') = s ⊗ Im(F freq ), where ⊗ represents element-wise multiplication along the channel dimension.

[0031] The encoded input features simultaneously enter both the global spatial feature extraction module and the Fourier feature module, sharing the same input source. Therefore, although spatial domain features and frequency domain features reside in different domains, they have a one-to-one correspondence in the channel dimension. Each channel represents the response of the same latent pattern or semantic information in the input image under different domains. The weight vector *s* reflects the importance of these latent patterns to the overall road structure in the spatial domain. Applying this importance measure directly to the corresponding channel of the frequency domain features adaptively enhances those patterns deemed important in the spatial domain, corresponding to low-frequency information of the road structure, while suppressing unimportant patterns, corresponding to background or noise.

[0032] Specifically, a Fourier feature module needs to be constructed first. The Fourier feature module includes a frequency domain information feature processing branch and a spatial information feature processing branch. Both the frequency domain information feature processing branch and the spatial information feature processing branch directly receive the encoded input features and perform a fast Fourier transform to divide the frequency domain signal into multiple groups of signals in the channel dimension. Through different operations, low-frequency main information and high-frequency edge information are extracted and enhanced to achieve multi-scale frequency domain feature fusion.

[0033] In one possible implementation, the Fourier feature module is designed to fully leverage the complementarity of frequency and spatial domain information. The frequency domain feature processing branch converts the coded input features into a frequency domain signal using a Fast Fourier Transform (FFT), then divides it into multiple signal groups along the channel dimension, processing low-frequency and high-frequency information separately. Low-frequency information typically corresponds to the main structure of the road, such as a wide main road or a long straight section, while high-frequency information corresponds to the edge details of the road, such as subtle changes at road boundaries or intersections. By employing different processing methods for different frequency information, multi-scale feature extraction and fusion can be achieved, thereby improving the accuracy of road segmentation.

[0034] Furthermore, a two-dimensional fast Fourier transform is performed on the encoded input features, i.e., the high-dimensional feature map output from the last layer of the encoder, to map it from the spatial domain to the frequency domain, obtaining complex frequency domain features. The frequency domain information feature processing branch divides the complex frequency domain features into six groups of signals. Three groups of signals are processed through average pooling layers with different window sizes to simulate low-pass filters and extract low-frequency main information, while the other three groups of signals are processed through convolutional layers with different kernel sizes to simulate high-pass filters and extract high-frequency edge information. Specifically, the total number of channels in the encoded input features is 576, which is uniformly divided into six groups of signals, with each group having 96 channels. For the processing of low-frequency main information, average pooling layers with window sizes of 2×2, 4×4, and 8×8 are used. This design can simulate the effect of a low-pass filter and extract the smooth information of the main road structure. For the processing of high-frequency edge information, convolutional layers with kernel sizes of 3×3, 5×5, and 7×7 are used to simulate the effect of a high-pass filter and capture the detailed changes in the road edges. This layered processing method can ensure the multi-scale characteristics of the frequency domain features.

[0035] To ensure that the subsequent inverse Fourier transform can accurately reconstruct the spatial domain features, this invention imposes strict size constraints on all convolution and pooling operations that act on the frequency domain features. All convolutional layers use 1×1 convolution kernels, and a padding strategy is set to make the spatial size (H×W) of the output feature map completely consistent with the input. In one embodiment, to maintain the integrity of the feature representation within each subgroup and avoid excessive information compression, a 1×1 convolutional layer is applied to each group of 96-channel signals, with the output channel count also configured to 96. This design ensures that the information capacity of each subgroup feature is preserved after nonlinear transformation; subsequently, when concatenating the six groups of signals along the channel dimension, a complete frequency domain feature map of the original 576 channels can be seamlessly reconstructed, thus ensuring the structural consistency of the inverse Fourier transform input. It should be noted that, in this invention, all convolutional operations performed after frequency domain feature grouping, unless specifically specified for channel fusion or dimensionality reduction, are designed with the default principle of the same channel number transformation, i.e., the input and output channel numbers remain consistent. This principle applies to 96-channel grouping processing and similar grouping processing scenarios that may occur subsequently, to ensure dimensional alignment and lossless information transfer during feature separation and merging.

[0036] The extraction process for low-frequency subject information can be adjusted according to different road types. For example, in remote sensing imagery of urban areas, main roads typically have greater width and smoother structures, so larger window average pooling layers, such as 8×8 windows, can be prioritized to smooth noise and highlight the main road subject. In contrast, in imagery of rural areas, roads are narrower and have more complex structures, so smaller window average pooling layers, such as 2×2 windows, can be relied upon more to retain more local information. This scene-specific approach enhances the adaptability of low-frequency information extraction.

[0037] The extraction process for high-frequency edge information can also be optimized based on the application scenario. In one possible implementation, for urban imagery with relatively regular road edges, a smaller convolutional kernel, such as a 3×3 kernel, can be used to capture subtle edge variations. Conversely, for rural imagery with irregular road edges and significant noise, a larger convolutional kernel, such as a 7×7 kernel, can be used to smooth the noise and extract more significant edge features. In this way, the extraction of high-frequency information can be better adapted to different road characteristics.

[0038] Specifically, after low-frequency and high-frequency processing, the six signals are concatenated along the channel dimension to form a fused signal containing multi-scale frequency information. Since all the aforementioned processing operations do not change the spatial dimensions of the feature map, the spatial dimension (H×W) of the concatenated complex frequency domain feature map is completely consistent with the initial frequency domain feature obtained by FFT transformation, satisfying the input requirements of inverse Fourier transform. Subsequently, the fused signal is converted from the frequency domain back to the spatial domain through two-dimensional inverse Fourier transform to generate enhanced frequency domain features. This feature not only preserves the smooth information of the main road structure but also contains detailed information of the road edges, providing an important supplement for subsequent spatial domain feature fusion.

[0039] In another implementation, three sets of signals are processed through an average pooling layer to obtain a low-frequency main information group. Another three sets of signals are processed through a convolutional layer to obtain a high-frequency edge information group. The generation of the low-frequency main information group can be achieved through layer-by-layer pooling. For example, the first set of signals is processed through a 2×2 window average pooling layer to extract finer-grained low-frequency information; the second set of signals is processed through a 4×4 window average pooling layer to extract medium-scale low-frequency information; and the third set of signals is processed through an 8×8 window average pooling layer to extract large-scale low-frequency information. This hierarchical extraction method ensures that the low-frequency main information group covers road structures at different scales. The generation of the high-frequency edge information group also employs a hierarchical design. For example, the first set of signals is processed through a 3×3 convolutional kernel to capture subtle changes in road edges; the second set of signals is processed through a 5×5 convolutional kernel to extract medium-scale edge features; and the third set of signals is processed through a 7×7 convolutional kernel to highlight edge information over a larger area. In this way, the high-frequency edge information group can comprehensively characterize the detailed features of road edges. The low-frequency main information group and the high-frequency edge information group are concatenated along the channel dimension to form a fused frequency domain signal. In one possible implementation, the concatenation process ensures a balance in the number of channels for low-frequency and high-frequency information to prevent one type of information from dominating the fused signal. For example, the low-frequency main information group contains 288 channels, and the high-frequency edge information group also contains 288 channels, which are then concatenated to form a fused frequency domain signal with 576 channels. This balanced design ensures that the contributions of the two types of information are roughly equal in subsequent processing. An inverse Fourier transform is performed on the fused frequency domain signal to obtain enhanced frequency domain features. Specifically, the inverse Fourier transform process converts the fused frequency domain signal from the frequency domain back to the spatial domain, forming an enhanced frequency domain feature map containing multi-scale features. This feature map contains both the smooth structural information of the road body and the detailed information of the road edges, providing important supplementary information for subsequent fusion with spatial domain features. Then, a pre-defined convolutional layer is used to process the input feature map, and the result is added to the original input residual to compensate for local spatial details. Specifically, the input feature map first passes through a 3×3 convolutional layer to extract local spatial features, with a stride of 1, zero-padding at the edges, and the number of channels remaining constant at 96. The processed result is then added to the original input feature map as a residual. This residual connection effectively compensates for local details that may be lost during global feature extraction, generating spatial domain detail-enhanced features. Next, the spatial domain detail-enhanced features and the enhanced frequency domain features are concatenated along the channel dimension, and the final result of the Fourier feature module is output through a fusion convolution, used for parallel processing in subsequent encoding stages. Before concatenation, alignment is performed on the two feature paths. Channel alignment uses a 1×1 convolutional layer to unify the two features to 96 channels; size alignment uses bilinear interpolation to adjust the feature map size to be consistent. After concatenation along the channel dimension, the spatial domain detail-enhanced features and the enhanced frequency domain features form a comprehensive feature map containing both spatial and frequency domain information.Subsequently, the concatenated features are further integrated through a 1×1 convolutional layer to generate the final output of the Fourier feature module. This result preserves local details in the spatial domain and integrates multi-scale features in the frequency domain, providing rich feature information for parallel processing in the subsequent encoding stage.

[0040] The spatial information feature processing branch performs convolution on the input feature map and adds it to the original input residual to obtain spatial domain detail enhancement features. These features are then concatenated with the output of the frequency domain information feature processing branch along the channel dimension, fusing spatial and frequency domain features. In one possible implementation, the spatial information feature processing branch extracts local spatial information from the input feature map through convolution, such as texture features of road edges or details of small road structures. This information is then added to the original input via residual concatenation to form spatial domain detail enhancement features. This feature is then concatenated with the enhanced frequency domain features output by the frequency domain information feature processing branch along the channel dimension, achieving the fusion of spatial and frequency domain features. This fusion method fully utilizes the complementarity of the two types of features, improving the accuracy of road segmentation.

[0041] The process involves concatenating the outputs of spatial domain detail enhancement features and frequency domain information feature processing branches along the channel dimension. This concatenation operation combines multi-scale frequency domain features with spatial compensation information. The concatenated result is then used as the output of the Fourier feature module and passed to the encoding stage. Specifically, the spatial domain detail enhancement features contain local texture and detail information of the road, while the output of the frequency domain information feature processing branch contains multi-scale frequency information of the road body and edges. By concatenating along the channel dimension, these two types of features can be integrated into a unified feature map, providing comprehensive information support for subsequent processing. The concatenation operation not only achieves simple feature stacking but also further fuses the two types of features through subsequent convolutional processing. For example, a set of 1×1 convolutional layers can be used to perform inter-channel interaction processing on the concatenated feature map, allowing multi-scale frequency domain features and spatial compensation information to be better combined, forming a more consistent feature representation. After fusion processing, the concatenated feature map serves as the final output of the Fourier feature module and is passed to the subsequent encoding stage. This output contains both local details in the spatial domain and multi-scale features in the frequency domain, providing important support for feature extraction in the encoding stage.

[0042] In the encoding stage, Fourier feature modules and global spatial feature extraction modules are embedded and processed in parallel. The fused features are then passed to the decoding stage via skip connections, progressively restoring image resolution and outputting road segmentation results. Specifically, the encoding stage is designed to gradually compress the spatial resolution of the input image through multi-level feature extraction while preserving key information about the road structure. The Fourier feature module and the global spatial feature extraction module work in parallel during the encoding stage; the former is responsible for fusing spatial and frequency domain features, while the latter focuses on capturing long-range dependencies and global contextual information. Through a skip connection mechanism, the features extracted in the encoding stage are passed to the decoding stage to progressively restore image resolution and ultimately generate road segmentation results. This parallel processing and feature passing method ensures that detailed road information and overall structural information are fully preserved during resolution restoration.

[0043] Furthermore, Fourier feature modules are embedded in the first two layers of the encoding stage, while the last two layers only use the global spatial feature extraction module, processing the feature maps through stride convolution downsampling. The number of feature channels in the four layers of the encoding stage are 96, 192, 384, and 768 respectively. Each layer performs downsampling through a 3×3 convolutional layer with a stride of 2, halving the feature map size and doubling the number of channels. In the first two layers, Fourier feature modules are embedded to fuse spatial and frequency domain features, extracting the main structure and edge details of the road. The last two layers rely solely on the global spatial feature extraction module, focusing on capturing deeper global contextual information, such as the overall layout of the road network or the relationship between the road and its surrounding environment. Stride convolution is used for downsampling; for example, a stride of 2 convolution can halve the spatial size of the feature map while increasing the number of feature channels to retain more semantic information.

[0044] In one embodiment, the design of the first two layers of the encoding stage, which embed Fourier feature modules, can be adjusted according to the characteristics of the remote sensing image. For example, when processing high-resolution images of urban areas, where road structures are complex and edge details are rich, the processing depth of the Fourier feature modules in the first two layers can be increased to capture subtle changes in road edges through multi-scale frequency domain feature extraction. Conversely, when processing images of rural areas, where road structures are simpler, the processing complexity of the Fourier feature modules can be appropriately reduced, relying more on the global spatial feature extraction module to highlight the overall direction of the road. This flexible design can adapt to the road extraction needs of different scenarios. For the implementation of the last two layers using only the global spatial feature extraction module, the feature extraction effect can be optimized by adjusting the downsampling step size. For example, in urban images, where there are many road intersections and branches, a smaller downsampling step size, such as 1, can be used to retain more spatial detail information. In rural images, where road structures are relatively simple, a larger downsampling step size, such as 2, can be used to quickly compress spatial resolution and reduce computation. In this way, the feature extraction in the last two layers can better focus on the extraction of global semantic information. The downsampling process is repeated twice during the encoding stage, and spatial and frequency domain features are extracted in parallel in the first two layers. These features are then passed to the corresponding layers in the decoding stage via skip connections, achieving step-by-step feature fusion. Specifically, the downsampling process in the encoding stage is typically repeated multiple times; for example, after two downsampling operations, the spatial resolution of the feature map is compressed to one-quarter of its original size. The parallel extraction of spatial and frequency domain features in the first two layers ensures that local details and multi-scale frequency information of the road are fully captured. Subsequently, through a skip connection mechanism, these features are directly passed to the corresponding layers in the decoding stage, where they are fused with the upsampling features from the decoding process. This step-by-step fusion approach gradually integrates shallow detail information and deep semantic information during resolution restoration.

[0045] In the decoder upsampling process, an optimized skip connection mechanism is introduced to fuse intermediate features from different levels in the encoding stage with the upsampled features from the corresponding levels in the decoder. To adapt to the characteristics of spatial domain branching and frequency domain branching structures, the skip connection design is differentiated and optimized according to the feature level. For shallow, high-resolution, skip connections: prioritize preserving spatial domain detail enhancement features.

[0046] The specific implementation involves selecting only the corresponding level of encoded features from the global spatial feature extraction module as the skip signal, while masking the output of the Fourier feature module. Since shallow encoded features are rich in high-frequency spatial details such as edges and textures, and the spatial domain branches further enhance this information through attention or dilated convolution, directly injecting them into the decoder helps to recover fine boundary structures.

[0047] For deep, low-resolution, skip connections, more attention is paid to the low-frequency main information in the frequency domain features. Specifically, this is achieved by explicitly preserving low-frequency components, such as those in the central region, through a frequency domain mask in the corresponding level output of the Fourier feature module. In the spectral region, high-frequency noise is suppressed, and the complex frequency domain features dominated by low frequency are transformed back to the spatial domain by IFFT and used as skip signals for fusion. At the same time, the skip connections of spatial domain branches at this level can be selectively weakened or discarded to avoid interference from redundant or conflicting semantic information.

[0048] The above strategy ensures that the decoder obtains the most suitable contextual guidance at different scales, with shallow focus on geometric detail reconstruction and deep focus on semantic subject localization.

[0049] In one possible implementation, the design of skip connections can be optimized based on feature hierarchy. For example, when passing the first layer of encoded features to the decoding stage, spatial domain detail enhancement features can be prioritized to ensure that subtle information about road edges is not lost. When passing the second layer of encoded features, more attention can be paid to low-frequency main information in the frequency domain features to highlight the overall structure of the road. Through this hierarchical transmission method, skip connections can achieve more effective feature fusion during the decoding stage.

[0050] The repetition of the downsampling process can be adjusted based on the image resolution. For example, when processing ultra-high resolution remote sensing imagery, such as images with 0.3 meters per pixel, the number of downsampling iterations can be increased to three to further compress the spatial resolution and reduce computational complexity. Conversely, when processing medium-resolution imagery, such as images with 1 meter per pixel, two downsampling iterations can be maintained to avoid over-compression and loss of detail. This flexible downsampling strategy can adapt to the processing needs of imagery at different resolutions.

[0051] In the decoding stage, features are upsampled and fused layer by layer using transposed convolutions. Features from the corresponding layers in the encoding stage are obtained through skip connections and fused with the upsampled features. The feature integration module performs multi-level semantic integration on the fused features to restore image resolution. Based on the integration result, the road segmentation result is output through a segmentation layer. Specifically, the goal of the decoding stage is to gradually restore the spatial resolution of the feature maps through upsampling operations, ultimately achieving the same size as the original image. Transposed convolutions are used for upsampling, integrating shallow detail information and deep semantic information through fusion with features from the corresponding layers in the encoding stage. Subsequently, the feature integration module integrates multi-level features to generate the final feature representation, which is used to output the road segmentation result. This step-by-step restoration method ensures that the details and overall layout of the road structure are reflected in the final result.

[0052] Furthermore, the number of feature channels in the four layers of the decoding stage corresponds to 768, 384, 192, and 96 in the encoding stage. Upsampling is performed using a 2×2 transposed convolution with a stride of 2, doubling the feature map size and halving the number of channels. After each layer of upsampling, it is fused with the features of the corresponding layer in the encoding stage through skip connections. For example, the first layer of decoded features is fused with the features of the last layer in the encoding stage, and the second layer of decoded features is fused with the features of the penultimate layer in the encoding stage. This layer-by-layer fusion method ensures that the detailed information of the road is gradually restored during the upsampling process. First, bilinear interpolation and a 1×1 convolution are performed on the encoded features to align the resolution and number of channels. Then, the features of the corresponding layer in the encoding stage are concatenated with the upsampled features along the channel dimension, and subsequently integrated through a 3×3 convolutional layer to form a new feature representation. This fusion method effectively combines shallow details and deep semantic information. The aggregated feature integration module sequentially includes a 1×1 convolutional layer, a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation function. A 1×1 convolutional layer reduces the number of channels to 96, and a 3×3 convolutional layer has a stride of 1 and uses zero-padding at the edges. Through multiple convolutional operations, the fused features are further processed, integrating semantic information from different levels to ultimately generate a feature map close to the original resolution. This multi-level semantic integration method ensures the integrity of the road structure in the final feature map. The segmentation layer adjusts the number of channels in the feature map to 2 using a 1×1 convolutional layer, outputs a road probability prediction map via Softma×, performs binarization using a fixed threshold of 0.5, and then eliminates noise and enhances connectivity through morphological dilation and erosion once each using 3×3 structuring elements, finally outputting the binarized road extraction result. This method can clearly identify road regions in the image.

[0053] The feature integration module receives upsampled features and fuses shallow frequency domain details with deep global semantics, outputting a binarized road extraction result. In one embodiment, the feature integration module is designed to integrate feature information from different levels during the decoding stage. Shallow frequency domain details typically contain subtle changes in road edges, while deep global semantics contain the overall road structure and contextual relationships. Through multi-layer convolution and feature fusion operations, the module generates a comprehensive feature map, which is then processed by a segmentation layer to output a binarized road extraction result. This fusion method ensures that both the details and overall structure of the road are accurately represented in the final result. In one possible implementation, the processing of shallow frequency domain details by the feature integration module can be adjusted according to the complexity of the road features. For example, in urban area images, where road edge details are abundant, the weight of shallow frequency domain features can be increased, and multi-layer convolution operations can be used to enhance the expression of edge information. In rural area images, where road edges are simpler, the focus on shallow features can be appropriately reduced, relying more on deep global semantic information to highlight the overall road layout. Through this adjustment method, the module can adapt to the road extraction needs of different scenarios. To optimize the fusion of deep global semantic information, the module can adjust the convolutional layer structure. For example, when processing urban imagery, convolutional layers with larger receptive fields, such as 5×5 layers, can be used to capture the complex layout of road networks. Conversely, when processing rural imagery, convolutional layers with smaller receptive fields, such as 3×3 layers, can be used to reduce computation and focus on simpler road structures. This flexible design approach enhances the fusion of deep semantic information. The binarized road extraction results output by the feature integration module can be further optimized through post-processing steps. For instance, morphological operations such as dilation and erosion can be performed on the segmentation results to eliminate small-scale noise points and enhance the connectivity of road areas. Alternatively, road width constraints can be incorporated to refine the segmentation results, retaining only areas that conform to the actual road width range. This post-processing method further improves the accuracy of the road extraction results.

[0054] In one embodiment, the output format of the binarized road extraction results can be adjusted according to application requirements. For example, in urban planning applications, the results can be output as a vectorized road network map for subsequent spatial analysis. In road monitoring applications, the results can be output as a rasterized binary map for direct identification of damaged road areas. Through this flexible output method, the road extraction results can better serve different application scenarios.

[0055] The road segmentation results can also be optimized for different application scenarios. In urban scenarios, the details of road edges and intersections are enhanced; in rural scenarios, the overall connectivity of roads is strengthened; and in mountainous scenarios, the road segmentation range is constrained by terrain features to improve the extraction accuracy in different scenarios.

[0056] Morphological optimization of the road segmentation results can also be performed by using 3×3 structural elements to sequentially perform expansion and erosion processing, eliminating noise and ensuring the continuity of the road area.

[0057] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

Claims

1. A high-resolution remote sensing road extraction method based on spatial frequency domain features, characterized in that, include: High-resolution remote sensing images are acquired, divided into preset pixel blocks through block embedding, and linearly mapped to obtain coded input features; The encoded input features are synchronously input into the global spatial feature extraction module and the Fourier feature module, and the spatial domain feature extraction and frequency domain feature enhancement processing are performed in parallel. The global spatial feature extraction module is built based on the Swing Transformer and obtains a spatial global feature map by capturing long-distance road dependencies and global context information. The Fourier feature module contains frequency domain and spatial information feature processing branches that are parallel and share encoded input features. The frequency domain information feature processing branch performs a two-dimensional fast Fourier transform on the encoded input features, divides the frequency domain signal into multiple groups according to the channel, extracts and enhances the low-frequency main body and high-frequency edge information, and obtains the enhanced frequency domain features through inverse Fourier transform; The spatial information feature processing branch convolves the encoded input features and adds them to the original feature residual to obtain spatial domain detail enhancement features; The spatial domain detail enhancement features and the enhanced frequency domain feature channels are concatenated and then fused through convolutional layer channels to obtain the frequency domain enhanced fused features; The spatial domain global feature map is concatenated with the frequency domain enhanced fusion feature channel to obtain a single-level fusion feature in the coding stage. In the encoding stage, corresponding modules are set up layer by layer to complete multi-level feature extraction. The encoded outputs of each level are transmitted to the decoding stage through skip connections to complete multi-level feature fusion. In the decoding stage, the feature map resolution is restored level by level. The road probability prediction map is output through the segmentation layer and obtained by binarization.

2. The method according to claim 1, characterized in that, The process of acquiring high-resolution remote sensing images, dividing them into preset pixel blocks through block embedding and linear mapping to obtain coded input features includes: Acquire high-resolution remote sensing images, and divide the high-resolution remote sensing images into image blocks of preset pixels as input; The image block is divided into pixel blocks of a preset pixel size by block embedding, and linear mapping is completed. After mapping the pixel dimension of the pixel block to the preset embedding dimension, the encoded input feature is obtained. The encoded input features are synchronously input into the global spatial feature extraction module and the Fourier feature module. The spatial domain global feature map output by the global spatial feature extraction module provides a spatial prior for the overall road structure for subsequent frequency domain feature processing.

3. The method according to claim 1, characterized in that, The frequency domain information feature processing branch performs a two-dimensional fast Fourier transform on the encoded input features, divides the frequency domain signal into multiple groups according to channels, extracts and enhances the low-frequency main body and high-frequency edge information, and obtains enhanced frequency domain features through inverse Fourier transform, including: Perform a two-dimensional fast Fourier transform on the encoded input features, and divide the resulting complex frequency domain features into a preset number of signal groups in the channel dimension; The signals of a predetermined number of groups are divided into a low-frequency processing group and a high-frequency processing group. The two groups of grouped signals, which are divided based on the complete frequency domain features of the same two-dimensional fast Fourier transform output, are processed synchronously and in parallel. The low-frequency processing group contains half of the signals of the predetermined number of groups, and each group is processed by simulating a low-pass filter through an average pooling layer with a different window size to extract the low-frequency main information. The high-frequency processing group contains the remaining half of the signals of the predetermined number of groups, and each group is processed by simulating a high-pass filter through a convolutional layer with a different kernel size to extract the high-frequency edge information. The signals of a preset number of groups corresponding to the low-frequency main information and the high-frequency edge information are spliced ​​together in the channel dimension to obtain a fused frequency domain signal after integrating the low-frequency fusion feature and the high-frequency fusion feature. An inverse Fourier transform is performed on the fused frequency domain signal to obtain enhanced frequency domain features.

4. The method according to claim 1, characterized in that, The process of concatenating spatial domain detail enhancement features with enhanced frequency domain feature channels and fusing them through convolutional layer channels to obtain frequency domain enhanced fused features includes: After aligning the number of channels and the size of the spatial domain detail enhancement features and the enhanced frequency domain features, the channel dimensions are then stitched together. After the splicing is completed, channel fusion is performed through a convolutional layer to obtain frequency domain enhanced fusion features.

5. The method according to claim 1, characterized in that, The encoding stage sets up corresponding modules layer by layer to complete multi-level feature extraction, including: The encoding stage is set up with a four-layer encoding structure arranged sequentially from input to output. The first two layers are configured with parallel global spatial feature extraction modules and Fourier feature modules to output single-layer fused features. The last two layers are configured with only global spatial feature extraction modules to output single-layer spatial features. The two types of features are unified as encoded output features. The output of the previous level coding structure is used as the input of the next level coding structure. The single-level coding structure completes feature extraction and fusion to obtain single-level fused features. Then, the features are downsampled by a convolutional layer with a preset stride, and the downsampled features are input into the next level coding structure. The features passed to the decoding stage through skip connections are the output features of each level in the encoding stage, which correspond one-to-one with the upsampling level resolution in the decoding stage.

6. The method according to claim 1, characterized in that, The decoding stage restores the feature map resolution step by step, outputs a road probability prediction map through the segmentation layer, and obtains the road segmentation result after binarization, including: The decoding stage adopts a four-layer decoding structure with the same number of layers as the encoding stage and progressively corresponding resolutions. Each layer of the decoding structure performs upsampling on the input features through transposed convolution. After upsampling is completed, the encoded output features of the corresponding level in the encoding stage are obtained through skip connections. The encoded output features are then concatenated with the upsampled features of the current level in the channel dimension to obtain the decoded fusion features. The decoded and fused features are input into a preset aggregated feature integration module. The aggregated feature integration module performs multi-level semantic information integration on the decoded and fused features to complete multi-level feature fusion. The integrated features are output and input into the next level of decoding structure. After the decoding stage completes upsampling and feature integration at all levels, the road probability prediction map is output through the segmentation layer, and then binarized post-processing is performed to obtain the binarized road segmentation result.

7. The method according to claim 2, characterized in that, The step of dividing the image block into pixel blocks of a preset pixel size through block embedding, completing linear mapping, and mapping the pixel dimension of the pixel block to a preset embedding dimension, yields the encoded input features, including: The image block is divided using a pixel size smaller than the preset pixel to generate multiple non-overlapping pixel blocks; Perform a linear mapping on each pixel block to map the pixel dimension of the pixel block to a preset embedding dimension, thereby obtaining the encoded input features.

8. The method according to claim 3, characterized in that, The step of concatenating the six sets of signals corresponding to the low-frequency main information and the high-frequency edge information along the channel dimension to obtain the fused frequency domain signal after integrating the low-frequency fusion features and the high-frequency fusion features includes: The signals output from the average pooling layer of half of the preset number of low-frequency processing groups are spliced ​​together along the channel dimension to obtain low-frequency fusion features. The signals output from the remaining half of the pre-set number of convolutional layers in the high-frequency processing group are concatenated along the channel dimension to obtain high-frequency fusion features. The low-frequency fusion feature and the high-frequency fusion feature are concatenated along the channel dimension to obtain a fused frequency domain signal.

Images