Remote sensing road image segmentation method and device based on geometric reasoning guidance, and equipment

Abstract

The application provides a remote sensing road image segmentation method and device based on geometric reasoning guidance and equipment, and relates to the technical field of remote sensing image segmentation. The application obtains a remote sensing road image, inputs the remote sensing road image into an encoder for spatial feature extraction and modulation to obtain multi-scale features after feature modulation; performs reasoning on the road geometry of the multi-scale features after feature modulation to generate a continuous geometric prior field, outputs reasoning fusion feature maps; performs dynamic trajectory sampling and aggregation on the reasoning fusion feature maps based on the geometric prior field, and outputs long-range connected aggregation features; and decodes and reconstructs the aggregation features and shallow features in the multi-scale features to output a binary road segmentation mask, that is, a road image segmentation result. The application can solve the technical defects of road topological fracture, edge blur adhesion and dynamic sampling drift in the existing remote sensing road segmentation technology.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image segmentation and processing technology, and more specifically, to a method, apparatus, and device for remote sensing road image segmentation based on geometric reasoning guidance. Background Technology

[0002] Road extraction and segmentation from remote sensing imagery is a crucial foundational task for applications such as Geographic Information Systems (GIS), intelligent transportation planning, and emergency response. With the widespread availability of high-resolution remote sensing data, deep learning-based road extraction and segmentation methods have made significant progress. Current mainstream technologies fall into three main categories: first, classic encoding / decoding structures based on fully convolutional networks (FCNs), such as the U-Net and DeepLab series, which extract local texture and edge information layer by layer using convolutional kernels with fixed geometric shapes; second, Transformer architectures incorporating self-attention mechanisms, such as Swin-UNet and SegFormer, which aim to enhance the perception of road connectivity by modeling long-range dependencies between pixels; and third, implicit dynamic sampling strategies such as deformable convolutions (DCNs) or serpentine convolutions, attempting to adaptively fit the receptive field to the winding shape of the road. While these methods perform well in conventional scenarios, they still face several insurmountable technical bottlenecks when processing remotely sensed roads with elongated topologies, complex background interference, and severe occlusion.

[0003] Specifically, traditional FCN methods are limited by the local receptive field of fixed convolutional kernels, making it difficult to establish effective contextual relationships across occluded areas such as trees and building shadows. This results in frequent topological breaks in the extraction results at occluded locations, failing to maintain road continuity. While Transformer-type methods can expand the receptive field, their isotropic global attention mechanism easily introduces a large amount of irrelevant background noise (such as farmland and rooftops) when processing extremely narrow roads, causing a decrease in feature contrast and leading to edge blurring and road adhesion problems. At the same time, their high computational cost makes it difficult to meet the efficient processing requirements of remote sensing big data. While implicit dynamic sampling-based methods improve morphological adaptability, their offsets are generated by black-box regression of the network, lacking explicit geometric and physical constraints. In scenarios with complex background textures (such as field ridges and rooftops that resemble roads), semantic drift is prone to occur, meaning that the sampling points deviate from the real road manifold, leading to unstable model training, especially in accurately capturing narrow rural roads.

[0004] In view of the above, this application is hereby submitted. Summary of the Invention

[0005] The present invention aims to provide a method, apparatus, device and medium for remote sensing road image segmentation based on geometric reasoning, so as to solve the technical defects of existing remote sensing road segmentation technology, such as road topological breakage, edge blurring and adhesion, and dynamic sampling drift.

[0006] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:

[0007] A remote sensing road image segmentation method based on geometric reasoning, comprising:

[0008] S1, acquire remote sensing road images;

[0009] S2, input the remote sensing road image into the encoder for spatial feature extraction and modulation to obtain multi-scale features after feature modulation;

[0010] S3 performs road geometry reasoning on the multi-scale features after feature modulation, generates a continuous geometric prior field, and outputs a reasoning fusion feature map.

[0011] S4, Based on the geometric prior field, perform dynamic trajectory sampling and aggregation on the inference fusion feature map, and output long-range connected aggregated features;

[0012] S5, decode and reconstruct the aggregated features and the shallow features in the multi-scale features, and output a binary road segmentation mask, i.e., the road image segmentation result.

[0013] Preferably, the process of inputting the remote sensing road image into the encoder for spatial feature extraction and outputting multi-scale features, followed by modulation, is as follows:

[0014] The deepest feature containing global semantic information in the multi-scale features is used to generate affine transformation parameters, as follows:

[0015] ;

[0016] in, This is a scaling factor used to enhance or suppress the characteristic response; This is the bias coefficient, used to adjust the baseline of the characteristic distribution; The deepest feature; This is a convolution operation; This is a channel-dimensional splitting operation;

[0017] Affine transformation parameters Upsampled to the same spatial size as the mid-level features in the multi-scale features by bilinear interpolation;

[0018] Based on the upsampled affine transformation parameters Spatial feature linear modulation is performed on the mid-level features that contain rich details but are noisy, and the modulated mid-level features are output as follows:

[0019] ;

[0020] in, These are the modulated mid-layer features; This is a mid-level feature; This indicates element-wise multiplication; To adjust the hyperparameters of intensity; It is the hyperbolic tangent activation function;

[0021] The shallow features in the multi-scale features are retained.

[0022] Preferably, the road geometry is inferred from the multi-scale features after feature modulation to generate a continuous geometric prior field, specifically:

[0023] The modulated mid-layer features are fused through convolutional layers. Output the initial orientation field and initial width field to obtain the initial fused feature map;

[0024] For each center pixel in the initial fused feature map Take one Local spatial neighborhood ;

[0025] For each center pixel in the initial fused feature map Extract the direction vector and local spatial neighborhood Each neighboring pixel Direction vector Perform a dot product operation to obtain a similarity scalar. The formula is:

[0026] ;

[0027] in, The direction vector representing the center pixel Direction vectors of neighboring pixels The angle between them;

[0028] For the similarity scalar Application with temperature coefficient Softmax generates normalized diffusion weights. The formula is:

[0029] ;

[0030] in, Represents an exponential function. , All are local spatial neighborhoods The neighboring pixels in;

[0031] Based on the initial width field and the initial direction field, using The initial features within the local spatial neighborhood are aggregated using a targeted weighted method, as shown in the formula:

[0032] ;

[0033] ;

[0034] in, The width feature after aggregation; The directional features after aggregation; This is the initial width field; This is the initial direction field;

[0035] The aggregated smooth features, i.e., the smooth width features With smooth direction features The initial orientation field and initial width field in the initial fused feature map are residually fused, and the orientation vector field after residual fusion is normalized to output the inference fused feature map and the final geometric prior field, i.e., the orientation field. Width field .

[0036] Preferably, the geometric prior field performs dynamic trajectory sampling and aggregation on the inference fusion feature map, specifically as follows:

[0037] Construct a base coordinate grid with the same size as the inference fusion feature map, and assign coordinates to the position of each pixel in the inference fusion feature map. Mapping to a normalized coordinate system yields the starting center point for trajectory generation. ;

[0038] The offset vector of the sampling step size is explicitly calculated based on the geometric prior field. A deterministic mapping between sampling geometry and road physical properties is established, using the following formula:

[0039] ;

[0040] in, The base step size is used to define the minimum physical interval between samples; For pixels The road width scalar at the location; This is a width adjustment factor used to control the degree to which the path width affects the sampling step size; For pixels The normalized direction vector at that point;

[0041] Then, using the center point Using the offset vector as a reference, Extending forward and backward, a set of linear trajectories is generated. The formula is:

[0042] ;

[0043] in, For pixels The corresponding set of linear trajectories; This is the kth sampling point; This is a pre-defined multi-scale scaling sequence;

[0044] Based on a fundamental coordinate grid, using a set of linear trajectories Sampling points in Bilinear interpolation sampling is performed on the inference fusion feature map to extract features from each sampling point, and the sampled features are output using the following formula:

[0045] ;

[0046] in, For the k-th sampling feature; The inference fusion feature map; It is a feature sampling function based on the basic coordinate grid;

[0047] The features and orientation field of the inference fusion feature map are combined. Width field Concatenate along the channel dimension, input into a lightweight convolutional network, and then... Normalization yields the confidence weight for each sampling point, using the following formula:

[0048] ;

[0049] in, The confidence weight for the k-th sampling point; This is a convolution operation; This is a normalization operation;

[0050] Finally, convolution optimization is performed on each group of sampled features, and then the features are weighted and summed using attention weights to obtain the final aggregated features. The formula is:

[0051] ;

[0052] in, This represents the total number of sampled features.

[0053] Preferably, the decoding and reconstruction process is as follows:

[0054] The aggregated features are spatially aligned and fused with the shallow features in the multi-scale features to generate a global channel descriptor vector.

[0055] The global channel descriptor vector is input into the MLP multilayer perceptron, and complementary channel attention vectors are generated through Softmax normalization.

[0056] The global channel descriptor vector is weighted and recombined based on the channel attention vector to output attention fusion features;

[0057] Based on the attention fusion features, horizontal and vertical bar pooling kernels are used for pooling to aggregate the global context and obtain horizontal flow features and vertical flow features.

[0058] The horizontal flow features and the vertical flow features are superimposed, and a topological attention weight map is generated using a sigmoid function. This map is then applied to the attention fusion features to repair road breaks caused by occlusion, resulting in an attention-weighted enhanced fusion feature, expressed as:

[0059] ;

[0060] ;

[0061] in, This is a topological attention weight graph; , These refer to the horizontal flow characteristics and the vertical flow characteristics, respectively. Attention-weighted enhancement of fusion features; For attention fusion features;

[0062] By applying the difference operator to the shallow features in the multi-scale features, high-frequency edge information is extracted to obtain the feature gradient map;

[0063] The feature gradient map is mapped to an edge attention map using a 1×1 convolution;

[0064] The edge attention map is superimposed on the attention-weighted enhanced fusion feature to sharpen the boundary, resulting in an explicit boundary enhanced fusion feature, expressed as:

[0065] ;

[0066] in, Explicitly enhance fusion features at the boundary; For edge attention maps;

[0067] By cascading upsampling, the explicit enhancement fusion features of the boundary are gradually restored to the original image resolution, and then a road segmentation probability map is output through 1×1 convolution.

[0068] The road segmentation probability map is processed by the Sigmoid function to obtain a binary road segmentation mask, i.e., the road segmentation result.

[0069] Preferably, the operation of spatially aligning the aggregated features with the shallow features in the multi-scale features specifically involves:

[0070] The shallow features and the aggregated features are mapped to the same channel dimension by 1×1 convolution, and the aggregated features are upsampled to the resolution of the shallow features by bilinear interpolation to obtain the aligned feature streams corresponding to the shallow features and the aggregated features, respectively.

[0071] The shallow features and the aligned feature streams corresponding to the aggregated features are added and fused together, and then a one-dimensional global channel descriptor vector is generated by global average pooling, as expressed in the following expression:

[0072] ;

[0073] in, A global channel descriptor vector; This is the global average pooling function; , These are the shallow features and the aligned feature streams corresponding to the aggregated features, respectively.

[0074] The present invention also provides a remote sensing road image segmentation device based on geometric reasoning guidance, comprising:

[0075] The acquisition unit is used to acquire remotely sensed road images;

[0076] The feature extraction and modulation unit is used to input the remote sensing road image into the encoder for spatial feature extraction and modulation to obtain multi-scale features after feature modulation.

[0077] The geometric inference unit is used to infer the road geometry from the multi-scale features after feature modulation, generate a continuous geometric prior field, and output an inference fusion feature map.

[0078] The sampling and aggregation unit is used to perform dynamic trajectory sampling and adaptive aggregation on the inference fusion feature map based on the geometric prior field, and output long-range connected aggregated features;

[0079] The decoding and reconstruction unit is used to decode and reconstruct the aggregated features and the shallow features in the multi-scale features, and output a binary road segmentation mask, i.e., the road image segmentation result.

[0080] The present invention also provides a remote sensing road image segmentation device based on geometric reasoning, including a processor and a memory. The memory stores a computer program that can be executed by the processor to implement the remote sensing road image segmentation method based on geometric reasoning as described above.

[0081] The present invention also provides a computer-readable storage medium storing computer-readable instructions, which, when executed by a processor of the device on which the computer-readable storage medium resides, implement the remote sensing road image segmentation method based on geometric reasoning as described above.

[0082] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0083] This invention introduces spatial feature linear modulation in the encoder stage to process multi-scale features, and uses the global semantic information contained in deep features to generate spatial affine parameters to perform gated modulation on mid-level features. This is equivalent to building a "semantic denoiser" in the feature extraction stage, using high-level knowledge to suppress complex texture noise in shallow features, and significantly improving the signal-to-noise ratio of the features.

[0084] This invention utilizes Softmax, controlled by a temperature coefficient, to sharpen directional similarity weights, forcing the model to smooth features only along the road's extension direction while blocking smoothing in the vertical direction. This resolves the contradiction between road breakage repair and edge preservation at the mathematical and physical level.

[0085] This invention proposes explicit trajectory navigation, which ensures that the sampling trajectory strictly conforms to the road manifold by establishing a functional relationship between the sampling offset and the predicted geometric properties (direction, width), thus solving the problem of sampling point drift in complex backgrounds.

[0086] In the decoding and reconstruction process, this invention combines strip pooling (to resolve topology) with edge flow injection (to resolve boundaries) and introduces selective fusion based on global descriptors to construct a decoding system specifically for slender target structures. Attached Figure Description

[0087] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0088] Figure 1 This is a schematic diagram of a remote sensing road image segmentation method based on geometric reasoning, provided in Example 1.

[0089] Figure 2 This is a schematic diagram of a remote sensing road image segmentation device based on geometric reasoning, provided in Embodiment 2.

[0090] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Detailed Implementation

[0091] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0092] Example 1

[0093] Embodiment 1 of the present invention provides a remote sensing road image segmentation method based on geometric reasoning, which can be implemented by a remote sensing road image segmentation device based on geometric reasoning (hereinafter referred to as the segmentation device), specifically, executed by one or more processors within the segmentation device.

[0094] In this embodiment, the segmentation device may be an electronic device equipped with a processor, which carries a computer program for the remote sensing road image segmentation method based on geometric reasoning and is executable, such as a computer, smartphone, smart tablet, workstation, etc., without limitation.

[0095] In existing technologies, methods based on fully convolutional networks, due to the use of convolutional kernels with fixed geometric shapes, cannot establish long-range dependencies across occluded regions when processing roads obscured by trees, shadows, or buildings, resulting in breaks in the output mask at the occlusion points. While Transformer-based attention mechanisms expand the receptive field, their isotropic global associations introduce a large amount of background noise, causing blurred road edges and computational redundancy. Methods based on implicit dynamic sampling (such as deformable convolution) lack explicit geometric and physical constraints, and the sampling offset is generated by implicit regression of the network, making them prone to semantic drift in complex texture backgrounds, causing sampling points to deviate from the main road. To address these issues, this invention transforms the semantic segmentation task into a continuous-field reasoning problem constrained by physical geometry, and introduces an explicit, interpretable trajectory navigation mechanism. Combining spatial modulation, anisotropic diffusion, explicit trajectory aggregation, and hybrid structure decoding, it achieves high-precision, high-connectivity extraction of slender road targets in complex occlusion and strong interference backgrounds.

[0096] like Figure 1As shown, a remote sensing road image segmentation method based on geometric reasoning includes steps S1 to S5.

[0097] S1, acquire remote sensing road images;

[0098] S2, the remote sensing road image is input into the encoder for spatial feature extraction and modulation to obtain multi-scale features after feature modulation.

[0099] The remote sensing road image input encoder (such as S-LSKNet) extracts spatial features through a hierarchical large kernel decomposition convolutional network and outputs multi-scale features.

[0100] For example, a hierarchical large-kernel decomposition convolutional network is used to extract feature maps of four different scales, denoted as follows: (1 / 4 scale) (1 / 8 scale) (1 / 16 scale) and (1 / 32 scale).

[0101] Then, the deepest feature containing global semantic information in the multi-scale features is used to generate affine transformation parameters, as shown in the formula:

[0102] ;

[0103] in, This is a scaling factor used to enhance or suppress the characteristic response; This is the bias coefficient, used to adjust the baseline of the characteristic distribution; For the deepest features, such as ; This is a convolution operation; This is a channel-dimensional splitting operation;

[0104] Affine transformation parameters Upsampled to the same spatial size as the mid-level features in the multi-scale features by bilinear interpolation;

[0105] Based on the upsampled affine transformation parameters For mid-level features that contain rich details but are noisy (such as...) , The spatial features are linearly modulated, and the modulated mid-layer features are output as follows:

[0106] ;

[0107] in, These are the modulated mid-layer features; This is a mid-level feature; This indicates element-wise multiplication; To adjust the intensity hyperparameter, set it to 0.1; It is the hyperbolic tangent activation function;

[0108] Preserve the shallow features in the multi-scale features (such as...) ).

[0109] This embodiment utilizes the high-level semantic information of the deepest features as the gating signal. Responsible for enhancing or suppressing characteristic responses, This is responsible for adjusting the feature distribution baseline, thereby significantly improving the signal-to-noise ratio of shallow features. Another reason is that orientation and width information are needed later; if a feature map is too small, important orientation and width information will be lost. Therefore, the deepest features are used to modulate the middle-layer features.

[0110] S3 performs road geometry reasoning on the multi-scale features after feature modulation, generates a continuous geometric prior field, and outputs a reasoning fusion feature map.

[0111] This step utilizes the features output in step S2. The geometric properties of a path are inferred through a differentiable anisotropic diffusion mechanism, including the direction field of the path. (including horizontal and vertical components) and width field ( ).

[0112] Specifically, S31, the modulated mid-layer features are fused through convolutional layers. Output the initial orientation field and initial width field to obtain the initial fused feature map;

[0113] S32, for each center pixel in the initial fused feature map Take one Local spatial neighborhood (For example (A region containing 9 pixels); spatially expand the directional and width features within this neighborhood to construct a local feature matrix.

[0114] S33, for each center pixel in the initial fused feature map Extract the direction vector and local spatial neighborhood Each neighboring pixel Direction vector Perform a dot product operation to obtain a similarity scalar. The formula is:

[0115] ;

[0116] in, The direction vector representing the center pixel Direction vectors of neighboring pixels The angle between them; if A value close to 1 indicates that the two directions are the same (high probability that they belong to the same road); a value close to 0 or a negative number indicates that the directions are perpendicular or opposite.

[0117] S34, regarding the similarity scalar Application with temperature coefficient Softmax generates normalized diffusion weights. The formula is:

[0118] ;

[0119] in, Represents an exponential function. , All are local spatial neighborhoods The neighboring pixels in.

[0120] This step involves setting a larger temperature coefficient (e.g., ...). =10.0), this formula can significantly sharpen the weight distribution. That is, it assigns high weights only to "same type" neighboring pixels with the same orientation height, while quickly suppressing the weights of pixels with large orientation deviations to zero, thereby achieving selective focusing.

[0121] S35, based on the initial width field and the initial direction field, using The initial features within the local spatial neighborhood are aggregated using a targeted weighted method, as shown in the formula:

[0122] ;

[0123] ;

[0124] in, The width feature after aggregation; The directional features after aggregation; This is the initial width field; This is the initial direction field.

[0125] This step aims to simulate the heat conduction process in physics to achieve directional smoothing of features.

[0126] In remote sensing imagery, roads appear as elongated manifold structures with a specific orientation. Traditional feature aggregation methods (such as standard convolution or Gaussian smoothing) are typically isotropic, meaning they indiscriminately aggregate information from all neighborhoods surrounding the central pixel.

[0127] Directly applying isotropic smoothing to the width feature leads to two serious problems:

[0128] (1) Edge blurring and feature diffusion: The width information inside the road is averaged with the background on both sides of the road (width is 0), which causes the predicted road edge to expand outward and the boundary to become blurred.

[0129] (2) Adjacent interference: In the case of overpasses or parallel roads, the width information of adjacent roads will be incorrectly merged, resulting in the "sticking" phenomenon.

[0130] This invention introduces a directional field ( The anisotropic diffusion mechanism guided by this principle essentially establishes an anisotropic heat conduction simulation of a "selective feature valve." Unlike traditional convolution (isotropic, uniform diffusion in all directions), this step achieves anisotropic diffusion of features.

[0131] Flow along the path (repairing fractures): due to weight On paths with consistent orientation, the width and orientation information "flow" and converge along the direction of the road's extension. This means that even if the center pixel loses features due to occlusion (such as a broken path), it can repair itself by "borrowing" strong features from its neighbors.

[0132] Vertical blocking (edge ​​fidelity): due to weight The noise level is extremely low (close to 0) in the direction perpendicular to the road, preventing background noise from flowing into the road area. Mathematically, this forms a "feature firewall" that effectively prevents blurring and merging of road edges.

[0133] S36, the aggregated smoothed features, i.e., the smoothed width features. With smooth direction features The initial orientation field and initial width field in the initial fused feature map are residually fused, and the orientation vector field after residual fusion is normalized to output the inference fused feature map and the final geometric prior field, i.e., the orientation field. (Including horizontal and vertical components), width field .

[0134] S4. Based on the geometric prior field, perform dynamic trajectory sampling and adaptive aggregation on the inference fusion feature map to output long-range connected aggregated features.

[0135] Using the refined geometric prior (normalized direction field) output by step S3 and width field This step involves dynamically resampling and aggregating the inference fusion feature map. An explicit trajectory navigation mechanism is employed to address the semantic drift issue caused by the lack of geometric constraints in deformable or serpentine convolutions.

[0136] Specifically, in step S41, a basic coordinate grid with the same size as the inference fusion feature map is constructed, and the coordinates of each pixel position in the inference fusion feature map are... Mapping to a normalized coordinate system yields the starting center point for trajectory generation. ;

[0137] S42, Explicitly calculate the offset vector of the sampling step size based on the geometric prior field. A deterministic mapping between sampling geometry and road physical properties is established, using the following formula:

[0138] ;

[0139] in, The base step size is used to define the minimum physical interval between samples; For pixels The road width scalar at the location; This is a width adjustment factor used to control the degree to which the path width affects the sampling step size; For pixels The normalized direction vector at that point;

[0140] This formula establishes a deterministic mapping relationship between sampling geometry and physical properties.

[0141] By forcing sampling points to extend along the predicted road flow direction, the possibility of sampling points deviating from the vertical background is mathematically eliminated.

[0142] when When the offset is large (wide path), Automatically enlarges to cover a wider range of road surface features; when When the offset is small (narrow path), Automatically shrinks to avoid sampling roadside noise.

[0143] S43, then from the center point Using the offset vector as a reference, Extending in the forward (along the road's direction) and backward, a set of linear trajectories is generated. The formula is:

[0144] ;

[0145] in, For pixels The corresponding set of linear trajectories; This is the kth sampling point; For a pre-defined multi-scale scaling sequence, such as .

[0146] This step can be achieved by introducing a scaling factor that is not an integer multiple. (like This enables sub-pixel dense sampling, allowing trajectories to fall precisely on the centerline of a slender, curved road, like "threading a needle," thus constructing a long-range context sequence along the road manifold.

[0147] S44, based on a fundamental coordinate grid, utilizes a set of linear trajectories. Sampling points in Bilinear interpolation sampling is performed on the inference fusion feature map to extract features from each sampling point, and the sampled features are output using the following formula:

[0148] ;

[0149] in, For the k-th sampling feature; The inference fusion feature map; This is a feature sampling function based on a fundamental coordinate grid.

[0150] The resulting sampling features represent texture information at different locations, such as the center point, forward extension point, and backward extension point.

[0151] To further filter out potential local noise on the trajectory (such as the trajectory ending slightly beyond the road boundary), a geometrically aware attention module is introduced to fuse the sampled features.

[0152] S45, combine the features and orientation field of the inference fusion feature map. Width field Concatenate along the channel dimension, input into a lightweight convolutional network, and then... Normalization yields the confidence weight for each sampling point, using the following formula:

[0153] ;

[0154] in, The confidence weight for the k-th sampling point; This is a convolution operation; This is a normalization operation.

[0155] This step allows the network to dynamically determine the validity of each sampling point based on geometric priors and texture information. If a sampling point accidentally falls into the background, the attention mechanism will assign it a very low weight, thereby ensuring that the final aggregated features are pure and have very strong long-range connectivity.

[0156] S46. Finally, convolution optimization is performed on each group of sampled features, and then the features are weighted and summed using attention weights to obtain the final aggregated features. The formula is:

[0157] ;

[0158] in, This represents the total number of sampled features.

[0159] S5, decode and reconstruct the aggregated features and the shallow features in the multi-scale features, and output a binary road segmentation mask, i.e., the road segmentation result.

[0160] This step will aggregate the features output from step S4. (Rich in long-range context) and shallow high-resolution features extracted in step S2 (Rich in texture details) Input to the hybrid structure decoder. This step achieves complete reconstruction of the road topology and fine-grained segmentation of the boundaries through a multi-branch collaborative mechanism.

[0161] The decoding and reconstruction process is specifically as follows:

[0162] S51, Spatially align and fuse the aggregated features with the shallow features in the multi-scale features to generate a global channel descriptor vector.

[0163] To address the slight spatial alignment discrepancies between deep semantic features and shallow detail features, this step introduces a dynamic selection mechanism.

[0164] The operation of spatially aligning the aggregated features with the shallow features in the multi-scale features specifically involves:

[0165] The shallow features and the aggregated features are mapped to the same channel dimension by 1×1 convolution, and the aggregated features are upsampled to the resolution of the shallow features by bilinear interpolation to obtain the aligned feature streams corresponding to the shallow features and the aggregated features, respectively.

[0166] The shallow features and the aligned feature streams corresponding to the aggregated features are added and fused together, and then a one-dimensional global channel descriptor vector is generated by global average pooling, as expressed in the following expression:

[0167] ;

[0168] in, A global channel descriptor vector; This is the global average pooling function; , These are the shallow features and the aligned feature streams corresponding to the aggregated features, respectively.

[0169] S52, input the global channel descriptor vector into the MLP multilayer perceptron, and generate complementary channel attention vectors through Softmax normalization, as shown in the formula:

[0170] ;

[0171] in, These are complementary channel attention vectors.

[0172] S53, based on the channel attention vector, the global channel descriptor vector is weighted and recombined to output the attention fusion feature. The formula is:

[0173] ;

[0174] S54, based on the attention fusion features, respectively adopt level ( ),vertical Pooling is performed using strip pooling kernels to aggregate global context and obtain horizontal flow features. With vertical flow characteristics ;

[0175] S55, the horizontal flow features and the vertical flow features are superimposed, and a topological attention weight map is generated via Sigmoid. This map is then applied to the attention fusion features to repair road breaks caused by occlusion, resulting in an attention-weighted enhanced fusion feature, expressed as:

[0176] ;

[0177] ;

[0178] in, This is a topological attention weight graph; , These refer to the horizontal flow characteristics and the vertical flow characteristics, respectively. Attention-weighted enhancement of fusion features; This refers to attention fusion features.

[0179] In this way, pixels at the breakpoint can "borrow" road features on the same straight line in the distance to activate themselves, thereby effectively connecting the occluded road segments and significantly improving topological integrity.

[0180] Traditional semantic segmentation often suffers from blurred or aliased road edges due to interpolation effects during the stepwise upsampling process. To address this issue, this step constructs an independent edge-aware branch that explicitly extracts and injects high-frequency boundary information.

[0181] S56 applies a difference operator to shallow features in multi-scale features to extract high-frequency edge information and obtain a feature gradient map.

[0182] This operator acts like an edge detection filter, identifying the areas in an image where the brightness (grayscale) of each pixel and its surrounding pixels changes most rapidly and drastically.

[0183] The most obvious abrupt change in grayscale occurs precisely at the boundary between the road and adjacent non-road areas (such as road shoulders, grass, dirt, and lane dividers).

[0184] S57, the feature gradient map is mapped to an edge attention map through a 1×1 convolution;

[0185] S58, the edge attention map is superimposed on the attention-weighted enhanced fusion feature to sharpen the boundary, resulting in an explicit boundary enhanced fusion feature, expressed as:

[0186] ;

[0187] in, Explicitly enhance fusion features at the boundary; This is an edge attention map.

[0188] This step is equivalent to adding a "golden edge" to the blurry road features. It forces the model to pay more attention to the classification accuracy of boundary pixels when calculating the loss, thus making the output road mask outline clear and sharp.

[0189] S59, through cascaded upsampling, the explicit enhancement fusion features of the boundary are gradually restored to the original image resolution, and then a road segmentation probability map is output through 1×1 convolution;

[0190] The road segmentation probability map is processed by the Sigmoid function to obtain a binary road segmentation mask, i.e., the road segmentation result.

[0191] Unlike ordinary multi-scale feature concatenation or addition, this invention introduces spatial feature linear modulation in the encoder stage. Spatial affine parameters are generated using the global semantic information contained in deep features. Gated modulation is applied to the mid-level features. This is equivalent to building a "semantic denoiser" into the feature extraction stage, using high-level knowledge to suppress complex texture noise in shallow features, and significantly improving the signal-to-noise ratio of the features.

[0192] Unlike traditional convolutional isotropic feature extraction, this patent innovatively constructs a differentiable anisotropic diffusion module. Its key lies in using "temperature-coefficient controlled Softmax" to sharpen directional similarity weights, forcing the model to smooth features only along the road's extension direction while blocking smoothing in the vertical direction. This resolves the contradiction between road fracture repair and edge preservation at a mathematical and physical level.

[0193] Unlike existing deformable convolutional networks (DCNs) that use an implicit regression offset "black box" approach, this patent proposes "explicit trajectory navigation." Its key lies in establishing a deterministic functional relationship between the sampling offset and the predicted geometric properties (direction, width), ensuring that the sampling trajectory strictly conforms to the road manifold and solving the problem of sampling point drift in complex backgrounds.

[0194] Unlike ordinary upsampling decoding, this patent combines strip pooling (to solve topology) with edge stream injection (to solve boundaries) and introduces selective fusion based on global descriptors to construct a decoding system specifically for slender target structures.

[0195] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0196] This invention utilizes a spatial modulation encoder to extract multi-scale features and generates spatial affine parameters through deep semantic features to perform linear modulation (FiLM) on shallow features to suppress background noise.

[0197] This invention constructs a geometric inference engine that uses modulated features to infer the continuous direction field and width field of a road. This process includes anisotropic diffusion operations based on direction consistency.

[0198] This invention uses the inferred direction field and width field to explicitly calculate the sampling offset, generate multi-scale sampling trajectories distributed along the road direction, and perform vectorized resampling and aggregation on the feature map;

[0199] This invention inputs the aggregated features into a hybrid structure decoder, repairs topological breaks through strip pooling, sharpens boundaries through edge flow injection, and finally outputs a road segmentation mask.

[0200] Example 2

[0201] like Figure 2 As shown, the second embodiment of the present invention also provides a remote sensing road image segmentation device based on geometric reasoning guidance, comprising:

[0202] The acquisition unit is used to acquire remotely sensed road images;

[0203] The feature extraction and modulation unit is used to input the remote sensing road image into the encoder for spatial feature extraction and modulation to obtain multi-scale features after feature modulation.

[0204] The geometric inference unit is used to infer the road geometry from the multi-scale features after feature modulation, generate a continuous geometric prior field, and output an inference fusion feature map.

[0205] The sampling and aggregation unit is used to perform dynamic trajectory sampling and adaptive aggregation on the inference fusion feature map based on the geometric prior field, and output long-range connected aggregated features;

[0206] The decoding and reconstruction unit is used to decode and reconstruct the aggregated features and the shallow features in the multi-scale features, and output a binary road segmentation mask, i.e., the road image segmentation result.

[0207] Example 3

[0208] The third embodiment of the present invention also provides a remote sensing road image segmentation device based on geometric reasoning, which includes a memory and a processor. The memory stores a computer program, which can be executed by the processor to implement the remote sensing road image segmentation method based on geometric reasoning as described above.

[0209] Example 4

[0210] The fourth embodiment of the present invention also provides a computer-readable storage medium storing computer-readable instructions. When the computer-readable instructions are executed by the processor of the device where the computer-readable storage medium is located, they implement the remote sensing road image segmentation method based on geometric reasoning as described above.

[0211] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A remote sensing road image segmentation method based on geometric reasoning, characterized in that, include: Acquire remote sensing images of roads; The remote sensing road image is input into an encoder for spatial feature extraction and modulation to obtain multi-scale features after feature modulation. The road geometry is inferred from the multi-scale features after feature modulation, a continuous geometric prior field is generated, and the inference fusion feature map is output. Based on the geometric prior field, dynamic trajectory sampling and aggregation are performed on the inference fusion feature map to output the final aggregated features; The final aggregated features and the shallow features in the multi-scale features after feature modulation are decoded and reconstructed to output a binary road segmentation mask, which is the road image segmentation result. The process of spatial feature extraction and modulation is as follows: The remote sensing road image is input into the encoder and spatial features are extracted through a hierarchical large kernel decomposition convolutional network to output multi-scale features. Generate affine transformation parameters from the deepest feature containing global semantic information in the multi-scale features; The affine transformation parameters are upsampled to the same spatial size as the mid-level features in the multi-scale features through bilinear interpolation. Based on the upsampled affine transformation parameters, the spatial features of the middle-level features are linearly modulated to output the modulated middle-level features while retaining the shallow features in the multi-scale features. The road geometry is inferred from the multi-scale features after feature modulation, generating a continuous geometric prior field, specifically: By fusing the modulated mid-layer features through convolutional layers, the initial orientation field and initial width field are output to obtain the initial fused feature map; For each center pixel in the initial fused feature map, take one The local spatial neighborhood; For each center pixel in the initial fused feature map, extract the direction vector and perform a dot product operation with the direction vector of each neighboring pixel in the local spatial neighborhood to obtain a similarity scalar. Applying a temperature-dependent Softmax to the similarity scalar generates normalized diffusion weights; Based on the initial width field and the initial direction field, the initial features in the local spatial neighborhood are aggregated in a directional weighted manner using normalized diffusion weights. The aggregated smooth features are residually fused with the initial orientation field and initial width field in the initial fused feature map, and the orientation vector field after residual fusion is normalized to output the inference fused feature map and the final geometric prior field, i.e., the orientation field. Width field .

2. The remote sensing road image segmentation method based on geometric reasoning as described in claim 1, characterized in that... The formula for the affine transformation parameters is: ； in, This is a scaling factor used to enhance or suppress the characteristic response; This is the bias coefficient, used to adjust the baseline of the characteristic distribution; The deepest feature; This is a convolution operation; This is a channel-dimensional splitting operation; The formula for the modulated middle-layer features is: ； in, These are the modulated mid-layer features; This is a mid-level feature; This indicates element-wise multiplication; To adjust the hyperparameters of intensity; It is the hyperbolic tangent activation function.

3. The remote sensing road image segmentation method based on geometric reasoning as described in claim 2, characterized in that... The formula for the similarity scalar is: ； in, It is a similarity scalar; Center pixel The direction vector; For neighboring pixels The direction vector; The direction vector representing the center pixel Direction vectors of neighboring pixels The angle between them; The formula for the normalized diffusion weight is: ； in, For normalized diffusion weights; Represents an exponential function. , All are local spatial neighborhoods The neighboring pixels in; Based on the initial width field and the initial direction field, using normalized diffusion weights The formula for targeted weighted aggregation of initial features within a local spatial neighborhood is as follows: ； ； in, The width feature after aggregation; The directional features after aggregation; This is the initial width field; This is the initial direction field.

4. The remote sensing road image segmentation method based on geometric reasoning as described in claim 3, characterized in that... The geometric prior field performs dynamic trajectory sampling and aggregation on the inference fusion feature map, specifically as follows: Construct a base coordinate grid with the same size as the inference fusion feature map, and assign coordinates to the position of each pixel in the inference fusion feature map. Mapping to a normalized coordinate system yields the starting center point for trajectory generation. ; The offset vector of the sampling step size is explicitly calculated based on the geometric prior field. A deterministic mapping between sampling geometry and road physical properties is established, using the following formula: ； in, The base step size is used to define the minimum physical interval between samples; For pixels The road width scalar at the location; This is a width adjustment factor used to control the degree to which the path width affects the sampling step size; For pixels The normalized direction vector at that location; Then, using the center point Using the offset vector as a reference, Extending forward and backward, a set of linear trajectories is generated. The formula is: ； in, For pixels The corresponding set of linear trajectories; This is the kth sampling point; This is a pre-defined multi-scale scaling sequence; Based on a fundamental coordinate grid, using a set of linear trajectories sampling points in Bilinear interpolation sampling is performed on the inference fusion feature map to extract features from each sampling point, and the sampled features are output using the following formula: ； in, For the k-th sampling feature; The inference fusion feature map; It is a feature sampling function based on the basic coordinate grid; The features and orientation field of the inference fusion feature map are combined. Width field Concatenate along the channel dimension, input into a lightweight convolutional network, and then... Normalization yields the confidence weight for each sampling point, using the following formula: ； in, The confidence weight for the k-th sampling point; This is a convolution operation; This is a normalization operation; Finally, convolution optimization is performed on each group of sampled features, and then the features are weighted and summed using attention weights to obtain the final aggregated features. The formula is: ； in, This represents the total number of sampled features in the set of linear trajectories.

5. The remote sensing road image segmentation method based on geometric reasoning as described in claim 4, characterized in that... The decoding and reconstruction process is as follows: The final aggregated features are spatially aligned and fused with the shallow features in the multi-scale features after feature modulation to generate a global channel descriptor vector. The global channel descriptor vector is input into the MLP multilayer perceptron, and complementary channel attention vectors are generated through Softmax normalization. The global channel descriptor vector is weighted and recombined based on the channel attention vector to output attention fusion features; Based on the attention fusion features, horizontal and vertical bar pooling kernels are used for pooling to aggregate the global context and obtain horizontal flow features and vertical flow features. The horizontal flow features and the vertical flow features are superimposed, and a topological attention weight map is generated using a sigmoid function. This map is then applied to the attention fusion features to repair road breaks caused by occlusion, resulting in an attention-weighted enhanced fusion feature, expressed as: ； ； in, This is a topological attention weight graph; , These refer to the horizontal flow characteristics and the vertical flow characteristics, respectively. Attention-weighted enhancement of fusion features; For attention fusion features; By applying the difference operator to the shallow features in the multi-scale features, high-frequency edge information is extracted to obtain the feature gradient map; The feature gradient map is mapped to an edge attention map using a 1×1 convolution; The edge attention map is superimposed on the attention-weighted enhanced fusion feature to sharpen the boundary, resulting in an explicit boundary enhanced fusion feature, expressed as: ； in, Explicitly enhance fusion features at the boundary; For edge attention maps; By cascading upsampling, the explicit enhancement fusion features of the boundary are gradually restored to the original image resolution, and then a road segmentation probability map is output through 1×1 convolution. The road segmentation probability map is processed by the Sigmoid function to obtain a binary road segmentation mask, i.e., the road segmentation result.

6. The remote sensing road image segmentation method based on geometric reasoning as described in claim 5, characterized in that... The operation of spatially aligning the final aggregated features with the shallow features in the feature-modulated multi-scale features is specifically as follows: The shallow features and the final aggregated features are mapped to the same channel dimension by 1×1 convolution. The final aggregated features are upsampled to the resolution of the shallow features by bilinear interpolation, and the aligned feature streams corresponding to the shallow features and the final aggregated features are obtained respectively. The shallow features and the aligned feature streams corresponding to the final aggregated features are added and fused together, and then a one-dimensional global channel descriptor vector is generated by global average pooling, as expressed in the following expression: ； in, A global channel descriptor vector; This is the global average pooling function; , These are the aligned feature flows corresponding to the shallow features and the final aggregated features, respectively.

7. A remote sensing road image segmentation device based on geometric reasoning guidance, used to implement the remote sensing road image segmentation method based on geometric reasoning guidance as described in any one of claims 1-6, characterized in that, include: The acquisition unit is used to acquire remotely sensed road images; The feature extraction and modulation unit is used to input the remote sensing road image into the encoder for spatial feature extraction and modulation to obtain multi-scale features after feature modulation. The geometric inference unit is used to infer the road geometry from the multi-scale features after feature modulation, generate a continuous geometric prior field, and output an inference fusion feature map. The sampling and aggregation unit is used to perform dynamic trajectory sampling and adaptive aggregation on the inference fusion feature map based on the geometric prior field, and output the final aggregated features; The decoding and reconstruction unit is used to decode and reconstruct the final aggregated features and the shallow features in the multi-scale features after feature modulation, and output a binary road segmentation mask, that is, the road image segmentation result.

8. A remote sensing road image segmentation device based on geometric reasoning guidance, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that can be executed by the processor to implement a remote sensing road image segmentation method based on geometric reasoning as described in any one of claims 1-6.

Images