Methods, apparatus, electronic equipment and storage media for pipeline extraction in chemical industrial parks

Abstract

This invention provides a method, apparatus, electronic device, and storage medium for extracting pipelines in chemical industrial parks, relating to the field of image processing technology. The method includes: acquiring remote sensing images of the chemical industrial park to be processed; inputting the remote sensing images into a pipeline extraction model; and obtaining a pipeline segmentation map of the chemical industrial park output by the pipeline extraction model. This invention, by setting up a pipeline decoder module with a striped receptive field, can enhance the characteristics of slender pipelines, effectively solving the problems of low extraction accuracy for extremely fine linear targets and easy omission of small branch pipes. Simultaneously, by setting up a cascaded refinement module, iteratively correcting the pipeline boundaries based on the high-resolution feature map and the current mask estimation, it improves the positioning accuracy of the segmentation boundaries and reduces the false detection rate caused by the similarity between pipeline and background features in complex industrial environments such as storage tanks and towers, thereby achieving high-precision extraction of pipelines in chemical industrial parks.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method, apparatus, electronic device, and storage medium for extracting pipelines in a chemical industrial park. Background Technology

[0002] Chemical industrial parks, as core carriers of modern industrial systems, undertake crucial functions in the production, storage, transportation, and conversion of chemical raw materials. The dense network of pipelines within these parks serves as the material transport channels for chemical production processes, and their accurate mapping is of significant practical importance for safety supervision, emergency response, and facility maintenance. Remote sensing technology, with its inherent advantages of wide coverage, non-contact observation, and multi-temporal monitoring, has become a primary means of monitoring industrial facilities and infrastructure, providing a feasible technical approach for the detailed mapping of pipelines in chemical industrial parks.

[0003] Deep semantic segmentation methods, represented by FCN (Fully Convolutional Networks) and U-shaped networks, have made significant progress in the field of high-resolution remote sensing image processing. Existing research mainly focuses on urban road network extraction, while there is relatively little research on the fine extraction of pipelines in chemical industrial parks.

[0004] Both roads and pipelines appear as long, thin lines in remote sensing images. However, they differ in several ways: chemical pipelines are much narrower than roads, occupying only a few pixels in the image; their ends connect to chemical plants in a tree-like, radiating pattern; and the dense concentration of storage tanks, towers, and other facilities within industrial parks, whose spectral characteristics are similar to pipelines, easily leads to false positives and false negatives. Therefore, directly applying road extraction methods to pipeline extraction tasks is unlikely to yield ideal results.

[0005] In summary, improving the accuracy of pipeline extraction in chemical industrial parks has become an urgent problem to be solved. Summary of the Invention

[0006] This invention provides a method, apparatus, electronic device, and storage medium for extracting pipelines in chemical industrial parks, thereby overcoming the shortcomings of low extraction accuracy in existing technologies and achieving high-precision extraction of pipelines in chemical industrial parks.

[0007] This invention provides a method for extracting pipelines in a chemical industrial park, comprising the following steps: Acquire remote sensing images of the chemical industrial park to be processed; The remote sensing image of the chemical industrial park is input into the pipeline extraction model to obtain the pipeline segmentation map of the chemical industrial park output by the pipeline extraction model. The pipeline extraction model includes: Multiple pipeline decoder modules with striped receptive fields are configured to progressively upsample and fuse multi-level feature maps of the remote sensing image of the chemical industrial park to enhance the feature response to pipeline orientation and output high-resolution pipeline feature maps. The cascaded refinement module is configured to correct the pipeline boundary based on the high-resolution pipeline feature map and the current mask estimation of the pipeline, and output a pipeline segmentation map with refined boundary.

[0008] According to the present invention, a method for extracting pipelines in a chemical industrial park is provided in which each pipeline decoder module receives a feature map from a multi-scale feature pyramid and a feature map output from the previous decoding stage, and outputs a feature map with a higher spatial resolution than its input feature map.

[0009] According to the present invention, a method for extracting pipelines in a chemical industrial park is provided. The pipeline decoder module includes a strip deformable convolutional unit. The strip deformable convolutional unit generates a strip receptive field through cascaded 1×N convolution and N×1 convolution, and combines deformable offset and modulation mask to adaptively adjust the sampling position and reweight features.

[0010] According to the method for extracting pipelines in a chemical industrial park provided by the present invention, the cascaded refinement module is further configured as follows: An edge mask is generated based on the high-resolution pipeline feature map; The pipeline boundary is corrected based on the high-resolution pipeline feature map, the current mask estimate of the pipeline, and the edge mask.

[0011] According to the method for extracting pipelines in a chemical industrial park provided by the present invention, the cascaded refinement module is further configured as follows: The high-resolution pipeline feature map, the current mask estimate processed by the activation function, and the edge mask are stitched together in the channel dimension; The concatenated features are processed sequentially through convolutional layers to output residual correction terms. The residual correction term is added to the current mask estimate to obtain the updated mask estimate.

[0012] According to the present invention, a method for extracting pipelines in a chemical industrial park is provided. The pipeline extraction model further includes a backbone network module for forward propagation of the input remote sensing image of the chemical industrial park to extract the multi-level feature map.

[0013] According to the present invention, a method for extracting pipelines in a chemical industrial park is provided. The pipeline extraction model further includes a feature pyramid construction module, which is used to project and scale the multi-level feature map, and output a multi-scale feature pyramid with a uniform number of channels and a spatial scale that is multiple of each other.

[0014] The present invention also provides a pipeline extraction device for chemical industrial parks, comprising the following modules: The acquisition module is used to acquire remote sensing images of the chemical industrial park to be processed; The pipeline extraction module is used to input the remote sensing image of the chemical industrial park into the pipeline extraction model and obtain the pipeline segmentation map of the chemical industrial park output by the pipeline extraction model. The pipeline extraction model includes: Multiple pipeline decoder modules with striped receptive fields are configured to progressively upsample and fuse multi-level feature maps of the remote sensing image of the chemical industrial park to enhance the feature response to pipeline orientation and output high-resolution pipeline feature maps. The cascaded refinement module is configured to correct the pipeline boundary based on the high-resolution pipeline feature map and the current mask estimation of the pipeline, and output a pipeline segmentation map with refined boundary.

[0015] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the chemical industrial park pipeline extraction method described above.

[0016] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the chemical industrial park pipeline extraction method as described above.

[0017] The present invention provides a method, apparatus, electronic device, and storage medium for extracting pipelines in chemical industrial parks. The method involves acquiring remote sensing images of the chemical industrial park to be processed; inputting the remote sensing images into a pipeline extraction model to obtain a pipeline segmentation map of the chemical industrial park output by the model; wherein the pipeline extraction model includes: multiple pipeline decoder modules with striped receptive fields, configured to progressively upsample and fuse multi-level feature maps of the remote sensing images of the chemical industrial park to enhance the feature response to pipeline orientation and output a high-resolution pipeline feature map; and a cascaded refinement module configured to correct the pipeline boundaries based on the high-resolution pipeline feature map and the current mask estimation of the pipelines, outputting a refined pipeline segmentation map. This invention, by setting up a pipeline decoder module with a strip-shaped receptive field, can enhance the characteristics of pipelines with slender orientations, effectively solving the problems of low extraction accuracy of extremely fine linear targets and easy omission of small branch pipes. At the same time, by setting up a cascaded refinement module, the pipeline boundary is iteratively corrected based on the high-resolution feature map and the current mask estimation, which improves the positioning accuracy of the segmentation boundary and reduces the false detection rate caused by the similarity between pipeline and background features in complex industrial backgrounds such as storage tanks and towers, thereby achieving high-precision extraction of pipelines in chemical industrial parks. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a schematic flowchart of the method for extracting pipelines in chemical industrial parks provided by the present invention.

[0020] Figure 2 This is a schematic diagram of the pipeline network in a chemical industrial park provided by the present invention.

[0021] Figure 3 This is the PDSDNet architecture diagram provided by the present invention.

[0022] Figure 4 This is the PDB architecture diagram provided by the present invention.

[0023] Figure 5 This is the CRM architecture diagram provided by the present invention.

[0024] Figure 6 This is a visual comparison diagram of the segmentation results of different models provided by this invention on images of a typical chemical industrial park.

[0025] Figure 7 This is a schematic diagram of the structure of the pipeline extraction device for chemical industrial parks provided by the present invention.

[0026] Figure 8 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0028] There are certain differences between pipeline extraction in chemical industrial parks and remote sensing road extraction, which means that existing road extraction algorithms cannot be simply transferred to pipeline extraction tasks. Compared to roads, pipelines have a smaller width, which makes the extraction of fine-grained features from the model more demanding (e.g., ...). Figure 1 (a) The tree-like branch structure also places higher demands on the model's ability to perceive details of branching (e.g. Figure 1 (b, c); the background inside the chemical industrial park also differs significantly from urban road scenes (e.g., Figure 1 (c in the text)

[0029] Based on this, this invention proposes a method for extracting pipelines in chemical industrial parks. This method is based on DINOv3 (Distillation with No Labels version 3) and proposes a dual-head segmentation based DINOv3Network (PDSDNet) specifically for remote sensing extraction of pipelines in chemical industrial parks. The network uses DINOv3 ViT-L (Distillation with No Labels version 3 Vision Transformer Large) pre-trained on the SAT-493M dataset as the backbone network, designs a pipeline decoding block (PDB), and introduces a cascading refinement module (CRM) to suppress false positive and false negative errors of small branches.

[0030] The training and testing experiments of the pipeline extraction model involved in this invention are both implemented based on the PyTorch deep learning framework and completed on a single NVIDIA GeForce RTX 3090 graphics processor (24 GB VRAM). The model adopts an end-to-end training method, and the parameters of the backbone network DINOv3 ViT-L are kept frozen during training, with only the learnable parameters in the decoder and cascaded refinement modules being updated.

[0031] The following is combined with Figures 1-8 The present invention describes a method, apparatus, electronic device, and storage medium for extracting pipelines in chemical industrial parks.

[0032] Figure 1 This is a schematic flowchart of the pipeline extraction method for chemical industrial parks provided by the present invention, as shown below. Figure 1 As shown, the method includes the following: Step 101: Obtain remote sensing images of the chemical industrial park to be processed.

[0033] High-resolution satellite or aerial imagery data covering the geographical area of ​​a target chemical industrial park is acquired or received from remote sensing data sources. For example, the latitude and longitude range or vector boundary of the target chemical industrial park is first determined as a spatial constraint for image acquisition; then, high-resolution remote sensing images covering the park are acquired through a remote sensing data service platform (such as a drone aerial photography system). The acquired image data typically includes three visible light bands: red, green, and blue. After preprocessing such as radiometric calibration, atmospheric correction, and geometric fine correction, a standard format remote sensing image file (such as GeoTIFF format) is formed. In the imagery, pipelines in the chemical industrial park appear as long, continuous linear features, typically occupying several pixels in width; storage tanks appear as circular or elliptical areas; and towers appear as bright point-like or small-area features.

[0034] Step 102: Input the remote sensing image of the chemical industrial park into the pipeline extraction model to obtain the pipeline segmentation map of the chemical industrial park output by the pipeline extraction model.

[0035] The pipeline extraction model is an end-to-end semantic segmentation neural network based on deep learning. It uses a DINOv3 visual Transformer with frozen parameters as the backbone network, multiple stacked pipeline decoder modules to form the decoding stage, and cascades edge-guided iterative refinement modules to map the input remote sensing image of the chemical industrial park into a pipeline segmentation map at the pixel level.

[0036] The pipeline extraction model includes: Multiple pipeline decoder modules with strip-shaped receptive fields are configured to progressively upsample and fuse multi-level feature maps from remote sensing images of the chemical industrial park to enhance the feature response to pipeline orientation and output high-resolution pipeline feature maps. Specifically, the decoding stage consists of multiple pipeline decoder modules (PDBs) connected in series. Each PDB generates a non-square strip-shaped receptive field through cascaded 1×N and N×1 convolutions, giving it a stronger feature response capability to slender pipeline targets. Multiple PDBs process the input multi-scale feature pyramid in a cascaded manner, with each level doubling the spatial resolution of the feature map through transposed convolution and then performing skip connections and fusion with the feature pyramid feature maps of the corresponding scale. After progressive upsampling and feature enhancement, the final output is a high-resolution pipeline feature map with a spatial resolution close to the original image size and rich semantic information.

[0037] The cascaded refinement module is configured to refine pipeline boundaries based on the high-resolution pipeline feature map and the current mask estimate of the pipeline, outputting a refined pipeline segmentation map. Specifically, the cascaded refinement module receives the high-resolution pipeline feature map and the current mask estimate output by the pipeline decoder module, extracts edge information from them, and fuses it with the original features and the current mask. Through a residual learning mechanism with multiple iterations, the boundary regions of the mask are progressively refined. Each iteration outputs a residual correction term, which is added to the current mask. After multiple iterations, a refined segmentation map with more accurate boundaries and a higher alignment with the real pipeline boundaries is output.

[0038] In one embodiment, the remote sensing extraction method for pipelines in a chemical industrial park based on a pipeline extraction model includes the following steps: Step S1: Acquire remote sensing images of the chemical industrial park to be processed; Step S2: Input the remote sensing image of the chemical industrial park into the parameter-frozen DINOv3 visual Transformer backbone network. The backbone network propagates from front to back and outputs multi-level feature maps from multiple preset Transformer layers of different depths. Step S3: Project and scale the multi-level feature maps to construct a multi-scale feature pyramid with a uniform number of channels and a spatial scale that is multiple of each other; Step S4: Input the multi-scale feature pyramid into the pipeline decoder module. The pipeline decoder module generates a high-resolution pipeline feature map by progressively upsampling and fusing multi-scale features. Each pipeline decoder module embeds a strip deformable convolution module. The strip deformable convolution module generates a strip receptive field through cascaded 1×N convolutions and N×1 convolutions, and performs feature enhancement on the slender pipeline structure based on the learned deformable sampling offset and modulation mask. Step S5: Input the high-resolution pipeline feature map into the cascaded refinement module. The cascaded refinement module generates an edge mask based on the high-resolution pipeline feature map, and then performs channel stitching of the high-resolution pipeline feature map, the edge mask and the current mask estimate. Through multiple iterations of residual learning, the pipeline boundary is gradually refined, and the final pipeline segmentation map is output.

[0039] In one specific embodiment, it is assumed that the size of the remote sensing image of the chemical industrial park used as the input pipeline extraction model is [size missing]. (Height × Width × RGB three bands), pixel value range normalized to [0,1] or [-1,1]. Image data first enters the DINOv3 ViT-L backbone network with frozen parameters for forward propagation. The backbone network outputs multi-level feature maps from the 8th, 16th, and 24th Transformer layers, with spatial resolutions of […]. , , , Then, the feature pyramid construction module projects and scales these multi-level feature maps to form a unified number of channels (256 dimensions) with spatial scales that are multiples of each other. , , , The system employs a multi-scale feature pyramid. Multiple stacked pipeline decoder modules progressively upsample and fuse the feature pyramid. Each pipeline decoder module contains a strip-shaped deformable convolutional unit. A strip-shaped receptive field is generated through cascaded 1×N and N×1 convolutions. Combined with deformable offsets and modulation masks, this enhances the response to features along the slender pipe orientation, outputting a high-resolution pipeline feature map (e.g., spatial resolution restored to [specific scale]). The cascaded refinement module generates an edge mask based on the high-resolution pipeline feature map and the current mask estimate of the pipeline, and gradually corrects the pipeline boundaries through multiple iterations of residual learning. Finally, a 1×1 convolution projects the features into a single-channel feature map, which, after passing through a Sigmoid activation function, outputs a probability map of pixel values ​​between [0,1]. A threshold segmentation is then used to generate a binary segmentation mask, where pixels with a value of 1 represent pipeline regions and pixels with a value of 0 represent background regions. This binary mask is the final pipeline segmentation map. Through the above process, the model maps the input high-resolution remote sensing image pixel-level to a segmentation result reflecting the topology of the pipeline network in the chemical industrial park.

[0040] The pipeline extraction method for chemical industrial parks provided in this invention involves acquiring remote sensing images of the chemical industrial park to be processed; inputting the remote sensing images into a pipeline extraction model to obtain a pipeline segmentation map of the chemical industrial park output by the pipeline extraction model; wherein, the pipeline extraction model includes: multiple pipeline decoder modules with strip receptive fields, configured to progressively upsample and fuse multi-level feature maps of the remote sensing images of the chemical industrial park to enhance the feature response to pipeline orientation and output a high-resolution pipeline feature map; and a cascaded refinement module, configured to correct the pipeline boundaries based on the high-resolution pipeline feature map and the current mask estimation of the pipeline, and output a pipeline segmentation map with refined boundaries. This invention, by setting up a pipeline decoder module with a strip-shaped receptive field, can enhance the characteristics of pipelines with slender orientations, effectively solving the problems of low extraction accuracy of extremely fine linear targets and easy omission of small branch pipes. At the same time, by setting up a cascaded refinement module, the pipeline boundary is iteratively corrected based on the high-resolution feature map and the current mask estimation, which improves the positioning accuracy of the segmentation boundary and reduces the false detection rate caused by the similarity between pipeline and background features in complex industrial backgrounds such as storage tanks and towers, thereby achieving high-precision extraction of pipelines in chemical industrial parks.

[0041] In one embodiment, the pipeline extraction model further includes a backbone network module for forward propagating the input remote sensing image of the chemical industrial park to extract the multi-level feature map.

[0042] The backbone network module is used to extract multi-level feature maps from the input remote sensing imagery of the chemical industrial park. For example... Figure 3 As shown, the backbone network module uses DINOv3 ViT-L pre-trained on the SAT-493M dataset as the basic feature extractor, which contains 24 Transformer layers.

[0043] The parameters of the backbone network modules remain frozen throughout the model training and inference process; that is, they do not participate in gradient updates and are used only as fixed feature extractors. Input: Remote sensing imagery of a chemical industrial park. After forward propagation through the backbone network, intermediate feature maps are extracted from the 8th, 16th, and 24th Transformer layers, as follows: 8th layer output feature map This feature map retains a lot of local texture details and spatial location information, and responds strongly to low-level geometric features such as the edges and orientation of pipelines.

[0044] 16th layer output feature map This feature map combines local details and intermediate semantic information, and responds well to intermediate features such as pipeline structure and branches.

[0045] Output feature map of layer 24 This feature map has a large receptive field and strong semantic abstraction ability, and responds strongly to high-level semantic features such as the overall topology of pipelines and long-range dependencies, but has low spatial resolution.

[0046] The three-layer feature maps mentioned above simultaneously contain hierarchical feature representations ranging from local texture details to global semantic information.

[0047] This invention employs a frozen DINOv3 ViT-L as the backbone network, which eliminates the need for training this part to obtain high-quality remote sensing image feature representations, thereby reducing training costs.

[0048] In one embodiment, the pipeline extraction model further includes a feature pyramid construction module, which is used to project and scale the multi-level feature maps to output a multi-scale feature pyramid with a uniform number of channels and a spatial scale that is multiple of each other.

[0049] The feature pyramid construction module is used to convert the multi-level feature maps output by the backbone network into multi-scale feature pyramids with uniform channel counts and spatial scales that are multiples of each other. For example... Figure 3As shown, the feature pyramid construction module receives multi-level feature maps output from layers 8, 16, and 24 of the backbone network. The specific processing flow is as follows: (1) Apply 1×1 convolution to the feature maps output from layers 8, 16 and 24 respectively, and project the number of channels of each feature map to 256.

[0050] (2) Perform spatial transformations on the feature map after channel projection at different scales, as follows: Layer 8 Feature Processing: The feature map output from Layer 8 of the backbone network is first subjected to a 1×1 convolutional projection, and then resampled to spatial resolution using a 4×4 deconvolution (ConvTranspose), batch normalization (BN), and ReLU activation function. The feature map, denoted as .

[0051] Feature processing at layer 16: The feature map output from layer 16 of the backbone network is first subjected to a 1×1 convolutional projection, and then resampled to spatial resolution through deconvolution with a kernel size of 2×2, batch normalization, and the ReLU activation function. The feature map, denoted as .

[0052] Feature processing at layer 24: The feature map output from layer 24 of the backbone network is directly used as the feature map with a resolution of [resolution value missing]. The feature map, denoted as Simultaneously, the feature map output from layer 2 is further subjected to max pooling downsampling with a stride of 2, resulting in a spatial resolution of [missing value]. The feature map, denoted as .

[0053] (3) Through the above processing, the feature pyramid construction module outputs four feature maps. Their spatial dimensions are respectively , , and Each feature map has 256 channels. The spatial scales of these four feature maps are multiples of each other, forming a bottom-up multi-scale feature pyramid. It has the highest resolution, retains rich spatial details, and is used to capture the end positions of tiny micro-branchs; It has the lowest resolution but the largest receptive field, used to capture the overall topology and long-range dependencies of pipelines. and It serves as an intermediate scale, balancing details and semantics.

[0054] The embodiments of the present invention provide the decoder with a clear multi-level feature structure by unifying the number of channels and aligning with multiple scales, thereby enhancing the model's ability to express features of pipelines at different scales and its anti-interference ability in complex backgrounds.

[0055] Based on the above embodiments, each pipeline decoder module receives a feature map from a multi-scale feature pyramid and a feature map output from the previous decoding stage, and outputs a feature map with a spatial resolution higher than its input feature map.

[0056] Each pipeline decoder module (PDB) is designed according to a unified input / output pattern. Multiple PDBs are cascaded to form a complete decoding stage, achieving step-by-step recovery of the feature map spatial resolution. For example... Figure 3 As shown, except for the first PDB, the input of each PDB contains two parts: ① feature maps from the multi-scale feature pyramid. This feature map has low spatial resolution. ② Feature map output from the previous decoding stage. This feature map has high spatial resolution. Indicates the index of the decoding stage. Representing the multi-scale feature pyramid and the first Feature maps corresponding to the decoding stage. Indicates the first The feature map output from the decoding stage. Indicates the previous decoding stage (the... The feature map output at the level of (level).

[0057] It should be noted that the first PDB does not contain the feature map output from the previous decoding stage. Therefore, the input to the first PDB consists of two feature maps from the multi-scale feature pyramid. ,like and .

[0058] The processing flow for each PDB is as follows: (1) Feature map output from the previous decoding stage By applying a transposed convolution with a stride of 2 (ConvTranspose), its spatial resolution is doubled, resulting in... This makes it compatible with the feature maps in the multi-scale feature pyramid. It matches the spatial resolution.

[0059] For example, let the input feature map Feature map output from the previous decoding stage Given the condition, where, Indicates the first The number of channels in the level feature map. Indicates the first The downsampling factor of the level feature map; That is, the downsampling factor relationship between two adjacent levels, for example , The feature map output from the previous decoding stage. Apply a transposed convolution with a stride of 2 to achieve the same spatial resolution as... match: .

[0060] It should be noted that the reference Figure 3 For the first-level PDB, an upsampling alignment step needs to be performed. For example, the first-level PDB receives feature maps from a multi-scale feature pyramid. (resolution) ) and feature map (resolution) At this point, the spatial resolutions of the two feature maps do not match, and an upsampling alignment step needs to be performed.

[0061] (2) Upsampled and By splicing along the channel dimension, a combined feature that integrates semantic information and spatial details is formed.

[0062] (3) The concatenated features are sequentially passed through the StripDCN module (strip deformable convolutional unit), batch normalization, ReLU activation function, and a 3×3 convolutional layer to output the feature map of this stage. . refer to Figure 3 Output feature map The spatial resolution is the input feature map and the output feature map of the previous decoding stage Twice the spatial resolution: Level 1 PDB: Receives feature maps from the multi-scale feature pyramid. (resolution) ) and feature map (resolution) Output feature map (resolution) ); Second-level PDB: Receives feature maps from the multi-scale feature pyramid. (resolution) ) and the feature map output from the previous decoding stage (resolution) Output feature map (resolution) ); Level 3 PDB: Receives feature maps from the multi-scale feature pyramid. (resolution) ) and the feature map output from the previous decoding stage (resolution) Output feature map (resolution) ).

[0063] The embodiments of the present invention gradually restore the spatial details of pipelines by upsampling and fusing multi-scale features at each level, thereby enhancing the response capability to slender routes and small branches, and thus improving the integrity and positioning accuracy of pipeline extraction.

[0064] Based on the above embodiments, the pipeline decoder module includes a strip deformable convolutional unit. The strip deformable convolutional unit generates a strip receptive field through cascaded 1×N convolutions and N×1 convolutions, and combines deformable offsets and modulation masks to adaptively adjust the sampling position and reweight features.

[0065] Deformable offset refers to the two-dimensional position offset vector learned for each sampling point in a deformable convolutional unit, denoted as: ,in, This represents the total number of sampling points. This offset is used to sample the regular grid of the standard convolution. Adjustments are made to allow the sampling points to adaptively shift according to the actual geometry of the target object. For pipeline targets, the shift can cause the sampling points to be distributed in a serpentine pattern along the pipeline, thereby achieving accurate sampling of curved and irregularly oriented pipelines.

[0066] The modulation mask refers to the scalar weight coefficients learned for each sampling point in a deformable convolutional unit, denoted as... This mask is used to reweight the features extracted from different sampling point locations, enabling the model to highlight key features of the pipeline area and suppress background interference, thereby enhancing the feature response to pipeline targets.

[0067] The pipeline decoder module incorporates a Strip Deformable Convolution Unit (StripDCN). This unit replaces traditional square convolution kernels with cascaded 1×N and N×1 convolution kernels, forming a strip-shaped receptive field covering both the horizontal and vertical directions. This enhances its feature response capability for elongated pipeline targets. Simultaneously, this unit learns deformable offsets and a modulation mask from the features of the preceding layers. The deformable offsets adaptively shift the convolutional sampling points along the pipeline's direction to accommodate its curvature, while the modulation mask adjusts the weights of features at different sampling points to highlight the pipeline region and suppress background interference. Together, they achieve adaptive feature enhancement based on the pipeline's orientation.

[0068] In one embodiment, reference Figure 4The StripDCN module is a specially designed deformable convolutional layer intended to enhance the model's ability to capture elongated structures, such as object boundaries and slender targets. Unlike standard deformable convolutions, StripDCN extracts features from the input... Simultaneous prediction of sampling offset and modulation mask weights Its operation is defined as follows: ; in, Indicates the position on the output feature map The eigenvalue at position, i.e., after StripDCN transformation, is... The output calculation results; This indicates the target location coordinates on the output feature map, i.e., the spatial location of the currently calculated feature point in the output feature map; This represents the total number of sampling points, i.e., the total number of sampling locations (e.g., for...). Convolution kernel, ); Indicates the index of the sampling point; Indicates the first Convolution weights for each sampling point; This represents the pixel value on the input feature map, that is, it means taking the feature value at a specified location in the input feature map; Indicates a predefined regular grid offset; This represents the learned deformable offset, that is, the first... Position offset of each sampling point; This indicates the actual sampling position, i.e., the final sampling coordinates after offset adjustment; This represents the learned modulation mask weights, that is, the weights of the first modulation mask. The feature reweighting coefficients for each sampling point. Inside StripDCN, the generation of deformable offsets and modulation masks is handled by a dedicated convolutional branch, which is... Convolution and The convolutions are concatenated to enhance the strip receptive field, and then passed through a... Convolutional regression yields the complete deformable offset and modulation mask.

[0069] It should be understood that the purpose of deformable offsets is to adaptively adjust the sampling position. Traditional convolution's sampling position is fixed to a regular grid. Deformable offsets are equivalent to adding a learnable offset vector to each sampling point, allowing it to move to a more suitable location. For pipeline targets, deformable offsets enable sampling points to undergo a serpentine offset along the pipeline's direction: when the pipeline is horizontal, the sampling point is stretched horizontally; when the pipeline is vertical, the sampling point is stretched vertically; when the pipeline is curved, the sampling point follows the same curved trajectory as the pipeline, thus always sampling close to the pipeline's centerline.

[0070] It should be understood that the function of the modulation mask is to reweight the features extracted from the sampling points. Sampling points located on the pipeline centerline receive higher weights (close to 1), while sampling points located in background areas or on interfering objects (such as the edges of storage tanks or towers) receive lower weights (close to 0), thereby highlighting the key features of the pipeline area and suppressing background interference.

[0071] Through the above design, the StripDCN module can adaptively sample and enhance features along the pipeline direction, improving the feature response capability for slender pipelines, curved pipelines and small branch pipes, and effectively solving the problems of low extraction accuracy of extremely fine linear targets and easy missed detection of small branch pipes.

[0072] refer to Figure 3 The features transformed by the StripDCN module will be further processed through an additional... Deformable convolutions are used to aggregate broader contextual cues. The final output of the PDB module is given by the following formula: ; in, Indicates the first The feature map output from the decoding stage, i.e. the final output of the PDB module, has a higher spatial resolution than the input. This represents the bar convolution operation. This represents the activation function. Indicates batch normalization, express convolution, This indicates a channel splicing operation.

[0073] The number of output channels is set separately according to the decoding stage (corresponding to...). to The number of channels is 128, 64, 32, and 16 respectively. Finally, through a... Convolution will Projection is a single-channel coarse prediction image The core idea of ​​StripConv is to decouple a two-dimensional square convolution into two orthogonal one-dimensional strip convolutions, which independently encode the spatial context along the horizontal and vertical directions, respectively, thereby obtaining an anisotropic elongated receptive field without increasing the number of parameters. Given an input feature map... ∈ R C×H×W The operation process of StripConv is defined as follows: ; ; in, Indicates the input feature map; This represents the output feature map, which is the result after StripConv processing; This represents the number of input channels, i.e., the channel dimension of the feature map; This represents the number of output channels, i.e., the channel dimension after convolution; Indicates the feature map height. Indicates the width of the feature map. Indicates the strip length, such as ; This represents horizontal stripe convolution, where the kernel size is... Encode along the horizontal direction; This represents vertical stripe convolution, where the kernel size is [value missing]. Encode along the vertical direction; This represents the output of the horizontal branch, which extracts contextual features in the horizontal direction. This indicates the vertical branch output, which extracts contextual features in the vertical direction; This indicates element-wise addition.

[0074] The outputs of the two branches are merged element-wise by addition, rather than concatenating, to ensure the number of channels remains unchanged and reduce computational overhead. Comparison with standard convolution: one The standard convolution parameter count is The number of parameters for StripConv is only It achieves a compression ratio of 7.5 while retaining long-range dependence along the main direction.

[0075] This invention improves the detection rate of thin pipes, bends, and small branches by fitting a strip-shaped receptive field to the slender direction of the pipeline and combining deformable offset with a modulation mask to achieve adaptive sampling and feature reweighting, thereby reducing missed detections and false detections.

[0076] Based on the above embodiments, the cascaded refinement module is further configured as follows: An edge mask is generated based on the high-resolution pipeline feature map; The pipeline boundary is corrected based on the high-resolution pipeline feature map, the current mask estimate of the pipeline, and the edge mask.

[0077] It should be understood that the edge mask refers to a single-channel probability map generated by the cascaded refinement module based on the high-resolution pipeline feature map, represented as... Each pixel value represents the confidence level that the location belongs to the pipeline boundary. The edge mask is used to explicitly guide the model to focus on the boundary region during iterative correction, enabling the thinning head to clearly perceive the edge location and enhance the correction magnitude of the boundary region in the residual prediction.

[0078] It should be understood that the current mask estimate refers to the model's current prediction of the pipeline segmentation result at the current iteration correction stage, expressed as: ,in This represents the current iteration number. Initial time. The current mask estimate is a rough prediction. After each iteration, the current mask estimate is updated to... The current mask estimation gradually approaches the actual pipeline boundary during the iteration process, and is the target and refinement objective of the residual correction operation.

[0079] The cascaded refinement module receives the high-resolution pipeline feature map output by the pipeline decoder module, performs convolution processing on the feature map through a dedicated convolution head, and then maps the output value to the [0,1] interval through the Sigmoid activation function to generate a single-channel edge probability map with the same spatial resolution as the input feature map. Each pixel value in this edge mask represents the confidence that the position belongs to the pipeline boundary, which is used to explicitly guide the model to focus on the boundary region in subsequent iterations and strengthen the correction magnitude at the boundary.

[0080] For example, refer to Figure 3 The process of generating the edge mask includes the following steps: (1) Obtain high-resolution pipeline feature maps This feature map is output by the previous stage pipeline decoder module, and its spatial resolution is [missing information]. The number of channels is 32; (2) To Applying a transposed convolution with a stride of 2 (ConvTranspose2×2) to reduce its spatial resolution from Upgraded to ; (3) The upsampled feature maps are passed through a batch normalization layer (BN) to accelerate training convergence and stabilize the feature distribution; (4) Introduce nonlinearity into the batch-normalized feature map through an activation function (such as the SiLU activation function); (5) Pass the activated feature map through a 1×1 convolutional layer (Conv1×1) to compress the number of channels from 32 to 1 and output a single-channel feature map; (6) This single-channel feature map is the edge mask. Its spatial resolution is Each pixel value represents the confidence level that the location belongs to the pipeline boundary.

[0081] The cascaded refinement module integrates the high-resolution pipeline feature map, the current mask estimate, and the edge mask. Through a residual learning mechanism with multiple iterations, it gradually corrects the boundary region of the mask. In each iteration, the three are concatenated and input into the convolutional network, outputting a residual correction term, which is added to the current mask estimate. This allows the model to gradually adjust the prediction confidence in a linear superposition manner. Under the guidance of the edge mask, the correction magnitude of the boundary region is strengthened, and finally, a refined segmentation map with more accurate boundaries and higher alignment with the real pipeline boundaries is output.

[0082] This invention improves the positioning accuracy and alignment of pipeline boundaries by using an edge mask to guide the model to focus on the boundary region and combining iterative correction with multi-source information.

[0083] Based on the above embodiments, the cascaded refinement module is further configured as follows: The high-resolution pipeline feature map, the current mask estimate processed by the activation function, and the edge mask are stitched together in the channel dimension; The concatenated features are processed sequentially through convolutional layers to output residual correction terms. The residual correction term is added to the current mask estimate to obtain the updated mask estimate.

[0084] The residual correction term refers to a single-channel correction matrix generated by the cascaded refinement module in each iteration through convolution processing of the sequence from the concatenated multi-source features, denoted as: Each pixel value represents the amount of correction (positive or negative) that the current mask estimate needs to be adjusted at that location. This correction term is added pixel by pixel to the current mask estimate to update the mask, allowing the model to progressively refine the segmentation boundaries in a linear, cumulative manner. It should be understood that the residual correction term is designed based on the idea of ​​residual learning. In each iteration, the model only learns the difference between the mask estimate and the previous result (i.e., the residual), rather than directly predicting the entire mask. This design allows the model to focus on the shortcomings of the previous prediction for targeted correction, avoiding repeated calculations of correctly predicted regions.

[0085] In each iteration of the cascaded refinement module, the high-resolution pipeline feature map, the current mask estimate activated by Sigmoid, and the edge mask are concatenated along the channel direction to form a joint feature tensor. This tensor is then passed through a 3×3 convolutional layer, a batch normalization layer, a ReLU activation function, and a 1×1 convolutional layer to output a residual correction term of the same size as the current mask estimate. This residual correction term is then added to the current mask estimate pixel by pixel to update the mask estimate in a linear superposition manner, thereby achieving targeted correction of the previous prediction.

[0086] In one embodiment, although a rough prediction graph While capable of capturing the general structure of a target, its boundaries often lack precise alignment and sharpness. To address this, this invention proposes a Cascaded Refinement Module (CRM) that explicitly utilizes edge information to iteratively refine the prediction results. The core of the CRM is a shared-weight refinement head, whose structure is as follows: Figure 5 As shown.

[0087] CRM inputs include a rough prediction graph. Edge mask and high-resolution pipeline feature maps from the decoder . refer to Figure 3 , Through Definitely, for example, for Applying a transposed convolution with a stride of 2 (ConvTranspose2×2) to reduce its spatial resolution from Upgraded to ,get .

[0088] refer to Figure 3 , The steps for determining this are as follows: (1) To Applying a transposed convolution with a stride of 2 (ConvTranspose2×2) to reduce its spatial resolution from Upgraded to ; (2) The upsampled feature maps are passed through a batch normalization layer (BN) to accelerate training convergence and stabilize the feature distribution; (3) Introduce nonlinearity into the batch-normalized feature map through an activation function (such as the ReLU activation function); (4) Pass the activated feature map through a 1×1 convolutional layer (Conv1×1) to compress the number of channels from 16 to 1 and output a single-channel feature map; (5) This single-channel feature map is the coarse prediction map. Its spatial resolution is .

[0089] The refinement process consists of three cascaded steps, each using the same refinement head structure, which is applied to progressively updated mask estimates.

[0090] make For the initial coarse mask, in the first... In the process of refinement, refine the head. Receive high-resolution pipeline feature map and current mask estimation It then outputs a residual correction term. Internally, the thinning head first applies a Sigmoid activation to the input mask to normalize it to... Intervals, and then, combine them with features. By splicing along the channel dimension, we obtain a [structure / object] with [specific characteristics]. The channel tensor. This splicing feature passes sequentially through a... Convolutional layers (keeping the number of channels at 16), batch normalization layers, and ReLU nonlinear activation functions, finally passed through a... Convolutional layers compress the channels to 1, resulting in residual predictions. The entire thinning head operation can be formally represented as: ; in, This represents the Sigmoid function.

[0091] Residual correction term Compared with current mask estimation The updated mask estimate is obtained by adding pixels one by one. : ; The residual correction term can be positive or negative. A positive value indicates that the predicted value needs to be increased at that location (expanding towards the pipeline), a negative value indicates that the predicted value needs to be decreased (shrinking inward), and zero indicates that no adjustment is needed. This allows the model to make targeted corrections only to the boundary deviation areas of the previous prediction, rather than re-predicting the entire mask.

[0092] For example, the mask estimate can be updated gradually through three iterations: First iteration: Correct the overall outline; Second iteration: Correct the branch structure boundaries; Third iteration: Refine pixel-level boundaries.

[0093] Through the above mechanism, only the residual correction term is learned in each iteration. Instead of re-predicting the entire mask, the residual correction term focuses on the shortcomings of the previous mask estimate (such as boundary offsets, punctures, breaks, etc.), and gradually adjusts the prediction confidence in a linear superposition manner. After multiple iterations (e.g., 3 times), the mask estimate is gradually refined from an initial coarse prediction to a segmentation result with accurate boundaries.

[0094] It should be noted that before splicing, Applying sigmoid activation is intended to map unnormalized logits to a probability space, thereby connecting them to high-resolution pipeline feature maps. The numerical scale is matched; and the residual update acts directly on the original logits space, allowing the model to gradually adjust the prediction confidence in a linear superposition manner. The entire refinement head shares weights in the three iterations, which not only reduces the number of parameters but also prompts the model to learn a general boundary correction strategy that is independent of the initial state.

[0095] In addition, to further enhance boundary awareness, edge masks... It is explicitly introduced during the cascading process. In each stitching operation, in addition to the high-resolution pipeline feature map... In addition to the current mask estimation, the edge mask It is also concatenated to the channel dimension, thus forming an 18-channel joint input (corresponding to...). This design enables the refinement head to clearly perceive edge locations and enhance the correction magnitude of boundary regions in residual prediction.

[0096] After three cascaded stages, the final output is... It combines high spatial accuracy with clear target boundaries. The entire CRM module is an end-to-end trainable structure, and it adds only a minimal amount of computational overhead compared to the backbone network inference process.

[0097] The embodiments of the present invention utilize a residual learning mechanism to correct only the boundary deviation of the previous mask in each iteration, avoiding repeated calculations. This process gradually refines the boundary while maintaining structural stability, thereby improving correction efficiency and positioning accuracy.

[0098] To further explain the method for extracting pipelines in chemical industrial parks proposed in this invention, please refer to the following embodiments.

[0099] To comprehensively verify the effectiveness and advancement of the proposed method in the task of segmenting facilities in chemical industrial parks, comparative experiments were conducted using several representative mainstream models and advanced methods in the current semantic segmentation field. The models compared included: the classic U-Net, D-LinkNet for remote sensing road extraction, DeepLabv3+ based on dilated convolution, Swin-Unet combined with Transformer, RCFSNet for remote sensing change detection, SGCN guided by graph convolution, CMTFNet for multi-task fusion, RS3Mamba based on a state-space model, and MADSNet for multi-scale attention. All comparative models were retrained and evaluated using the same dataset partitioning and training strategy as this invention to ensure fairness in the comparison.

[0100] Table 1 shows the comparison results of each model on the test set for six evaluation metrics.

[0101] Table 1 As shown in Table 1, PDSDNet achieved precision and recall of 0.8866 and 0.8678, respectively, representing improvements of approximately 1.9% and 2.8% compared to other models such as DeepLabv3+ and D-LinkNet. It effectively suppressed false alarms while providing more comprehensive pipeline coverage, achieving a good balance between precision and recall. The F-Score, reflecting overall performance, reached 0.8771, 0.0237 higher than D-LinkNet's 0.8534, validating the significant gains in pipeline feature extraction quality brought about by the pipeline decoding structure and cascaded refinement modules.

[0102] PDSDNet achieves foreground cross-union ratios (CUI) and average cross-union ratios (CUI) of 0.7812 and 0.8901, respectively, both significantly outperforming the comparison methods. Specifically, compared to D-LinkNet (FIoU of 0.7443), PDSDNet improves FIoU by 0.0369, demonstrating its superior ability to characterize the localization accuracy and shape integrity of chemical facility targets. On the ALPS metric, which measures boundary fit, PDSDNet reaches 0.7920, far exceeding all other comparison models. This demonstrates the edge mask guidance mechanism and StripDCN's sensitive modeling of strip structures in boundary reconstruction under complex backgrounds.

[0103] To further and more intuitively compare the segmentation performance of different models, Figure 6 The segmentation mask and error distribution of each model on typical test samples are shown. Figure 6The system adopts a unified visualization standard: white areas represent true positives (TP), which are correctly predicted foreground pixels; red areas represent false positives (FP), which are background pixels that are mistakenly detected as foreground pixels; green areas represent false negatives (FN), which are foreground pixels that are missed; and black areas represent true negatives (TN), which are correctly predicted background pixels.

[0104] As the visualization results show, U-Net's segmentation results contain a large number of green and red areas, indicating that its false negatives and false negatives are quite serious, making it difficult to cope with the densely packed facilities and complex background interference in chemical industrial parks. D-LinkNet and DeepLabv3+ show a significant reduction in FN regions, but there are still wide red FP bands at facility boundaries, indicating that their ability to finely locate edges is limited. Although Swin-Unet and MADSNet show some improvement in overall shape, they are still insufficient in capturing small pipes and auxiliary structures.

[0105] In contrast, the PDSDNet proposed in this invention exhibits the fewest red and green areas in the visualization results, with the white TP areas closely matching the actual annotations, especially at the sharp corners and slender connections of the facilities, where the boundaries are cleaner and sharper. This result, coupled with the quantitative advantages of the ALPS index, fully demonstrates the effectiveness of the cascaded refinement module and pipeline decoder module in boundary awareness and structure preservation.

[0106] The following describes the pipeline extraction device for chemical industrial parks provided by the present invention. The pipeline extraction device for chemical industrial parks described below can be referred to in correspondence with the pipeline extraction method for chemical industrial parks described above.

[0107] refer to Figure 7 The chemical industrial park pipeline extraction device provided by the present invention includes: The acquisition module 701 is used to acquire remote sensing images of the chemical industrial park to be processed. The pipeline extraction module 702 is used to input the remote sensing image of the chemical industrial park into the pipeline extraction model and obtain the pipeline segmentation map of the chemical industrial park output by the pipeline extraction model. The pipeline extraction model includes: Multiple pipeline decoder modules with striped receptive fields are configured to progressively upsample and fuse multi-level feature maps of the remote sensing image of the chemical industrial park to enhance the feature response to pipeline orientation and output high-resolution pipeline feature maps. The cascaded refinement module is configured to correct the pipeline boundary based on the high-resolution pipeline feature map and the current mask estimation of the pipeline, and output a pipeline segmentation map with refined boundary.

[0108] The pipeline extraction device for chemical industrial parks provided in this invention, by setting up a pipeline decoder module with a strip-shaped receptive field, can enhance the characteristics of the slender pipeline, effectively solving the problems of low extraction accuracy of extremely fine linear targets and easy omission of small branch pipes; at the same time, by setting up a cascaded refinement module, the pipeline boundary is iteratively corrected based on the high-resolution feature map and the current mask estimation, which improves the positioning accuracy of the segmentation boundary and reduces the false detection rate caused by the similarity between pipeline and background features in complex industrial backgrounds such as storage tanks and towers, thereby achieving high-precision extraction of pipelines in chemical industrial parks.

[0109] In one embodiment, each pipeline decoder module receives a feature map from a multi-scale feature pyramid and a feature map output from the previous decoding stage, with the output spatial resolution higher than its input feature map.

[0110] In one embodiment, the pipeline decoder module includes a strip deformable convolutional unit, which generates a strip receptive field through cascaded 1×N and N×1 convolutions, and combines deformable offsets and modulation masks to adaptively adjust the sampling position and reweight features.

[0111] In one embodiment, the cascaded refinement module is further configured as follows: An edge mask is generated based on the high-resolution pipeline feature map; The pipeline boundary is corrected based on the high-resolution pipeline feature map, the current mask estimate of the pipeline, and the edge mask.

[0112] In one embodiment, the cascaded refinement module is further configured as follows: The high-resolution pipeline feature map, the current mask estimate processed by the activation function, and the edge mask are stitched together in the channel dimension; The concatenated features are processed sequentially through convolutional layers to output residual correction terms. The residual correction term is added to the current mask estimate to obtain the updated mask estimate.

[0113] In one embodiment, the pipeline extraction model further includes a backbone network module for forward propagating the input remote sensing image of the chemical industrial park to extract the multi-level feature map.

[0114] In one embodiment, the pipeline extraction model further includes a feature pyramid construction module, which is used to project and scale the multi-level feature maps to output a multi-scale feature pyramid with a uniform number of channels and a spatial scale that is multiple of each other.

[0115] Figure 8 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 8As shown, the electronic device may include: a processor 810, a communication interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communication interface 820, and the memory 830 communicate with each other through the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute a pipeline extraction method for a chemical industrial park. The method includes: acquiring a remote sensing image of the chemical industrial park to be processed; inputting the remote sensing image of the chemical industrial park into a pipeline extraction model, and acquiring a pipeline segmentation map of the chemical industrial park output by the pipeline extraction model; wherein the pipeline extraction model includes: multiple pipeline decoder modules with strip receptive fields, configured to progressively upsample and fuse multi-level feature maps of the remote sensing image of the chemical industrial park to enhance the feature response to pipeline orientation and output a high-resolution pipeline feature map; and a cascaded refinement module, configured to correct the pipeline boundary based on the high-resolution pipeline feature map and the current mask estimation of the pipeline, and output a refined pipeline segmentation map.

[0116] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0117] On the other hand, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method for extracting pipelines in chemical industrial parks provided by the above methods. This method includes: acquiring a remote sensing image of a chemical industrial park to be processed; inputting the remote sensing image of the chemical industrial park into a pipeline extraction model to acquire a pipeline segmentation map of the chemical industrial park output by the pipeline extraction model; wherein the pipeline extraction model includes: multiple pipeline decoder modules with striped receptive fields, configured to progressively upsample and fuse multi-level feature maps of the remote sensing image of the chemical industrial park to enhance the feature response to pipeline orientation and output a high-resolution pipeline feature map; and a cascaded refinement module configured to correct the pipeline boundaries based on the high-resolution pipeline feature map and the current mask estimation of the pipeline, and output a refined pipeline segmentation map.

[0118] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0119] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for extracting pipelines in a chemical industrial park, characterized in that, include: Acquire remote sensing images of the chemical industrial park to be processed; The remote sensing image of the chemical industrial park is input into the pipeline extraction model to obtain the pipeline segmentation map of the chemical industrial park output by the pipeline extraction model. The pipeline extraction model includes: Multiple pipeline decoder modules with striped receptive fields are configured to progressively upsample and fuse multi-level feature maps of the remote sensing image of the chemical industrial park to enhance the feature response to pipeline orientation and output high-resolution pipeline feature maps. The cascaded refinement module is configured to correct the pipeline boundary based on the high-resolution pipeline feature map and the current mask estimation of the pipeline, and output a pipeline segmentation map with refined boundary.

2. The method for extracting pipelines in a chemical industrial park according to claim 1, characterized in that, Each of the pipeline decoder modules receives a feature map from a multi-scale feature pyramid and a feature map output from the previous decoding stage, with the output spatial resolution higher than its input feature map.

3. The method for extracting pipelines in a chemical industrial park according to claim 1, characterized in that, The pipeline decoder module includes a strip deformable convolutional unit. The strip deformable convolutional unit generates a strip receptive field through cascaded 1×N and N×1 convolutions, and combines deformable offsets and modulation masks to adaptively adjust the sampling position and reweight features.

4. The method for extracting pipelines in a chemical industrial park according to claim 1, characterized in that, The cascaded refinement module is further configured as follows: An edge mask is generated based on the high-resolution pipeline feature map; The pipeline boundary is corrected based on the high-resolution pipeline feature map, the current mask estimate of the pipeline, and the edge mask.

5. The method for extracting pipelines in a chemical industrial park according to claim 4, characterized in that, The cascaded refinement module is further configured as follows: The high-resolution pipeline feature map, the current mask estimate processed by the activation function, and the edge mask are stitched together in the channel dimension; The concatenated features are processed sequentially through convolutional layers to output residual correction terms. The residual correction term is added to the current mask estimate to obtain the updated mask estimate.

6. The method for extracting pipelines in a chemical industrial park according to claim 1, characterized in that, The pipeline extraction model also includes a backbone network module, which is used to perform forward propagation on the input remote sensing image of the chemical industrial park to extract the multi-level feature map.

7. The method for extracting pipelines in a chemical industrial park according to claim 1, characterized in that, The pipeline extraction model also includes a feature pyramid construction module, which is used to project and scale the multi-level feature maps to output a multi-scale feature pyramid with a uniform number of channels and a spatial scale that is multiple of each other.

8. A pipeline extraction device for a chemical industrial park, characterized in that, include: The acquisition module is used to acquire remote sensing images of the chemical industrial park to be processed; The pipeline extraction module is used to input the remote sensing image of the chemical industrial park into the pipeline extraction model and obtain the pipeline segmentation map of the chemical industrial park output by the pipeline extraction model. The pipeline extraction model includes: Multiple pipeline decoder modules with striped receptive fields are configured to progressively upsample and fuse multi-level feature maps of the remote sensing image of the chemical industrial park to enhance the feature response to pipeline orientation and output high-resolution pipeline feature maps. The cascaded refinement module is configured to correct the pipeline boundary based on the high-resolution pipeline feature map and the current mask estimation of the pipeline, and output a pipeline segmentation map with refined boundary.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the method for extracting pipelines in a chemical industrial park as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for extracting pipelines in chemical industrial parks as described in any one of claims 1 to 7.

Images