Remote sensing significant target detection method and device, computer device and storage medium

Abstract

This invention discloses a method, apparatus, computer equipment, and storage medium for detecting salient targets in remote sensing. The method includes: acquiring an optical remote sensing image to be detected; inputting the image into a target detection model for target detection, including: constructing a multi-scale feature pyramid from the optical remote sensing image to be detected and mapping it to a discrete feature field to be projected; using a Radon projection state-space encoder to perform anisotropic decoupling and sequence modeling on the discrete feature field to be projected, and generating geometrically enhanced features through inverse transformation; using a hysteresis-gated dynamics fusion module to modulate the fusion path between deep semantics and shallow details to generate a predicted saliency map; calculating the topological difference between the predicted saliency map and the ground truth map, and performing end-to-end optimization using a hybrid loss function. By implementing the method of this invention, problems such as structural fragmentation, edge blurring, topological breaks, and background noise interference encountered by traditional methods when processing complex remote sensing images can be effectively solved.

Description

Technical Field

[0001] This invention relates to computer vision, and more specifically to remote sensing salient target detection methods, apparatus, computer equipment, and storage media. Background Technology

[0002] With the continuous improvement of imaging resolution, accurately separating targets from complex geographical backgrounds has become a core challenge in the field of optical remote sensing saliency detection. Although existing deep learning methods, through convolutional neural networks or visual Transformers, have made progress in this field to some extent, these methods still exhibit fundamental limitations when faced with the unique geometric heterogeneity and environmental uncertainty of remote sensing images. Traditional convolution operators, limited by their local receptive field, struggle to capture global geometric topologies with clear directions, such as long-span bridges, rivers, or tilted ships, often resulting in detection results exhibiting severe structural fragmentation.

[0003] Although the Transformer attempts to establish long-range dependencies through self-attention, its high computational cost and insufficient extraction of local details make it prone to edge blurring or topological breaks when dealing with slender targets. More seriously, cloud cover, abrupt changes in illumination, and ground clutter, prevalent in remote sensing images, introduce significant transient noise into the feature fusion stage. Current feature fusion paradigms mostly rely on simple channel stacking or weighted summation, lacking dynamic modeling of signal evolution. This makes saliency signals highly susceptible to background noise interference and oscillations during the backpropagation from deep semantics to shallow details, failing to maintain the consistency of connectivity and saliency within the target.

[0004] Therefore, it is necessary to design a new method that can effectively solve the problems encountered by traditional methods when processing complex remote sensing images, such as structural fragmentation, edge blurring, topological breaks, and background noise interference, so as to improve detection accuracy and robustness. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, apparatus, computer equipment and storage medium for detecting prominent targets in remote sensing.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a remote sensing salient target detection method, comprising:

[0007] Acquire the optical remote sensing image to be detected;

[0008] The optical remote sensing image to be detected is input into a target detection model for target detection to obtain a salient target mask, which includes:

[0009] A multi-scale feature pyramid is constructed for the optical remote sensing image to be detected, and the multi-scale feature pyramid is mapped to a discrete feature field to be projected.

[0010] The discrete feature field to be projected is anisotropically decoupled and sequence modeled using a Radon projection state space encoder, and then geometrically enhanced features are generated through inverse transformation.

[0011] Based on the aforementioned geometric enhancement features, a hysteresis-gated dynamics fusion module is used to modulate the fusion path between deep semantics and shallow details, filter out background noise, and generate a prediction saliency map.

[0012] The topological difference between the predicted saliency map and the ground truth map is calculated, and end-to-end optimization is performed using a hybrid loss function to obtain the salient target mask.

[0013] The present invention also provides a remote sensing salient target detection device, comprising:

[0014] The acquisition unit is used to acquire the optical remote sensing image to be detected;

[0015] The detection unit is used to input the optical remote sensing image to be detected into the target detection model for target detection, so as to obtain a salient target mask;

[0016] The detection unit includes:

[0017] The mapping subunit is used to construct a multi-scale feature pyramid for the optical remote sensing image to be detected, and to map the multi-scale feature pyramid into a discrete feature field to be projected.

[0018] The projection coding subunit is used to perform anisotropic decoupling and sequence modeling on the discrete feature field to be projected using the Radon projection state space encoder, and then generate geometrically enhanced features through inverse transformation;

[0019] An enhancement subunit is used to modulate the fusion path of deep semantics and shallow details based on the geometric enhancement features using a hysteresis-gated dynamics fusion module, filter out background noise, and generate a prediction saliency map.

[0020] The computational subunit is used to calculate the topological difference between the predicted saliency map and the ground truth map, and performs end-to-end optimization using a hybrid loss function to obtain the salient target mask.

[0021] The present invention also provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the above-described method.

[0022] The present invention also provides a storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0023] The advantages of this invention compared to existing technologies are as follows: This invention constructs a multi-scale feature pyramid and maps these features to a discrete feature field to be projected. It then utilizes a Radon projection state-space encoder to achieve anisotropic decoupling and sequence modeling, thereby geometrically repairing the feature continuity problem of slender and tilted targets. Next, a hysteresis-gated dynamic fusion module modulates the fusion path between deep semantics and shallow details, effectively filtering background noise and maintaining the consistency of target saliency, solving problems such as structural fragmentation and edge blurring in traditional methods. Finally, by calculating the topological difference between the predicted saliency map and the ground truth map and using a hybrid loss function including topological consistency loss for end-to-end optimization, the output results maintain high accuracy and topological integrity even in complex backgrounds, greatly improving the system's noise robustness and detection accuracy. This method achieves deep fusion from geometric decoupling and dynamic filtering to global topological constraints, effectively overcoming various challenges encountered by traditional technologies in processing complex remote sensing images.

[0024] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 A schematic flowchart of the remote sensing salient target detection method provided in an embodiment of the present invention;

[0027] Figure 2 This is a schematic diagram of the network architecture of the target detection model provided in an embodiment of the present invention;

[0028] Figure 3 This is a schematic diagram of the projection decoupling and reconstruction logic of the Radon projection state space encoder provided in an embodiment of the present invention;

[0029] Figure 4 Microscopic evolution diagram for anisotropic state space modeling provided in embodiments of the present invention;

[0030] Figure 5 A schematic diagram of the projection path of the discrete Radon transform at different angles and a geometric interpretation diagram of the mask sine diagram filling strategy provided in the embodiments of the present invention;

[0031] Figure 6 The persistent homology evolution diagram of topological consistency constraints provided in the embodiments of the present invention;

[0032] Figure 7 This is a visualization comparison of the results of this invention with other methods in different scenarios, provided by the embodiments of the present invention.

[0033] Figure 8 A schematic block diagram of a remote sensing salient target detection device provided in an embodiment of the present invention;

[0034] Figure 9 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0037] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0038] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0039] Please see Figure 1 , Figure 1This is a flowchart illustrating the remote sensing salient target detection method provided in an embodiment of the present invention. This method is applied in a server. The method constructs a feature pyramid through a multi-scale feature extraction layer and utilizes a Radon projection state-space encoder to anisotropically decouple and sequentially model the feature field. An inverse transformation is then performed to generate geometrically enhanced features, thereby enhancing the model's ability to perceive information from different directions. Next, a hysteresis-gated dynamics fusion module modulates the fusion path between deep semantics and shallow details, effectively filtering out background noise and generating a predicted saliency map. Finally, by calculating the topological difference between the predicted saliency map and the ground truth map, end-to-end optimization is performed using a hybrid loss function to ensure the connectivity and structural integrity of the target. This method effectively solves the problems encountered by traditional methods when processing complex remote sensing images, such as structural fragmentation, edge blurring, topological breaks, and background noise interference, thereby improving detection accuracy and robustness.

[0040] Figure 1 This is a schematic flowchart of the remote sensing salient target detection method provided in an embodiment of the present invention. Figure 1 As shown, the method includes the following steps S110 to S120.

[0041] S110. Acquire the optical remote sensing image to be detected.

[0042] In this embodiment, the optical remote sensing image to be detected refers to an appropriate image selected from a series of available high-resolution optical remote sensing image datasets as the processing object. These images are typically captured by sensors carried by satellites or high-altitude aircraft, and have high spatial resolution and rich spectral information.

[0043] Specifically, this process involves the following aspects:

[0044] First, determine which data source to use. This could include public remote sensing databases, services provided by commercial satellite imagery providers, or custom-acquired datasets for a specific project. For example, Landsat and Sentinel-2 are common publicly available remote sensing image resources.

[0045] The selected remote sensing images undergo necessary preprocessing to ensure they are suitable for subsequent analysis. This may include steps such as radiometric correction (eliminating the effects of atmospheric conditions), geometric correction (correcting distortions caused by topography and the curvature of the Earth), cropping (extracting the area of ​​interest), and resampling (adjusting the spatial resolution of the image).

[0046] Constructing a multi-scale feature pyramid using a feature extraction network. This means using deep learning models such as the Pyramid Vision Transformer (PVT) or convolutional neural networks (e.g., ResNet) to compress and extract semantic information from different levels of the input image layer by layer, thus forming a pyramid structure containing feature maps at multiple scales. This multi-scale representation is crucial for capturing different levels of detail of the target and helps improve the ability of detection algorithms to adapt to targets of varying sizes.

[0047] Finally, the multi-scale feature maps obtained in the above process are mapped to discrete feature fields to be projected, so that the anisotropic decoupling and sequence modeling can be performed in the next step using the Radon projection state-space encoder. This step involves transforming the original two-dimensional image data into a form more suitable for subsequent processing, so as to effectively capture the directional and structural information of the target.

[0048] In summary, step S110 is not limited to simple image acquisition; it is a process that integrates data selection, preprocessing, and feature engineering, aiming to prepare high-quality input data for subsequent object detection tasks.

[0049] S120. Input the optical remote sensing image to be detected into the target detection model to perform target detection and obtain a salient target mask.

[0050] In this embodiment, the salient target mask refers to the binary image generated after processing by the target detection model, in which regions belonging to salient targets (represented by 1) and regions not belonging to salient targets (represented by 0) are marked, so as to accurately identify and segment important objects in the image.

[0051] The target detection model includes a multi-scale feature extraction layer, a Radon projection state-space encoder, a hysteresis-gated dynamics fusion module, and a prediction head.

[0052] In a specific embodiment of the present invention, step S120 describes the process of inputting the optical remote sensing image to be detected into a target detection model for processing to obtain a salient target mask. This process involves the coordinated work of multiple stages and components, including a multi-scale feature extraction layer, a Radon projection state-space encoder, a hysteresis-gated dynamics fusion module, and a prediction head.

[0053] First, the input optical remote sensing image passes through a multi-scale feature extraction layer. This is typically implemented using deep learning architectures such as the Pyramid Vision Transformer (PVT) or convolutional neural networks (e.g., ResNet). Through this layer, the original image is transformed into a feature pyramid containing different resolution levels. Each layer represents semantic information at a different scale of the image, enabling the model to simultaneously capture broad contextual information as well as fine-grained detailed features.

[0054] Next, the deep feature maps output from the multi-scale feature extraction layer (typically the highest-level semantic and lowest-resolution feature maps) are fed into the Radon projection state-space encoder. Within this encoder, the two-dimensional feature field is decoupled along multiple preset directions using the discrete Radon transform operator, transforming it into a series of one-dimensional projection sequences. Global long-range dependencies are captured through an anisotropic parameterized selective state-space model, and finally, this directional information is aggregated through an inverse transform operation to generate geometrically enhanced features.

[0055] Then, the geometrically enhanced features are passed as the initial dynamic state to the hysteresis-gated dynamics fusion module. This module optimizes the feature fusion process using the hysteresis effect from nonlinear control theory, simulating physical hysteresis, and modulates the fusion path between deep semantics and shallow details through a dual-threshold switching mechanism. This mechanism effectively suppresses background noise and maintains the consistency of target saliency. Each layer's hysteresis-gated dynamics fusion module uses the state variables output from the previous layer as a reference, recursively downsampling until the appropriate resolution of the original image is restored.

[0056] Finally, the topologically complete feature map, after multiple fusions, is fed into the prediction head to generate a saliency probability map. This probability map represents the probability that each pixel belongs to a salient object. To further improve the accuracy and structural integrity of the results, the predicted map is compared with the ground truth map. Figure 1 The process begins with topological consistency constraint optimization based on persistent cohomology theory, calculating the differences in topological features and constructing a hybrid loss function for end-to-end optimization.

[0057] A salient target mask is a binary image generated after all the above processing steps, where each pixel is labeled as belonging to a salient target (usually represented by 1) or not belonging to a salient target (usually represented by 0). This mask not only accurately marks the salient target regions in the image, but also ensures the topological connectivity and edge smoothness of these regions, thereby improving the robustness and accuracy of the detection results.

[0058] In summary, step S120 is not merely a simple input-output process; it is a complex and meticulously designed data processing flow aimed at maximizing the accuracy, completeness, and noise resistance of salient target detection in remote sensing images. By combining geometric decoupling, dynamic filtering, and topological global constraints, this embodiment provides a novel perspective and technical approach to address the problems inherent in traditional methods.

[0059] Please see Figure 2 By deeply embedding geometric perception and nonlinear control theory into the feature transfer process, precise target locking of remote sensing is achieved. The entire detection process begins with multi-scale feature extraction of an input optical remote sensing image with dimensions H×W×3. Using a Pyramid Vision Transformer (PVT) or a Convolutional Neural Network (ResNet) as the backbone network, semantic information is extracted layer by layer in the spatial dimension, thereby generating a sequence of features from shallow to deep with resolutions varying from the original. Figure 1 The four-level feature pyramid of / 4, 1 / 8, 1 / 16, and 1 / 32 Among them, the feature map represents the highest level of semantics and has the lowest resolution. First, the input signal is passed to the Radon projection state space. Inside the encoder, the discrete Radon transform operator decouples the two-dimensional feature manifold along multiple preset projection directions into a one-dimensional sequence set, capturing global long-range dependencies in an anisotropic parameterized sequence modeling space. Then, the inverse transform operator re-aggregates these directional information, ultimately outputting enhanced features with strong geometric awareness. Following this step, The initial momentum as the dynamic state, and the sub-deep features from the backbone network. They all converge in the first hysteresis-gated dynamics fusion module. This module utilizes... The deep semantic momentum contained within The characteristic flow direction is modulated, and the insignificant signals generated by background clutter are suppressed through a dual threshold switching mechanism to generate the fused intermediate layer dynamic state variables. This fusion mechanism exhibits a recursive downward trend in the decoding stage, where the output of each hysteresis-gated dynamic fusion module serves as the state reference for the next higher-resolution module, and so on. and Multi-scale feature fusion is performed until the spatial resolution is restored to 1 / 4 of the original image. Finally, the topologically complete feature map output by the last decoding module is fed into the prediction head to generate a saliency probability map. To ensure the structural accuracy of the final mask, this probability map is compared with the ground truth. Figure 1The algorithm then enters the topology consistency constraint optimization stage. By calculating the persistent cohomology feature differences under the sub-level set filtering, a hybrid loss function containing the topological Wasserstein distance is constructed. This allows for end-to-end closed-loop fine-tuning of all model parameters from the backbone network to the decoder, outputting a high-precision and topologically connected salient target mask.

[0060] In one embodiment, step S120 may include steps S121 to S124.

[0061] S121. Construct a multi-scale feature pyramid for the optical remote sensing image to be detected, and map the multi-scale feature pyramid into a discrete feature field to be projected.

[0062] In this embodiment, multi-scale feature pyramids are a common deep learning architecture strategy used to extract features at different resolutions and levels of abstraction from input images. Typically, the spatial resolution of the feature maps is gradually reduced by downsampling (e.g., using convolutional layers, pooling layers) while increasing their number of channels to capture higher-level semantic information.

[0063] The discrete feature field to be projected refers to the feature representations at different scales obtained through the above process. These feature representations can be regarded as "feature maps" of the original image at multiple scales, and each map contains important information at the corresponding scale.

[0064] Specifically, the optical remote sensing image to be detected is processed by a multi-scale feature extraction layer to extract semantic information at different levels by compressing layer by layer, generating a multi-scale feature pyramid, and mapping the multi-scale feature pyramid to a discrete feature field to be projected. The multi-scale feature extraction layer includes a pyramid visual Transformer or a convolutional neural network as the backbone network.

[0065] In this embodiment, firstly, the optical remote sensing image to be detected is acquired as input. To effectively capture multi-level information in the image, an appropriate backbone network is selected to construct a multi-scale feature pyramid. In this case, a Pyramid Visual Transformer (PVT) or a traditional convolutional neural network (such as ResNet) can be chosen as the basic architecture. Through the outputs of different layers of these networks, a series of feature maps with different resolutions and semantic levels can be obtained, forming a feature pyramid.

[0066] Low-level features contain rich spatial details but weak semantic information.

[0067] High-level features: rich in semantic information, but lack detailed information due to low spatial resolution.

[0068] This multi-scale feature pyramid can effectively combine information from different levels, which is crucial for subsequent tasks such as object detection.

[0069] Once the multi-scale feature pyramid is obtained, the next step is to map these feature maps into a form suitable for subsequent processing, known as a "discrete feature field." This process typically includes the following aspects:

[0070] Feature aggregation: Sometimes, features at different scales are upsampled or downsampled, and then added or concatenated element by element to generate new feature representations.

[0071] Feature encoding: This may involve applying additional convolutional layers or other types of transformations (such as non-linear activation functions) to further enhance the expressiveness of the features.

[0072] Preparing for projection: Ensure that the final feature representation is suitable for operations such as the Radon transform mentioned in subsequent steps. This means that the features should be organized in a way that allows them to be "projected" or analyzed along a specific direction.

[0073] In summary, step S121, by constructing a multi-scale feature pyramid from the input remote sensing image and mapping it to a discrete feature field to be projected, lays the foundation for subsequent in-depth feature analysis using the Radon projection state-space encoder. This process not only helps improve the model's understanding of complex scenes but also enhances the accuracy of recognizing targets of different sizes and shapes.

[0074] S122. The discrete feature field to be projected is anisotropically decoupled and sequence modeled using the Radon project state space encoder (RPSS), and then geometrically enhanced features are generated through inverse transformation.

[0075] In this embodiment, geometrically enhanced features refer to two-dimensional feature maps with strong geometric perception and topological integrity generated by performing a series of processes such as Radon transform, sequence modeling, and inverse transform on the original feature map.

[0076] In one embodiment, step S122 described above may include steps S1221 to S1225.

[0077] S1221. Select a specific set of angles to perform Radon transform on the discrete feature field to be projected, and generate a sequence of sine graphs representing information in different directions.

[0078] In this embodiment, the sine map sequence refers to the set of one-dimensional projection sequences obtained by performing line integration on the input feature map along a specific angle through discrete Radon transform, which is used to represent feature information in different directions.

[0079] First, define a set of preset angles Θ={ For a given deep feature map Performing the Discrete Radon Transform (DRT) along these angles transforms the two-dimensional features into a set of one-dimensional sine wave sequences. The sine curve at each angle represents the line integral response of the characteristic tensor in that direction.

[0080] S1222. Adjust all the sine sequence to the same length by zero padding and create a validity mask.

[0081] In this embodiment, the validity mask is a matrix that identifies the actual signal region. It is used to ensure that only the part containing the valid signal participates in the calculation during the standardization process and serves as a spatial constraint for subsequent parameter adjustments.

[0082] Since the projection path lengths may vary at different angles, leading to calculation distortion, it is necessary to perform zero-padding on all sine wave sequences to unify their lengths to Lmax, and simultaneously generate a set of validity mask matrices. It is used to identify the actual signal area and serve as a spatial constraint for subsequent parameter adjustments.

[0083] S1223. Based on the adaptive parameter mapping and the discretization of the continuous system, the effective mask is guided by recursive state evolution to evolve the standardized projection sequence in the hidden state space to obtain an enhanced one-dimensional state sequence.

[0084] In this embodiment, the enhanced one-dimensional state sequence is generated in the anisotropic selective state space model after adaptive parameter mapping and recursive state evolution, and is a one-dimensional feature sequence that can capture long-distance dependencies and repair structural fractures.

[0085] At this stage, the standardized projection sequence Its corresponding validity mask The input is fed into an anisotropic selective state-space model. The system generates a discretized step size for each projection angle through an adaptive parameter mapping process. Then, using this step size, a zero-order preserve transformation is performed on the continuous state matrix to obtain a discretized transition matrix specific to the current projection direction. and Then, these matrices are used to guide the normalized projection sequence. Evolving in the implicit state space, it generates an enhanced one-dimensional state sequence. .

[0086] S1224. The enhanced one-dimensional state sequence is remapped to a two-dimensional space using the inverse Radon transform, and attention weights are dynamically allocated according to the information entropy of each direction to perform feature aggregation, so as to obtain multi-angle features.

[0087] In this embodiment, multi-angle features refer to the feature map obtained by remapping one-dimensional state sequences from multiple different angles back to two-dimensional space through inverse Radon transform, and then dynamically allocating attention weights according to the information entropy of each direction.

[0088] Next, the enhanced one-dimensional state sequence By remapping back to two-dimensional space through the inverse Radon transform, component features in multiple directions are formed. In this process, to achieve selective information enhancement, attention weights are dynamically allocated based on the information entropy of each projection direction. These weights are then used to perform nonlinear aggregation on the reconstructed multi-angle features.

[0089] In one embodiment, step S1224 described above may include steps S12241 to S12243.

[0090] S12241. Generate an adaptive discretization step size for each projection angle based on the position encoding and mask weighted pooling features of the enhanced one-dimensional state sequence;

[0091] S12242. Use an adaptive step size to transform the preset continuous state matrix to obtain a discretized transition matrix specific to the current projection direction.

[0092] In this embodiment, the discretized transition matrix is ​​a state transition matrix specific to the current projection direction, obtained by performing a zero-order preserve transformation on the continuous state matrix using an adaptive step size. It is used to guide the evolution of the standardized projection sequence in the implicit state space.

[0093] S12243. Based on the discretization matrix, guide the normalized projection sequence to evolve in the hidden state space to obtain an enhanced one-dimensional state sequence.

[0094] S1225. The multi-angle features are fused into an enhanced feature map with strong geometric perception capabilities through weighted summation and nonlinear mapping to obtain geometrically enhanced features.

[0095] The final step involves weighted summation of the multi-angle features and non-linear mapping using a 1×1 convolutional layer, thereby integrating them into a geometrically enhanced feature map F4' with strong geometric awareness. This feature map not only contains rich structural information but also enhances the connectivity and topological integrity of the target.

[0096] Through the above process, this embodiment can effectively solve the problem of detecting significant targets caused by complex background interference in remote sensing images, and improve detection accuracy and robustness.

[0097] In this embodiment, please refer to Figure 3 The core logic of the Radon projection state space encoder lies in transforming 2D spatial features into anisotropic 1D state flow through integral geometric operators, thereby explicitly repairing topological breaks in remote sensing targets during the imaging process.

[0098] The entire processing flow begins with the projection decoupling stage. For the deep feature F4 output by the backbone network, a set of projection angles Θ={ In this embodiment, K=4, corresponding to Θ={0°, 45°, 90°, 135°}. These four angles cover the horizontal, vertical, and two diagonal principal axes of the image, effectively capturing the orientation of most man-made and natural features. The discrete Radon transform operator is used to process the input features along these directions. Perform line integrals to generate an initial set of sine wave sequences. On a discrete feature map, this is achieved by accumulating pixel values ​​along a specific path. Let the feature map coordinates be (i, j) and the projection axis coordinates be t. The Radon transform is determined by the discrete approximation of the Dirac function δ: ;

[0099] To address computational distortion caused by inconsistent projection path lengths, a set of sine wave sequences was used. The system then proceeds to the mask normalization module. In this module, the system normalizes the sequence to its maximum length using zero-padding. Sequence set Through padding operations, they are uniformly transformed into normalized projection sequences of constant length. It also outputs a set of validity mask matrices simultaneously. This mask matrix not only identifies the real signal region, but also serves as a spatial constraint for the subsequent generation of dynamic parameters.

[0100] Following that, the normalized projection sequence Its corresponding validity mask They collectively enter an anisotropic selective state-space model. This modeling process is not a simple linear transformation, but consists of three interconnected sub-stages:

[0101] The first step is parameter adaptive mapping. The system encodes the position vector of the projection angle θ. With the process The constrained masked weighted pooling features are nonlinearly concatenated, and an adaptive discretization step size is generated using the learned weight matrix. Its analytical expression is: ;in, and It is the weight matrix that the system automatically learns during the end-to-end training process. This is ensured. The element-wise product of the feature sequence and the mask. This is achieved by using the mask... By introducing normalization into the denominator, the system ensures the step size. The generation of [the signal] is driven solely by the effective signal concentration. The introduction of the Softplus activation function ensures that the step size is always positive, thus maintaining the stability of the system's dynamic evolution. This formula ensures the step size [is constant / stable]. The generation is driven solely by the effective signal concentration, thereby dynamically adjusting the model's ability to preserve historical states based on the target's geometric orientation.

[0102] The second step is the discretization of the continuous system. The system utilizes the output from the previous step... A zero-order preserve transformation is performed on the preset continuous state matrix A and input matrix B to calculate the discretized transition matrix specific to the current projection direction. With discretized input matrix This step transforms static weights into dynamic evolution parameters bound to geometric orientation.

[0103] The third step is recursive state evolution. Using the obtained... and The system guides the standardized projection sequence. Temporal evolution is performed in the latent state space. Specifically, each feature in the sequence... All according to The equation updates its hidden state The final projected output is an enhanced 1D state sequence. This evolutionary mechanism can capture long-range linear dependencies along the long axis of the target, achieving structural stitching at the feature level.

[0104] The final stage of the processing is the inverse mapping and fusion phase. This set of enhanced 1D state sequences The information is fed into the inverse Radon transform operator, which repaints the enhanced information in 1D space back onto the 2D coordinate plane, forming component features in multiple directions. The inverse mapping process is as follows, where (i, j) are the pixel coordinates of the target 2D feature map. Through this operator, the system generates 2D component feature maps that correspond one-to-one with the projection angle. . ;

[0105] To achieve dynamic information filtering, the system assigns weights based on the spatial information entropy of each directional component and performs nonlinear aggregation on the reconstructed multi-angle features. The information entropy H( of each projection sequence is calculated.) Based on the principle of minimizing entropy, the system generates dynamic attention weights in each direction using a temperature-sensitive Softmax operator. : ;

[0106] in, The temperature parameter is used to adjust the smoothness of the weight distribution. This formula ensures that directions that capture continuous, well-defined geometry (low-entropy directions) achieve higher weight gains. With corresponding 2D component features ;

[0107] The final system will use dynamic attention weights With corresponding 2D component features Perform element-wise weighted summation and then... Feature remixing convolutional layers perform nonlinear mapping to generate geometrically enhanced features with strong geometric awareness and connectivity. . .

[0108] This embodiment generates It will be passed to the decoding process as the core semantic seed, serving as the initial momentum for dynamic state updates.

[0109] S123. Based on the geometric enhancement features, a hysteresis-gated kinetic fusion (HGKF) module is used to modulate the fusion path of deep semantics and shallow details, filter out background noise, and generate a prediction saliency map.

[0110] In this embodiment, the predicted saliency map refers to the probability map generated after initialization by geometric enhancement features and processing by the hysteresis-gated dynamics fusion module, which can reflect the probability of each region in the image becoming a salient target and ensure the integrity of the topological structure.

[0111] In one embodiment, step S123 described above may include steps S1231 to S1234.

[0112] S1231. Initialize the dynamic state variables using the geometric enhancement features.

[0113] In this embodiment, the dynamic state variable refers to the feature map with strong geometric perception capability obtained after processing by the Radon projection state-space encoder. This feature map serves as initial momentum, guiding feature fusion and noise suppression in subsequent decoding processes. This ensures that target structural information is effectively preserved and background interference is suppressed during decoding.

[0114] S1232. In each decoding layer, starting from the dynamic state variables, the combination of deep state variables and current layer encoder features is modulated by the momentum preservation coefficient to generate intermediate dynamic state variables.

[0115] In this embodiment, intermediate dynamic state variables refer to the deep state variables (such as...) propagated from the previous layer during each decoding process. (and multi-scale encoder features from the backbone network) The result after weighted summation. This process modulates the feature flow through the momentum preservation coefficient μ, ensuring the consistency and stability of the features while effectively suppressing the influence of background noise.

[0116] S1233. Based on the intermediate dynamic state variables, control the gating state according to the set opening and closing thresholds to obtain the processed feature map.

[0117] In this embodiment, the processed feature map refers to the topologically complete feature map after processing by the double-threshold hysteresis activation function. Specifically, when the intermediate dynamic state variable exceeds a set opening threshold, the gating is activated; when it falls below a closing threshold, the gating is deactivated; and when it is between the two thresholds, the gating maintains the state of the previous moment. This mechanism ensures that the target features maintain consistency even under weak disturbances, avoiding spurious responses caused by transient noise.

[0118] S1234. Generate a saliency probability map based on the processed feature map, and optimize the system parameters by calculating the topological loss gradient.

[0119] In this embodiment, the predicted saliency map refers to the saliency probability map generated after the above steps. This map not only reflects the probability that each region in the image is a salient target, but also ensures the integrity of the prediction result in terms of topological structure by introducing topological consistency constraint optimization based on persistent cohomology theory. By calculating the Wasserstein distance between the predicted map and the ground truth map as a topological loss term, the system can dynamically adjust the model parameters, thereby increasing the probability value at the break to eliminate redundant connected components or reducing the probability value at background clutter to fill in erroneous holes, achieving high-precision and topologically complete salient target detection.

[0120] In one embodiment, such as Figure 4 and Figure 5 As shown, the feature fusion process in the decoding stage is optimized by utilizing the hysteresis effect in nonlinear control theory to replace the traditional simple stitching method, thereby achieving effective filtering of transient noise in the remote sensing background.

[0121] The entire processing flow begins with the initialization and propagation phase of the dynamic state. The system receives the final result from the previous step—geometrically enhanced features with strong geometry awareness. At the deepest layer of the decoding chain (layer 4), this feature... Directly defined as the initial dynamic state variables driving the evolution of the system Subsequently, the system initiates a recursive downlink fusion mechanism at each decoding level l ( In this context, the deep state variables propagated from the upper layers are... First, spatial alignment is performed using a bilinear upsampling operator, followed by integration with multi-scale encoder features from the backbone network (Example 1). They are jointly fed to the momentum update operator.

[0122] The next step is momentum update and residual correction. To ensure consistency in significance detection, the system uses a momentum retention coefficient. (Value 0.9) Modulates the feature stream. Specifically, the system calculates a weighted sum of the upsampled deep state and the current layer's detailed features after convolutional transformation to generate the intermediate dynamic state variables of the current layer. : ;

[0123] This step establishes a semantic protection mechanism: isolated noise points in the background are in a deep state. If there is no response, its energy will be significantly diluted during the weighting process, making it difficult to [react]. Sufficient strength was established to suppress feature flickering at its source.

[0124] Subsequently, the intermediate state variables were generated. Entering the core dual-threshold hysteresis activation module. This module simulates the characteristics of a physical hysteresis loop, using a preset activation threshold τ. high With the closing threshold τ low (where τ) high > τ low To maintain a gated state with memory properties. .when The cumulative momentum exceeds τ high At that time, the system will Flip to the on state; when When the threshold is between two thresholds, the system forcibly maintains the gating logic from the previous time step; only when the threshold is between two thresholds... Completely fell to τ low The gating system only executes the closing command under the following conditions. This inertial switching mechanism ensures that the target's feature channels remain open through a memory effect even when it is slightly obscured by clouds or has uneven texture, thus effectively repairing the internal voids of significant targets.

[0125] The final stage of the processing flow is the feature output and closed-loop feedback phase. The system will output the complete topological feature map after hysteresis-gated modulation. (calculated as) · The signal is passed down to the next-level decoding unit. When the spatial resolution is restored to 1 / 4 of the original image, the output signal of the last layer of hysteresis-gated dynamics fusion module is fed into the prediction head to generate the final saliency probability map.

[0126] This probabilistic graph is then passed to the topology consistency constraint optimization stage. By calculating the persistent cohomology difference between the predicted result and the true value, the topology loss gradient generated by the system is backpropagated along the dynamic path of the aforementioned hysteresis-gated dynamics fusion module, affecting the momentum coefficient. The system performs end-to-end dynamic optimization of the hysteresis threshold, ultimately producing a high-precision and topologically connected salient target mask.

[0127] S124. Calculate the topological difference between the predicted saliency map and the ground truth map, and use the hybrid loss function for end-to-end optimization to obtain the salient target mask.

[0128] In one embodiment, step S124 described above may include steps S1241 to S1244.

[0129] S1241. The predicted saliency probability map and the truth map are processed by the sub-level set filtering operator to convert the predicted saliency probability map and the truth map into a persistent map, so as to obtain a persistent map of the prediction map and a persistent map of the truth map.

[0130] In this embodiment, in this step, the predicted saliency probability map (a two-dimensional matrix generated by the model representing the probability that each pixel in the image is a salient object) and the ground truth map (i.e., the actual saliency annotation map) are first converted into a persistent map. Specifically:

[0131] Persistent Predictive Graph: This is the result obtained by processing the predicted saliency probability graph under a series of dynamically changing thresholds using a subset level set filtering algorithm. This process records the time points at which connected components and holes appear and disappear with different thresholds, thus forming a persistent graph describing the lifecycle of these topological features.

[0132] Truth graph persistent graph: Similarly, this persistent graph is extracted from the truth graph and reflects the topological structure that a salient target should ideally have.

[0133] S1242. Calculate the Wasserstein distance between the predicted persistent graph and the ground persistent graph to obtain the topological loss gradient.

[0134] In this embodiment, it is necessary to quantify the difference between the predicted persistent graph and the ground truth persistent graph. Therefore, the Wasserstein distance is used as the metric, defined as follows:

[0135] Topological loss gradient: The Wasserstein distance measures the cost of the best match between two persistent graphs. By minimizing this distance, we can ensure that the connectivity and hole features in the prediction are as close as possible to the reality. The gradient information generated in this process guides the direction of network parameter updates, aiming to reduce the topological differences between the predicted and ground truth graphs.

[0136] S1243. Using the backpropagation of the topological loss gradient, optimize the network parameters and adjust the output of the target detection model to obtain the optimized model output.

[0137] In this embodiment, based on the topological loss gradient obtained in the previous step, the following operations are performed:

[0138] Optimized model output: The network weights are adjusted based on the calculated topological loss gradient using the backpropagation algorithm. This not only affects the processing of the current batch of data but also gradually improves the model's overall performance in detecting salient targets in remote sensing images through iterative learning across the entire training set. This process emphasizes the pursuit of topological accuracy, enabling the model to better capture and maintain the integrity of salient targets.

[0139] S1244. Generate a salient target mask based on the optimized model output.

[0140] The final step is to generate a salient target mask based on the final model output obtained after the above optimization process:

[0141] Salient Target Mask: This is a binary image where foreground pixels correspond to the parts identified as salient targets, while background pixels represent non-salient regions. This mask comprehensively considers information from multiple aspects, including geometric perception, dynamic evolution, and topological consistency, providing high-precision and topologically complete salient target detection results.

[0142] In this embodiment, please refer to Figure 6By utilizing the persistent cohomology theory in algebraic topology to impose global constraints on the prediction results, and by introducing topological consistency loss, the structural connectivity of remote sensing targets (such as slender roads and broken bridges) is explicitly repaired in addition to pixel classification accuracy.

[0143] The entire processing flow begins with the extraction and filtering of topological features. The system receives the final saliency probability map P generated by the decoder and the corresponding ground truth map G. The system treats these two maps as a 2D scalar field and initiates a sub-level set filtering operator. This operator monitors the evolution of topological features in the binarized image at different thresholds v∈[0,1]. Specifically, the system records the (birth, death) coordinate pairs of each connected component and hole, thereby transforming the complex 2D geometric connectivity into an abstract set of persistent graphs.

[0144] The next step is topological distance quantization and loss calculation. The system utilizes the obtained predicted persistent graph. With truth graphs and persistent graphs Calculate the Wasserstein distance between the two. It is the best bijective matching between two point sets. ;

[0145] In this calculation process, the system searches for the optimal bijective match between two point sets: if an unexpected break appears in the predicted map, the feature is... The result will appear as a noisy point off the main diagonal. By minimizing the Wasserstein distance, the system establishes a topological loss term. This term differs from the standard pixel-level cross-entropy loss. and structural crossover ratio loss Together they constitute the system's hybrid objective function . ;

[0146] Through this loss, the network receives explicit gradient signals to increase the probability value at the breakpoint to eliminate redundant connected components, or to decrease the probability value at the background to eliminate false holes. This mechanism directly affects the topology of the target, rather than just pixel classification accuracy.

[0147] Subsequently, the system enters the closed-loop feedback and dynamic tuning phase of the gradient. The generated topological loss gradient is no longer limited to the correction of local pixels, but serves as a clear signal to guide the network to increase the response probability value at the break point to eliminate redundant connected components, or to reduce the probability value at background clutter to fill erroneous holes. This gradient signal is backpropagated along the dynamic path of the hysteresis-gated dynamic fusion module and the Radon projection evolution path, dynamically fine-tuning the momentum coefficient. Hysteresis threshold τ and transition matrix parameters of the state-space model, etc.

[0148] The final stage of the processing flow is the mask output and performance locking phase. After the closed-loop optimization of the above topology consistency constraints, the system finally outputs a salient target mask with a complete topology structure.

[0149] In this embodiment, as Figure 7 As shown, this embodiment was validated on the mainstream remote sensing saliency detection dataset ORSSD. Evaluation metrics include S-measure ( ), F-measure ( ), mean absolute error (M), and frames per second (FPS).

[0150] The verification environment was PyTorch 2.1 and NVIDIA RTX 4090 GPU (24GB).

[0151] The parameter settings are as follows: the input image size is adjusted to 352×352. The optimizer used is AdamW, with an initial learning rate of 1e-4 and a weight decay of 1e-4. The momentum coefficient in the hysteresis-gated dynamics fusion module is also specified. =0.9, hysteresis threshold initialized to =0.7, =0.3 and set as a learnable parameter.

[0152] Training strategy: Due to the high computational cost of topology loss, a warm-up strategy is adopted: only BCE+IoU loss is used for training in the first 10 epochs, and after the prediction graph is initially formed, other methods are added. Make minor adjustments.

[0153] The results of the comparative experiment are shown in Table 1.

[0154] Table 1. Experimental Results

[0155]

[0156] The metrics reached 0.9621 and 0.9298 respectively, demonstrating that this embodiment performs well on the ORSSD dataset. , , M = 0.9208, 0.9125, 0.0051, significantly outperforming existing remote sensing salient target detection methods. Although multi-angle projection is introduced, due to the linear complexity of Mamba and optimization of parallel computing, the inference speed of this embodiment reaches 42 FPS, meeting the requirements of real-time processing and outperforming other computationally intensive methods.

[0157] The method in this embodiment utilizes integral geometric transformation and nonlinear control theory to enhance the detection of salient targets in high-resolution optical remote sensing images. This embodiment is particularly suitable for fine-grained segmentation and localization of targets with large-scale spans, arbitrary directional distributions, and complex topological structures (such as bridges, airport runways, ships, and narrow rivers) in complex terrain backgrounds (such as ports, urban clusters, mountains, and sea surfaces).

[0158] To address the aforementioned issues, this embodiment proposes a detection method based on the fusion of Radon projection state space and hysteresis gated dynamics. This method utilizes the integral geometric properties of the Radon transform to explicitly decouple the 2D spatial structure into a multi-directional 1D projection sequence, fundamentally solving the spatial discontinuity problem of the state space model in 2D images. Simultaneously, it introduces a physics-inspired hysteresis dynamics mechanism, endowing the network with robustness against background noise. Finally, it combines topological data analysis to constrain the topological integrity of the target at the loss function level. The core logic lies in transforming the static feature mapping task into a dynamic evolutionary process guided by geometric priors. First, a Radon projection state space encoder is constructed. By introducing the discrete Radon transform operator, the feature tensors originally in high-dimensional space are geometrically decoupled along multiple preset angles, transforming them into a directional projection sequence. Based on this, an anisotropic parameterized selective state space model is used to capture long-range linear structural dependencies within the projection domain. This design not only gives the model an intuition to perceive complex geometries but also achieves explicit repair of the target's topological structure through adaptive discretization step size.

[0159] In the decoding and feature integration stages, this embodiment abandons the static fusion approach and constructs a hysteresis-gated dynamics fusion module. This module models the propagation of saliency signals as a dynamic process with state memory, simulating physical hysteresis, and modulates the feature stream through a dual-threshold switching mechanism. By introducing deep semantic momentum as the initial state, hysteresis-gated dynamics fusion can effectively suppress spurious responses caused by background clutter, and maintain the logical coherence of saliency in damaged or occluded areas by utilizing state inertia, ensuring the purity and integrity of the output results.

[0160] To achieve global-level precision locking, this embodiment further introduces a topological consistency constraint based on persistent cohomology theory. By calculating the differences in topological features between the predicted probability graph and the ground truth graph at different level sets, a topological loss function incorporating Wasserstein distance is constructed. This constraint mechanism directly acts on the connected components and hole features of the target, driving the network to examine the integrity of the target structure from an algebraic topological perspective.

[0161] The advantage of this embodiment is that it establishes a complete closed loop from geometric feature decoupling and dynamic state evolution to global topology optimization, which significantly improves the detection accuracy and topological robustness of the system in complex and high-noise remote sensing scenarios.

[0162] In this embodiment, the anisotropic Radon domain state encoding specifically includes: defining a set of projection angles, calculating the line integral response of the feature tensor at each angle to generate a sine wave sequence; applying a masking strategy to the sequence to unify the length; inputting the padded sequence into a selective state space model, using a discretization step size parameter that is adaptively adjusted with the projection angle to capture linear structural dependencies; and finally mapping back to two-dimensional space through the inverse Radon transform operator and performing weighted fusion based on the entropy value of the projection sequence.

[0163] The selective state-space model adopts the discretization form of a linear time-invariant system, and the discretization process of its state transition matrix and input matrix is ​​controlled by the dynamic step size parameter generated by the projection angle feature.

[0164] Define a dynamic state variable that is updated by the momentum of the deep features and the residual of the current layer features; set an on threshold and a off threshold, where the on threshold is greater than the off threshold; activate gating when the dynamic state variable exceeds the on threshold, and deactivate gating if and only if the dynamic state variable is below the off threshold, and maintain the gating state of the previous time step when the dynamic state variable is between the two thresholds.

[0165] Persistent graphs are constructed based on sub-level set filtering to compare the prediction graph and the ground truth graph. The Wasserstein distance between the two persistent graphs is calculated as a topological loss term to constrain the number of connected components and the hole structure of the prediction target.

[0166] In addition, an attention mechanism based on projection entropy is used in the inverse transform aggregation process to automatically adjust the feature fusion weights according to the information entropy of each projection direction, giving higher weights to projection directions with richer structural information.

[0167] By employing an innovative Radon projection state-space encoder, the system successfully overcomes the spatial continuity bottleneck of traditional state-space models in 2D image modeling, achieving precise capture of the anisotropic geometric structure of remote sensing targets using integral geometry principles. Combined with a hysteresis-gated dynamic decoder and topological consistency loss, the system transforms the static pixel classification task into a dynamic evolution process guided by geometric priors. This ensures extremely high detection accuracy while fundamentally solving the problems of structural fragmentation and edge oscillations that easily occur in remote sensing targets against complex backgrounds. This deep fusion architecture, from geometric decoupling to dynamic filtering and then to global topological constraints, not only theoretically expands the application boundaries of state-space models in computer vision but also provides robust and efficient technical support for the refined monitoring of complex terrain features in practical engineering. This embodiment belongs to the field of remote sensing image processing technology. To address the issues of topological breaks and edge blurring in high-resolution remote sensing images caused by varying target orientations and complex background interference, this invention constructs a Radon projection state-space encoder. It utilizes the discrete Radon transform operator to decouple the two-dimensional feature field into a multi-angle projection sequence and captures global long-range dependencies through an anisotropic parameterized selective state-space model, explicitly restoring the feature continuity of slender and tilted targets at the geometric level. In the decoding stage, this invention designs a hysteresis-gated dynamics fusion module, modeling feature fusion as a recursive evolution process driven by initial state momentum. It introduces a nonlinear hysteresis effect and a dual-threshold switching mechanism to effectively filter transient noise in the background and maintain the consistency of target saliency. Finally, combined with a topological consistency loss function based on persistent cohomology theory, it constrains the connected components and hole features of the predicted graph at the algebraic topological level, generating a high-precision, topologically complete saliency mask. Experimental results show that the method in this embodiment significantly outperforms existing technologies in terms of detection accuracy and structural integrity, exhibiting strong noise robustness.

[0168] The aforementioned remote sensing salient target detection method constructs a multi-scale feature pyramid and maps these features to a discrete feature field to be projected. It then utilizes a Radon projection state-space encoder to achieve anisotropic decoupling and sequence modeling, thereby geometrically addressing the feature continuity issues of slender and tilted targets. Next, a hysteresis-gated dynamic fusion module modulates the fusion path between deep semantics and shallow details, effectively filtering background noise and maintaining the consistency of target saliency, thus solving problems such as structural fragmentation and edge blurring in traditional methods. Finally, by calculating the topological difference between the predicted saliency map and the ground truth map and using a hybrid loss function including topological consistency loss for end-to-end optimization, the method ensures high accuracy and topological integrity in the output even in complex backgrounds, significantly improving the system's noise robustness and detection accuracy. This method achieves deep fusion from geometric decoupling and dynamic filtering to global topological constraints, effectively overcoming various challenges encountered by traditional techniques in processing complex remote sensing images.

[0169] Figure 8 This is a schematic block diagram of a remote sensing salient target detection device 300 provided in an embodiment of the present invention. Figure 8 As shown, corresponding to the above-described remote sensing salient target detection method, the present invention also provides a remote sensing salient target detection device 300. This remote sensing salient target detection device 300 includes a unit for performing the above-described remote sensing salient target detection method, and the device can be configured in a server. Specifically, please refer to... Figure 8 The remote sensing salient target detection device 300 includes an acquisition unit 301 and a detection unit 302.

[0170] The acquisition unit 301 is used to acquire the optical remote sensing image to be detected; the detection unit 302 is used to input the optical remote sensing image to be detected into the target detection model to perform target detection, so as to obtain a salient target mask.

[0171] In one embodiment, the detection unit 302 includes:

[0172] The mapping subunit is used to construct a multi-scale feature pyramid for the optical remote sensing image to be detected, and map the multi-scale feature pyramid into a discrete feature field to be projected; the projection coding subunit is used to perform anisotropic decoupling and sequence modeling on the discrete feature field to be projected using a Radon projection state space encoder, and then generate geometrically enhanced features through inverse transformation; the enhancement subunit is used to modulate the fusion path of deep semantics and shallow details using a hysteresis-gated dynamics fusion module based on the geometrically enhanced features, filter out background noise and generate a predicted saliency map; the computation subunit is used to calculate the topological difference between the predicted saliency map and the ground truth map, and perform end-to-end optimization using a hybrid loss function to obtain a salient target mask.

[0173] In one embodiment, the mapping subunit is used to extract semantic information at different levels from the optical remote sensing image to be detected by using a multi-scale feature extraction layer in a layer-by-layer compression manner, generate a multi-scale feature pyramid, and map the multi-scale feature pyramid into a discrete feature field to be projected. The multi-scale feature extraction layer includes a pyramid visual Transformer or a convolutional neural network as the backbone network.

[0174] In one embodiment, the projection coding subunit includes:

[0175] The transformation module is used to select a specific set of angles to perform Radon transform on the discrete feature field to be projected, generating a sequence of sine graphs representing information in different directions; the adjustment module is used to adjust all the sine graph sequences to the same length by zero-padding and create an effectiveness mask; the evolution module is used to guide the normalized projection sequence to evolve in the latent state space using recursive state evolution based on adaptive parameter mapping and discretization of the continuous system, so as to obtain an enhanced one-dimensional state sequence; the aggregation module is used to remap the enhanced one-dimensional state sequence to a two-dimensional space using inverse Radon transform, and dynamically allocate attention weights according to the information entropy of each direction to perform feature aggregation, so as to obtain multi-angle features; the fusion module is used to fuse the multi-angle features into an enhanced feature map with strong geometric perception capability through weighted summation and nonlinear mapping, so as to obtain geometrically enhanced features.

[0176] In one embodiment, the projection coding subunit includes:

[0177] The step size generation module is used to generate an adaptive discretized step size for each projection angle based on the position encoding and mask weighted pooling features of the enhanced one-dimensional state sequence; the transition matrix generation module is used to transform the preset continuous state matrix using the adaptive step size to obtain a discretized transition matrix specific to the current projection direction; the sequence evolution module is used to guide the normalized projection sequence to evolve in the latent state space based on the discretized matrix to obtain the enhanced one-dimensional state sequence.

[0178] In one embodiment, the enhancement subunit includes:

[0179] An initialization module is used to initialize the dynamic state variables using the geometric enhancement features; an intermediate variable generation module is used to generate intermediate dynamic state variables at each decoding layer, starting from the dynamic state variables and modulating the combination of deep state variables and current layer encoder features using momentum preservation coefficients; a state control module is used to control the gating state based on the intermediate dynamic state variables and according to set on and off thresholds to obtain the processed feature map; and a probability map generation module is used to generate a saliency probability map based on the processed feature map and optimize system parameters by calculating the topological loss gradient.

[0180] In one embodiment, the computing subunit includes:

[0181] The persistent graph conversion module processes the predicted saliency probability map and the ground truth map using a sublevel set filtering operator, converting them into persistent graphs to obtain a predicted persistent graph and a ground truth persistent graph. The distance calculation module calculates the Wasserstein distance between the predicted persistent graph and the ground truth persistent graph to obtain the topological loss gradient. The output optimization module uses the topological loss gradient for backpropagation to optimize network parameters and adjust the output of the target detection model to obtain an optimized model output. The mask generation module generates a salient target mask based on the optimized model output.

[0182] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned remote sensing significant target detection device 300 and each unit can be referred to the corresponding description in the foregoing method embodiments. For the sake of convenience and brevity, it will not be repeated here.

[0183] The aforementioned remote sensing salient target detection device 300 can be implemented as a computer program, which can, for example... Figure 9 It runs on the computer device shown.

[0184] Please see Figure 9 , Figure 9 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0185] See Figure 9 The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0186] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform a remote sensing salient target detection method.

[0187] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0188] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can perform a remote sensing salient target detection method.

[0189] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0190] The processor 502 is used to run a computer program 5032 stored in a memory to implement all the steps of the remote sensing salient target detection method.

[0191] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0192] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0193] Therefore, the present invention also provides a storage medium. This storage medium may be a computer-readable storage medium. The storage medium stores a computer program, wherein when executed by a processor, the computer program causes the processor to perform all the steps of the remote sensing salient target detection method.

[0194] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0195] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0196] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0197] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the device of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0198] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or 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, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0199] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for detecting salient objects in remote sensing images, characterized in that, include: Acquire the optical remote sensing image to be detected; The optical remote sensing image to be detected is input into a target detection model for target detection to obtain a salient target mask, which includes: A multi-scale feature pyramid is constructed for the optical remote sensing image to be detected, and the multi-scale feature pyramid is mapped to a discrete feature field to be projected. The discrete feature field to be projected is anisotropically decoupled and sequence modeled using a Radon projection state space encoder, and then geometrically enhanced features are generated through inverse transformation. Based on the aforementioned geometric enhancement features, a hysteresis-gated dynamics fusion module is used to modulate the fusion path between deep semantics and shallow details, filter out background noise, and generate a prediction saliency map. The topological difference between the predicted saliency map and the ground truth map is calculated, and end-to-end optimization is performed using a hybrid loss function to obtain the salient target mask.

2. The remote sensing salient target detection method according to claim 1, characterized in that, The target detection model includes a multi-scale feature extraction layer, a Radon projection state-space encoder, a hysteresis-gated dynamics fusion module, and a prediction head.

3. The remote sensing salient target detection method according to claim 2, characterized in that, Constructing a multi-scale feature pyramid for the optical remote sensing image to be detected, and mapping the multi-scale feature pyramid to a discrete feature field to be projected, includes: The optical remote sensing image to be detected is processed by a multi-scale feature extraction layer to extract semantic information at different levels in a layer-by-layer compression manner, generating a multi-scale feature pyramid, and mapping the multi-scale feature pyramid to a discrete feature field to be projected. The multi-scale feature extraction layer includes a pyramid visual Transformer or a convolutional neural network as the backbone network.

4. The remote sensing salient target detection method according to claim 2, characterized in that, The process of using a Radon projection state-space encoder to perform anisotropic decoupling and sequence modeling on the discrete feature field to be projected, and then generating geometrically enhanced features through inverse transformation, includes: Select a specific set of angles to perform Radon transform on the discrete feature field to be projected, and generate a sequence of sine graphs representing information in different directions; All the sine wave sequences are adjusted to the same length using zero padding, and a validity mask is created. Based on adaptive parameter mapping and the discretization of continuous systems, the effective mask is guided by recursive state evolution to evolve the standardized projection sequence in the hidden state space, so as to obtain an enhanced one-dimensional state sequence. The enhanced one-dimensional state sequence is remapped to a two-dimensional space using the inverse Radon transform, and attention weights are dynamically allocated based on the information entropy of each direction to perform feature aggregation, so as to obtain multi-angle features. The multi-angle features are fused into an enhanced feature map with strong geometric perception by weighted summation and nonlinear mapping to obtain geometrically enhanced features.

5. The remote sensing salient target detection method according to claim 2, characterized in that, The enhanced one-dimensional state sequence is remapped to a two-dimensional space using the inverse Radon transform, and attention weights are dynamically allocated based on the information entropy in each direction to perform feature aggregation, thereby obtaining multi-angle features, including: An adaptive discretization step size for each projection angle is generated for the enhanced one-dimensional state sequence based on position encoding and mask weighted pooling features; The preset continuous state matrix is ​​transformed using an adaptive step size to obtain a discretized transition matrix specific to the current projection direction. Based on the discretized matrix, the normalized projection sequence is guided to evolve in the hidden state space to obtain an enhanced one-dimensional state sequence.

6. The remote sensing salient target detection method according to claim 2, characterized in that, The process, based on the geometric enhancement features, employs a hysteresis-gated dynamics fusion module to modulate the fusion path between deep semantics and shallow details, filters out background noise, and generates a prediction saliency map, including: Initialize the dynamic state variables using the aforementioned geometric enhancement features; At each decoding layer, starting from the dynamic state variables, the combination of deep state variables and current layer encoder features is modulated by momentum preservation coefficients to generate intermediate dynamic state variables. Based on the intermediate dynamic state variables, the gating state is controlled according to the set opening and closing thresholds to obtain the processed feature map; A saliency probability map is generated based on the processed feature map, and the system parameters are optimized by calculating the topological loss gradient.

7. The remote sensing salient target detection method according to claim 2, characterized in that, The calculation of the topological difference between the predicted saliency map and the ground truth map, using a hybrid loss function for end-to-end optimization to obtain a salient target mask, includes: The predicted saliency probability map and the truth map are processed by a sublevel set filtering operator to convert the predicted saliency probability map and the truth map into persistent maps, so as to obtain a persistent map of the predicted map and a persistent map of the truth map. Calculate the Wasserstein distance between the predicted persistent graph and the ground persistent graph to obtain the topological loss gradient; By using the backpropagation of the topological loss gradient, the network parameters are optimized and the output of the target detection model is adjusted to obtain the optimized model output; A salient target mask is generated based on the output of the optimized model.

8. A remote sensing salient target detection device, characterized in that, include: The acquisition unit is used to acquire the optical remote sensing image to be detected; The detection unit is used to input the optical remote sensing image to be detected into the target detection model for target detection, so as to obtain a salient target mask; The detection unit includes: The mapping subunit is used to construct a multi-scale feature pyramid for the optical remote sensing image to be detected, and to map the multi-scale feature pyramid into a discrete feature field to be projected. The projection coding subunit is used to perform anisotropic decoupling and sequence modeling on the discrete feature field to be projected using the Radon projection state space encoder, and then generate geometrically enhanced features through inverse transformation; An enhancement subunit is used to modulate the fusion path of deep semantics and shallow details based on the geometric enhancement features using a hysteresis-gated dynamics fusion module, filter out background noise, and generate a prediction saliency map. The computational subunit is used to calculate the topological difference between the predicted saliency map and the ground truth map, and performs end-to-end optimization using a hybrid loss function to obtain the salient target mask.

9. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images