High-resolution remote sensing image target detection and intelligent interpretation system

Abstract

The invention discloses a high-resolution remote sensing image target detection and intelligent interpretation system, and belongs to the technical field of remote sensing image processing and computer vision. The system comprises a multi-scale feature adaptive extraction module, a cross-modal semantic fusion module, a context association enhancement module, a dynamic weight closed-loop adjustment module and a target accurate positioning and attribute extraction module, and deep coupling and collaborative optimization among the modules are realized by constructing a multi-layer closed-loop feedback mechanism. Multi-scale feature extraction is combined with a layered attention mechanism, spectrum and space information are integrated through cross-modal fusion, context enhancement utilizes a global-local relation graph and dynamic weight adjustment to achieve parameter self-adaptive optimization, the system effectively overcomes the complex background interference and multi-scale target detection problems, the average target detection precision is improved by 15%-25% compared with the prior art, and the target detection efficiency is improved by 20%-30%. The method has excellent performance in multi-category target detection and attribute interpretation of buildings, roads, vehicles and the like, and has wide application value.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing and computer vision technology, specifically to a high-resolution remote sensing image target detection and intelligent interpretation system. Background Technology

[0002] High-resolution remote sensing imagery has wide applications in urban planning, land resource surveys, disaster monitoring, and precision agricultural management. With the rapid development of remote sensing imaging technology, the spatial resolution of remote sensing images has improved from meter-level to sub-meter-level and even centimeter-level, enabling them to present richer details of ground features. However, the increased image complexity brought about by high resolution also places higher demands on target detection and intelligent interpretation technologies.

[0003] Traditional remote sensing image target detection methods mainly rely on manually designed feature extraction operators and classifiers, such as using scale-invariant feature transformation algorithms to extract features and then combining them with support vector machines for classification. These methods are effective for simple scenes, but their detection accuracy and robustness fall short of practical application requirements when dealing with complex land cover types, diverse target scales, and complex background interference. In recent years, with the rapid development of deep learning technology, target detection methods based on convolutional neural networks have been widely used in the field of remote sensing image processing.

[0004] Existing technologies, exemplified by CN113689445A, propose a high-resolution remote sensing building extraction method combining semantic segmentation and edge detection. This method employs a semantic segmentation network such as U-Net to predict building roof surfaces and an edge detection network such as BDCN to predict building edges. Then, the edge map undergoes denoising, skeleton extraction, deburring, and line break repair to ultimately obtain the building's vector result. This method addresses, to some extent, the problem of discontinuous and non-closed building edges predicted by deep convolutional neural networks, improving the accuracy of building edge extraction. However, this method still has the following shortcomings:

[0005] First, this method uses separate semantic segmentation and edge detection networks, lacking deep interaction and collaborative optimization mechanisms between them. This results in insufficient fusion of semantic and edge information, easily leading to semantic inconsistencies when handling complex scenes. Second, the processing flow is a unidirectional linear process, lacking a feedback mechanism and unable to optimize and adjust the initial feature extraction based on subsequent processing results, limiting the potential for overall performance improvement. Third, this method primarily focuses on extracting single-category buildings, lacking sufficient ability for simultaneous detection and attribute interpretation of multi-category targets, and failing to fully utilize the rich spectral and spatial contextual information contained in high-resolution remote sensing imagery. Furthermore, this method uses fixed parameter configurations, lacking adaptability to different scenes and target scales, and its detection accuracy and robustness need improvement when facing complex and variable remote sensing imagery.

[0006] Current cutting-edge research in remote sensing image target detection mainly focuses on the following directions: First, replacing traditional convolutional neural networks with Transformer architecture and utilizing self-attention mechanisms to model global contextual relationships, thereby improving the ability to understand large-scale targets and scenes; second, designing multi-scale feature fusion strategies, using methods such as feature pyramid networks to simultaneously process targets of different scales; third, introducing multi-modal information fusion, collaboratively processing multi-source features such as spectral, spatial, and texture information; and fourth, exploring lightweight model design to reduce computational complexity while ensuring detection accuracy and improving real-time processing capabilities. These research directions provide new ideas for the development of high-resolution remote sensing image target detection technology, but a complete technical system with deep inter-module coupling and closed-loop collaborative optimization has not yet been formed.

[0007] Therefore, how to design a technical solution that can fully utilize the multi-scale features, multi-modal information and spatial context of high-resolution remote sensing images, achieve deep coupling and closed-loop collaborative optimization of various processing modules, and significantly improve the accuracy of target detection and intelligent interpretation capabilities is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a high-resolution remote sensing image target detection and intelligent interpretation system. By constructing core modules such as multi-scale feature adaptive extraction, cross-modal semantic fusion, context association enhancement, and dynamic weight closed-loop adjustment, the system achieves deep coupling and collaborative optimization among the modules, significantly improving the detection accuracy and attribute interpretation accuracy of multi-category targets in high-resolution remote sensing images.

[0009] To achieve the above objectives, the technical solution adopted in this invention is: a high-resolution remote sensing image target detection and intelligent interpretation system, comprising a multi-scale adaptive feature extraction module, a cross-modal semantic fusion module, a context association enhancement module, a dynamic weight closed-loop adjustment module, and a target precise localization and attribute extraction module. The multi-scale adaptive feature extraction module extracts features from remote sensing images at different scales based on a hierarchical attention mechanism, generating a multi-scale feature map set containing shallow detail features, mid-level texture features, and deep semantic features. The cross-modal semantic fusion module performs cross-modal fusion of spectral and spatial features and feeds back the statistical characteristics of the fused feature map to the multi-scale adaptive feature extraction module to adjust the weight allocation of feature extraction. The context association enhancement module constructs a global context relationship graph and a local context relationship graph, models the spatial relationship between the target and its surrounding environment based on a graph convolutional network, and adjusts the fusion weights of the cross-modal semantic fusion module inversely based on the quality evaluation results of the context enhancement features. The dynamic weight closed-loop adjustment module generates a dynamic weight coefficient set based on the contribution score of each module, realizing the adaptive adjustment of the parameters of each module. The target precise localization and attribute extraction module generates candidate target regions based on a weighted region proposal network, performs precise segmentation and category recognition, and feeds back the accuracy of the detection results to the dynamic weight closed-loop adjustment module to update the weight coefficients.

[0010] Through the above technical solution, the present invention has the following beneficial effects:

[0011] First, this invention constructs a deeply coupled architecture consisting of a multi-scale feature adaptive extraction module, a cross-modal semantic fusion module, a context association enhancement module, a dynamic weight closed-loop adjustment module, and a target precise localization and attribute extraction module. This achieves deep coupling between the parameter level, state level, and logic level of each module. The output of the previous module directly serves as the key input parameter of the next module, forming a tight technical chain. This avoids the information loss and semantic inconsistency problems caused by independent processing of modules in traditional methods.

[0012] Second, this invention constructs a complete closed-loop mechanism of "forward transmission → performance evaluation → reverse feedback → parameter adjustment" through a dynamic weight closed-loop adjustment module. The detection results of the downstream module affect the parameter configuration of the upstream module in reverse, realizing global optimization rather than local optimization, which significantly improves the overall performance of the system and its adaptability to different scenarios.

[0013] Third, this invention achieves a multi-dimensional synergistic effect, including: mutual promotion of spectral features and spatial features in the cross-modal semantic fusion module, superposition and enhancement of global and local contexts in the context association enhancement module, synergistic optimization of the performance of each module by the dynamic weight closed-loop adjustment module, and complementary advantages of multi-scale feature extraction and target localization, ensuring that the technical effect presents a non-linear growth characteristic of 1+1>2, and the average accuracy of target detection is improved by 15%-25% compared with the existing technology.

[0014] Fourth, this invention makes full use of the multi-scale spatial information, multispectral band information and global-local context information contained in high-resolution remote sensing images. Through cross-modal fusion and context enhancement mechanisms, it effectively overcomes technical challenges such as complex background interference, target scale diversity and difficulty in detecting small targets. It performs excellently in the simultaneous detection and attribute interpretation of multiple types of targets such as buildings, roads, vehicles and vegetation.

[0015] Fifth, this invention employs a lightweight hierarchical attention mechanism and an efficient graph convolutional network, which controls computational complexity while ensuring detection accuracy. For high-resolution remote sensing images of 2048×2048 pixels, the processing time for a single image can be controlled within 2 seconds on a workstation equipped with GPU acceleration, meeting the real-time requirements in practical applications. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall architecture of the high-resolution remote sensing image target detection and intelligent interpretation system of the present invention;

[0017] Figure 2 This is a schematic diagram of the multi-scale feature adaptive extraction module of the present invention;

[0018] Figure 3 This is a schematic diagram of the cross-modal semantic fusion module of the present invention;

[0019] Figure 4 This is a schematic diagram of the context association enhancement module of the present invention;

[0020] Figure 5 This is a schematic diagram of the workflow of the dynamic weight closed-loop adjustment module of the present invention;

[0021] Figure 6 This is a schematic diagram of the target precise positioning and attribute extraction module of the present invention. Detailed Implementation

[0022] Please refer to the attached document. Figures 1-6 The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0023] Reference Figure 1The high-resolution remote sensing image target detection and intelligent interpretation system provided by this invention includes a multi-scale feature adaptive extraction module 1, a cross-modal semantic fusion module 2, a context association enhancement module 3, a dynamic weight closed-loop adjustment module 4, and a target precise localization and attribute extraction module 5. These five core modules form an organic whole through deep coupling and a closed-loop feedback mechanism, realizing the precise detection and intelligent interpretation of multi-category targets in high-resolution remote sensing images.

[0024] Reference Figure 2 The multi-scale adaptive feature extraction module 1 is the front-end feature extraction unit of the system of this invention, responsible for extracting multi-scale spatial features from the input high-resolution remote sensing image. This module adopts an improved feature pyramid network architecture, combined with a hierarchical attention mechanism, to achieve adaptive feature extraction for targets at different scales.

[0025] The multi-scale adaptive feature extraction module 1 includes a feature pyramid construction unit, a hierarchical attention weighting unit, and a feature fusion unit. The feature pyramid construction unit downsamples the input remote sensing image through five sets of convolutional layers, generating five feature levels with different resolutions. The first level maintains the original resolution and mainly contains shallow details such as edges and textures; the second level downsamples to half the original resolution and contains local shape and structural features; the third level downsamples to one-quarter and contains overall target contour features; the fourth level downsamples to one-eighth and contains mid-level semantic features of the target category; and the fifth level downsamples to one-sixteenth and contains deep semantic features of the global scene. Each convolutional layer group contains 3×3 convolutions, batch normalization, and a ReLU activation function.

[0026] The hierarchical attention weighting unit applies spatial and channel attention to each feature level to enhance important features and suppress redundant features. The spatial attention branch generates a spatial attention map through global average pooling and global max pooling operations, highlighting the spatial locations containing target information. The channel attention branch generates channel attention weights through 1×1 convolution and a sigmoid activation function, adaptively selecting the feature channels most useful for the current task. The outputs of the two branches are then fused element-wise and applied to the original feature map, achieving adaptive feature weighting.

[0027] The feature fusion unit fuses the weighted features from each level to generate a multi-scale feature map set. Specifically, the fifth-level feature is upsampled by a factor of 2 and added to the fourth-level feature; the result is then upsampled by a factor of 2 and added to the third-level feature, and so on, ultimately generating a feature map set containing multi-scale information. This feature map set simultaneously preserves shallow detail features and deep semantic features, providing rich feature representations for subsequent processing.

[0028] The multi-scale feature adaptive extraction module 1 receives feedback signals from the dynamic weight closed-loop adjustment module 4 and dynamically adjusts the weight configuration of features at each level according to the overall detection performance. Preferably, the initial weight configuration is equal at each level, and adaptive adjustments are made during system operation based on the target scale distribution and detection error. For scenarios dominated by small targets, the weight of shallow features is increased; for scenarios dominated by large targets, the weight of deep features is increased.

[0029] In one embodiment of the present invention, the input remote sensing image has a size of 2048×2048 pixels and a spatial resolution of 0.5m, containing four bands: red, green, blue, and near-infrared. The feature pyramid construction unit generates five levels of feature maps with sizes of 2048×2048, 1024×1024, 512×512, 256×256, and 128×128, and feature channels of 64, 128, 256, 512, and 512, respectively. After the hierarchical attention weighting unit applies an attention mechanism to the features at each level, the effective information density of the features is increased by approximately 30%, meaning that more discriminative information is contained within the same feature dimension.

[0030] Reference Figure 3 The cross-modal semantic fusion module 2, located after the multi-scale feature adaptive extraction module 1, is responsible for deeply fusing spatial and spectral features to generate a semantically richer fused feature map. High-resolution remote sensing images not only contain rich spatial structure information but also spectral information in multiple spectral bands, with different land cover categories exhibiting unique response characteristics in the spectral domain. Through cross-modal fusion, these two complementary types of information can be fully utilized to improve the accuracy of target identification.

[0031] The cross-modal semantic fusion module 2 includes a spectral feature extraction unit, an adaptive channel attention unit, a weighted fusion unit, and a fusion quality assessment unit. The spectral feature extraction unit extracts spectral feature vectors from the multispectral bands of the original remote sensing image. For common 4-band remote sensing images, spectral indices such as the Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), and Normalized Difference Building Index (NDBI) are calculated. NDVI is used to enhance vegetation information, NDWI to enhance water information, and NDBI to enhance building information. These spectral indices map the original 4-dimensional spectral space to a more discriminative feature space.

[0032] The adaptive channel attention unit calculates the channel correlation between spectral and spatial features to generate channel attention weights. Specifically, the correlation between the spectral feature vector and each channel in the multi-scale feature map group is calculated, and cosine similarity is used to measure the degree of correlation between them. Channels with high correlation indicate that the spatial and spectral features contained in that channel have strong semantic consistency and should be assigned higher weights. The channel attention weights are normalized using the Softmax function to ensure that the sum of all channel weights is 1.

[0033] The weighted fusion unit performs weighted fusion of spectral and spatial features based on channel attention weights. The fusion process combines feature concatenation and convolutional fusion. First, the spectral feature vector is mapped to the same channel dimension as the spatial features through a 1×1 convolution. Then, it is concatenated with the spatial feature map along the same channel dimension. Finally, a 3×3 convolution is used to integrate the features, generating a fused feature map. This fused feature map simultaneously contains spatial structure information and spectral response information, providing a more comprehensive feature representation for target semantic recognition.

[0034] The fusion quality assessment unit evaluates the quality of the fused feature map and generates feedback signals to the multi-scale adaptive feature extraction module 1. The quality assessment uses two metrics: information entropy and feature separability. Information entropy measures the amount of information contained in the fused feature map, calculated by statistically analyzing the probability distribution of each pixel value in the feature map and then calculating the Shannon entropy. Feature separability measures the degree of separation between different categories of targets in the feature space, using the ratio of inter-class variance to intra-class variance as the metric. If the information entropy is too low or the feature separability is insufficient, it indicates that the current feature extraction weight configuration is not ideal, and feedback is needed to adjust the weight allocation of the multi-scale adaptive feature extraction module 1.

[0035] The cross-modal semantic fusion module 2 and the multi-scale feature adaptive extraction module 1 form the first closed loop, and the result of the fusion quality evaluation drives the dynamic optimization of the feature extraction strategy. This closed-loop mechanism ensures deep synergy between spatial features and spectral features, avoiding the semantic inconsistency problem caused by the independent processing of the two types of features in traditional methods.

[0036] In a preferred embodiment of the present invention, the NDVI calculated by the spectral feature extraction unit is defined as (NIR-R) / (NIR+R), where NIR is the near-infrared reflectance and R is the red light reflectance. NDWI is defined as (G-NIR) / (G+NIR), where G is the green light reflectance. NDBI is defined as (SWIR-NIR) / (SWIR+NIR). For images lacking the short-wave infrared band, (B-NIR) / (B+NIR) can be used as an approximation, where B is the blue light reflectance. These spectral indices exhibit significant differences across different target categories. For example, NDVI is positive in vegetated areas and negative or close to 0 in non-vegetated areas; NDWI is positive in water areas and negative in non-water areas. The adaptive channel attention mechanism dynamically adjusts the weights of each spectral feature based on the differences in sensitivity of different target categories to spectral features. When detecting vegetation targets, NDVI has a higher weight; when detecting water targets, NDWI has a higher weight; and when detecting building targets, NDBI has a higher weight.

[0037] Reference Figure 4 The contextual association enhancement module 3, located after the cross-modal semantic fusion module 2, is responsible for modeling the spatial relationships between the target and its surrounding environment, enhancing target features using contextual information. In high-resolution remote sensing imagery, target identification depends not only on the target's own features but also on the contextual information of its surrounding environment. For example, vehicles are typically found on roads, buildings are often surrounded by roads and greenery, and bodies of water may be surrounded by vegetation. By explicitly modeling these contextual relationships, this prior knowledge can be effectively utilized to improve the accuracy and robustness of target identification.

[0038] The context association enhancement module 3 includes a global context modeling unit, a local context modeling unit, a graph convolution enhancement unit, and a quality assessment feedback unit. The global context modeling unit constructs long-range dependencies between all pixels in the fused feature map based on a self-attention mechanism. The self-attention mechanism flattens the feature map into a sequence, calculates the correlation between each position in the sequence and all other positions, and generates an attention weight matrix. This matrix describes the global associations between pixels; even if two pixels are spatially far apart, as long as they are semantically related, a connection can be established through the self-attention mechanism. The global context graph preserves the overall structural information of the scene, providing macroscopic contextual support for object recognition.

[0039] The local context modeling unit extracts local neighborhood features of each target region based on deformable convolution. Deformable convolution is an extension of standard convolution; its sampling positions are not fixed grids but can adaptively adjust according to input features. This allows the convolution kernel to flexibly cover irregularly shaped target regions, extracting more accurate local features. The local context graph captures the fine-grained relationships between the target and its neighboring regions, providing micro-contextual support for the precise localization of target boundaries.

[0040] The graph convolutional augmentation unit constructs a graph structure from the global and local context graphs, and updates node features through a graph convolutional network. In the graph structure, each pixel or target region is a node, and the relationships between nodes are edges. The graph convolutional network aggregates the features of each node with the features of its neighboring nodes through a message-passing mechanism, achieving contextual augmentation of features. Specifically, a graph attention network is used to assign different attention weights to different neighboring nodes, making semantically related nodes have a greater influence on the current node. After multiple layers of graph convolutional operations, node features not only contain their own local information but also incorporate contextual information from multi-hop neighbors, generating context-enhanced features.

[0041] The quality assessment feedback unit evaluates the quality of the context-enhanced features and generates feedback signals to the cross-modal semantic fusion module 2. The quality assessment uses two metrics: target boundary clarity and class confidence. Target boundary clarity is measured by calculating the feature gradient at the target boundary; a larger gradient indicates a clearer boundary. Class confidence is calculated by performing a preliminary classification on the context-enhanced features and calculating the confidence score of the classification result. If the target boundary is blurry or the class confidence is low, it indicates that the current cross-modal fusion strategy has failed to generate sufficiently discriminative features, requiring feedback to adjust the fusion weights of the cross-modal semantic fusion module 2.

[0042] The context association enhancement module 3 and the cross-modal semantic fusion module 2 form a second closed loop, with the results of context quality assessment driving the optimization of the cross-modal fusion strategy. This multi-layer closed-loop mechanism ensures the layer-by-layer refinement of feature representations, with each module receiving feedback from subsequent modules, achieving global collaborative optimization.

[0043] In the innovative algorithm design of this invention, the graph convolution enhancement unit uses the following innovative formula to update node features:

[0044] .

[0045] in, For nodes In the The feature vector of the layer, Neighboring nodes In the The feature vector of the layer, For nodes The set of neighboring nodes, For nodes and nodes Attention weights between For the first The learnable weight matrix of the layer, These are self-loop weighting coefficients, used to preserve the node's own feature information. The ReLU activation function is used. This formula introduces self-loop weighting coefficients. This approach aggregates neighbor features while preserving the original features of each node, avoiding over-smoothing. Attention weights The calculation uses the following formula:

[0046] .

[0047] in, For nodes and nodes The similarity score between them is calculated as follows:

[0048] .

[0049] in, A learnable attention parameter vector, The vector concatenation operation is represented by LeakyReLU, which is a ReLU activation function with leakage and a negative slope of 0.2. Preferably, the graph convolutional network contains 3 layers, each with a hidden dimension of 256, and self-loop weight coefficients... It is initialized to 0.5 and adaptively adjusted during training.

[0050] In one embodiment of the present invention, for a 512×512 pixel feature map, the global context modeling unit flattens it into a sequence of 262,144 pixels, and the complexity of calculating the self-attention matrix is ​​O(n log n). ,in To reduce computational complexity, a local window self-attention mechanism is employed. The feature map is divided into 32×32 windows, each containing 16×16=256 pixels. Self-attention is calculated only within each window, reducing the complexity to [missing value]. ,in For the number of windows, The number of pixels within the window is denoted as . The local context modeling unit uses a 3×3 deformable convolution, with the 9 sampling positions of the convolution kernel adaptively shifted according to the input features, with a maximum shift of 3 pixels. The graph structure constructed by the graph convolutional enhancement unit contains approximately 100,000 nodes (corresponding to candidate target regions), with each node connecting an average of 20 neighboring nodes. The graph convolutional network achieves 3-hop neighbor information aggregation through 3 layers of message passing.

[0051] Reference Figure 5 The dynamic weight closed-loop adjustment module 4 is the core control unit of the system of this invention. It is responsible for dynamically calculating the contribution of each module based on the output performance of each module and the final detection result, generating adaptive weight coefficients, and feeding back the weight coefficients to each front-end module to achieve global closed-loop optimization. This module is the key innovation of this invention, which distinguishes it from the prior art. By introducing a closed-loop feedback mechanism, it breaks the limitations of the traditional unidirectional processing flow and realizes the collaborative optimization of each module and the adaptive adjustment of parameters.

[0052] The dynamic weight closed-loop adjustment module 4 includes a contribution calculation unit, a weight optimization unit, a weight distribution unit, and a convergence judgment unit. The contribution calculation unit calculates the contribution score of each module based on the correlation between the output features of each module and the final detection result. Specifically, for the multi-scale feature adaptive extraction module 1, it calculates the spatial consistency between its output multi-scale feature map set and the finally detected target location; for the cross-modal semantic fusion module 2, it calculates the semantic consistency between its output fused feature map and the finally detected target category; for the context association enhancement module 3, it calculates the correlation between its output context enhancement features and the boundary accuracy of the final detection result. The contribution score uses a weighted combination of the intersection-over-union ratio (IoU) and classification confidence.

[0053] The weight optimization unit scores contributions and uses gradient descent to optimize the weight coefficients of each module. The optimization objective is to maximize overall detection accuracy while balancing the contribution distribution among modules to avoid performance bottlenecks caused by any single module becoming overly dominant. The weight coefficients are normalized using the Softmax function to ensure that the sum of the weights of all modules is 1. The weight optimization process employs the momentum gradient descent algorithm, with the learning rate initialized to 0.001 and gradually decreasing with each iteration.

[0054] The weight distribution unit distributes the optimized dynamic weight coefficients to the multi-scale feature adaptive extraction module 1, the cross-modal semantic fusion module 2, and the context association enhancement module 3, respectively. Each module adjusts its internal parameter configuration and feature processing strategy based on the received weight coefficients. For example, if the weight coefficient of the multi-scale feature adaptive extraction module 1 is high, it indicates a greater demand for multi-scale features in the current scene, and this module will enhance the extraction strength of features at each level; if the weight coefficient of the cross-modal semantic fusion module 2 is high, it indicates that spectral information plays an important role in the current target recognition, and this module will strengthen the fusion strength of spectral and spatial features.

[0055] The convergence judgment unit is used to determine whether the dynamic weight coefficients have converged. The judgment criterion is whether the change in the weight coefficients over five consecutive iterations is less than a preset threshold of 0.01. If convergence has not occurred, the next round of iteration adjustment continues; if convergence has occurred, the current weight coefficients are fixed and iteration stops. The converged weight coefficients reflect the optimal contribution distribution of each module to the current scenario, and the system enters a stable, high-performance operating state.

[0056] The dynamic weight closed-loop adjustment module 4 forms the global closed loop of the system of this invention, feeding back the detection results from the backend to the feature extraction and fusion module at the frontend, thereby achieving end-to-end collaborative optimization. This closed-loop mechanism transforms the system from a simple pipeline process into an adaptive intelligent system that can dynamically adjust the working mode of each module according to actual detection needs and scenario characteristics, significantly improving the robustness and generalization ability of the system.

[0057] In the innovative algorithm design of this invention, the contribution calculation adopts the following innovative formula:

[0058] ,

[0059] in, For module Contribution score For module The intersection-union ratio (IoU) of the output features and the final detected target. For module The classification confidence of the output features. For module The reciprocal of the processing speed (normalized feature processing time) , and These are weighting coefficients, representing the importance of positioning accuracy, classification accuracy, and computational efficiency, respectively. In a preferred embodiment, , , This indicates that positioning accuracy and classification accuracy are prioritized while also considering computational efficiency. Weight optimization uses the following momentum gradient descent formula:

[0060] ,

[0061] in, For module In the The weight coefficients of the next iteration, The current weighting coefficient, This is the momentum coefficient, with a value of 0.9. For momentum term, The learning rate is initially set to 0.001. loss function Weights gradient, Defined as the weighted sum of detection accuracy loss and computational efficiency loss. The normalization formula for the weighting coefficients is:

[0062] ,

[0063] in, These are unnormalized weighting coefficients. This represents the total number of modules, with a value of 3 (corresponding to the multi-scale feature adaptive extraction module, cross-modal semantic fusion module, and context association enhancement module). The normalized weight coefficients satisfy... .

[0064] In one embodiment of the present invention, when processing a remote sensing image of an urban scene containing buildings, roads, and vehicles, the weight coefficients converge after 10 iterations. The weight of the multi-scale feature adaptive extraction module 1 is 0.35, the weight of the cross-modal semantic fusion module 2 is 0.30, and the weight of the context association enhancement module 3 is 0.35. This indicates that in this scene, multi-scale features and context association contribute significantly to target detection, while spectral information plays a relatively minor role. When processing a remote sensing image of a natural scene containing vegetation and water, the weight distribution changes to 0.25, 0.45, and 0.30, with the weight of the cross-modal semantic fusion module 2 significantly increasing, indicating that spectral information plays a more important role in target recognition in natural scenes. This adaptive weight adjustment capability enables the system to maintain excellent detection performance in different scenes.

[0065] Reference Figure 6 The target precise localization and attribute extraction module 5 is the output unit of the system of this invention. It is responsible for performing precise target localization, segmentation, and attribute extraction based on the context enhancement features and dynamic weight coefficients generated by the aforementioned modules, and generating the final detection result. This module combines high-level semantic features with low-level detail features to achieve pixel-level precise segmentation and rich attribute interpretation.

[0066] The target precise localization and attribute extraction module 5 includes a region proposal generation unit, a target segmentation unit, a category recognition unit, an attribute extraction unit, and a result integration unit. The region proposal generation unit generates candidate target regions using a Region Proposal Network (RPN) based on context-enhanced features and dynamic weight coefficients. The RPN is a lightweight fully convolutional network that predicts multiple candidate boxes of different scales and aspect ratios at each location in the feature map and outputs a target probability score for each candidate box. The target probability score represents the probability that the candidate box contains the target; candidate boxes with scores higher than a threshold of 0.7 are retained as candidate target regions. During candidate box generation, dynamic weight coefficients are used to adjust the contribution of features at different scales, ensuring that the candidate box generation strategy is consistent with the front-end feature extraction strategy.

[0067] The target segmentation unit performs pixel-level segmentation on candidate target regions, extracting the precise contours of the targets. The segmentation process employs a fully convolutional network, generating a binary segmentation mask for each candidate region. Pixels with a value of 1 in the mask belong to the target, while pixels with a value of 0 belong to the background. To improve segmentation accuracy, the segmentation network uses residual connections and skip connections to fuse high-level semantic features with low-level detail features, preserving the boundary details of the target. The segmentation results undergo morphological post-processing, including opening operations to remove small noise, closing operations to fill internal holes, and boundary smoothing, generating complete and smooth target contours.

[0068] The category recognition unit performs semantic classification on the segmented targets to determine their category labels. Category recognition employs a multilayer perceptron classifier, mapping the feature vectors of the target region to a category space and outputting confidence scores for each category. The category with the highest confidence score is used as the target's category label. In embodiments of this invention, the system supports the recognition of 10 common land cover categories, including buildings, roads, vehicles, vegetation, water bodies, and bare soil, with an average accuracy rate exceeding 92%.

[0069] The attribute extraction unit extracts the target's spatial coordinates, geometric parameters, spectral properties, and texture features to generate an attribute description. The spatial coordinates include the coordinates of the center point of the target's bounding rectangle. ,width and height And the precise contour point coordinate sequence of the target segmentation mask. Geometric parameters include the target area. ,perimeter Circularity Rectangularity and principal direction angle Circularity is defined as... The value ranges from 0 to 1. The roundness of a circular target is close to 1, while the roundness of an irregular target is close to 0. Rectangularity is defined as... The area represented by represents the proportion of the target area within its bounding rectangle. The principal orientation angle is obtained by calculating the eigenvectors of the covariance matrix of the target region, representing the main extension direction of the target. Spectral properties include the average reflectance, standard deviation, and spectral index values ​​of the target region in each band. Texture features are extracted using the gray-level co-occurrence matrix, including statistics such as contrast, energy, homogeneity, and entropy.

[0070] The results integration unit integrates target location, category, and attribute information to generate the final detection results. The detection results are output in a structured data format, with each detected target including category label, confidence score, location coordinates, segmentation mask, and attribute description. The detection results can be visualized and overlaid on the original remote sensing image, or exported as a vector or database format for subsequent analysis and application. The results integration unit also calculates the overall detection accuracy, including precision, recall, and F1 score, and feeds the accuracy back to the dynamic weight closed-loop adjustment module 4 for the next round of weight optimization.

[0071] The target precise localization and attribute extraction module 5 and the dynamic weight closed-loop adjustment module 4 form the final closed loop, and the accuracy of the detection results directly drives the parameter optimization of the entire system. This end-to-end closed-loop mechanism ensures continuous improvement and performance enhancement of the system, enabling it to learn and optimize from each detection task and continuously adapt to new scenarios and target types.

[0072] In one embodiment of the present invention, for a 2048×2048 pixel remote sensing image of an urban scene, the region proposal generation unit generates approximately 2000 candidate target regions. After non-maximum suppression (NMS) processing, approximately 500 candidate regions are retained. The target segmentation unit performs pixel-level segmentation on these 500 candidate regions, with an average segmentation time of 5ms per region. The category recognition unit classifies the targets into categories such as buildings, roads, and vehicles. The confidence threshold for category recognition is set to 0.8, meaning that only detection results with a confidence score higher than 0.8 are output. Among the geometric shape parameters extracted by the attribute extraction unit, the circularity of buildings is typically between 0.3 and 0.6, and the circularity of vehicles is between 0.6 and 0.9. These statistical regularities can be used to further optimize the classification results. Ultimately, 150 buildings, 8 roads, and 45 vehicles were detected, with a precision of 94%, a recall of 89%, and an F1 score of 91.4%.

[0073] The data flow of this invention begins with the input high-resolution remote sensing image, and sequentially passes through the forward processing of the multi-scale feature adaptive extraction module 1, the cross-modal semantic fusion module 2, the context association enhancement module 3, and the target precise localization and attribute extraction module 5 to generate the final detection result. Simultaneously with the forward processing, the output performance indicators of each module are fed back to the dynamic weight closed-loop adjustment module 4, which calculates the contribution of each module and generates dynamic weight coefficients. These dynamic weight coefficients are then backpropagated to each preceding module, driving the adaptive adjustment of the parameters of each module.

[0074] Specifically, there are three closed-loop feedback mechanisms:

[0075] The first closed loop: the cross-modal semantic fusion module 2 feeds back the quality evaluation results of the fused feature map to the multi-scale adaptive feature extraction module 1, adjusting the weight allocation of feature extraction. This closed loop ensures deep synergy between spatial and spectral features and optimizes the low-level feature representation.

[0076] The second closed loop: The context association enhancement module 3 feeds back the quality evaluation results of the context enhancement features to the cross-modal semantic fusion module 2, adjusting the fusion weights. This closed loop ensures the effective integration of contextual information and multimodal features, improving the discriminative power of mid-level features.

[0077] The third closed loop: Target precise localization and attribute extraction module 5 feeds back the accuracy of the detection results to the dynamic weight closed loop adjustment module 4, which coordinates the weight coefficients of each module. This closed loop achieves global optimization, ensuring the collaborative work of each module and maximizing overall performance.

[0078] This multi-layered closed-loop feedback mechanism breaks through the limitations of traditional unidirectional processing flows, achieving a complete closed loop of "forward transmission → performance evaluation → reverse feedback → parameter adjustment," ensuring deep coupling and collaborative optimization among modules. The output of each module not only serves as the input to the next module but also influences the working state of preceding modules through the feedback mechanism, forming an organic whole system. This closed-loop collaborative mechanism is the core innovation that distinguishes this invention from existing technologies, significantly improving the system's intelligence level and adaptability.

[0079] During system operation, the closed-loop feedback mechanism achieves three synergistic effects among the modules: First, a mutually reinforcing effect, where spectral features enhance the semantic discriminative power of spatial features, while the optimized spatial features, in turn, guide the extraction of spectral features; second, a superimposed synergistic effect, where the superimposed processing of multi-scale features, cross-modal fusion, and contextual enhancement produces an enhancement effect far exceeding that of a single technology; and third, a complementary advantage effect, where the advantages of each module in different scenarios and target types are fully utilized through weight adjustments, compensating for the limitations of a single module. These three synergistic effects work together to result in a non-linear growth characteristic of 1+1>2 in the overall system performance, with an average target detection accuracy improvement of 15%-25% compared to traditional methods using independent modules, and a more significant improvement of over 30% in small target detection accuracy.

[0080] To verify the effectiveness of the system of this invention, experiments were conducted on several publicly available remote sensing image datasets, including DOTA, NWPU VHR-10, and RSOD datasets. The hardware platform used in the experiments was a workstation equipped with an NVIDIA RTX 3090 GPU and 32GB of video memory, and the software platform was the Ubuntu 20.04 operating system and the PyTorch 1.10 deep learning framework.

[0081] The DOTA dataset contains 2806 high-resolution remote sensing images, covering 15 categories of targets, including aircraft, ships, oil tanks, baseball fields, tennis courts, basketball courts, athletic fields, ports, bridges, large vehicles, small vehicles, helicopters, roundabouts, football fields, and swimming pools. Image sizes range from 800×800 to 4000×4000 pixels, with spatial resolutions ranging from 0.1m to 1.0m. The system of this invention achieves an average precision (mAP) of 78.6% on this dataset, a 13.4 percentage point improvement over the comparison method based on U-Net and BDCN (mAP 65.2%), and a 7.3 percentage point improvement over the method based on Faster R-CNN (mAP 71.3%). In small target detection, for targets smaller than 32×32 pixels, the detection accuracy of the system of this invention is 62.5%, a 14.4 percentage point improvement over the comparison method (48.1%), representing an improvement of 29.9%.

[0082] On the NWPU VHR-10 dataset, which contains 800 high-resolution remote sensing images covering 10 categories of targets including aircraft, ships, oil tanks, baseball fields, tennis courts, basketball courts, athletic fields, ports, bridges, and vehicles, the system of this invention achieves an average accuracy (mAP) of 91.8% on this dataset, an improvement of 9.2 percentage points compared to comparative methods. In terms of processing speed, for a 2048×2048 pixel image, the processing time per image is 1.8 seconds, meeting the real-time requirements of practical applications.

[0083] Regarding attribute extraction accuracy, tests were conducted on 100 urban scene images containing buildings, and the extraction errors of building geometric parameters were statistically analyzed. The average relative error for building area was 4.2%, the average relative error for perimeter was 3.8%, and the average absolute error for principal orientation angle was 5.1°, all achieving high extraction accuracy. In terms of spectral characteristic extraction, the calculation errors for spectral indices such as NDVI, NDWI, and NDBI were all less than 0.05, effectively supporting fine-grained target classification and attribute interpretation.

[0084] Ablation experiments further validated the roles of each core module and the closed-loop mechanism. Removing the dynamic weight closed-loop adjustment module decreased the system's average accuracy (mAP) by 8.3 percentage points, indicating that the closed-loop feedback mechanism plays a crucial role in improving overall performance. Removing the cross-modal semantic fusion module decreased the average accuracy by 5.7 percentage points, demonstrating that the introduction of spectral information significantly improved target recognition accuracy. Removing the context association enhancement module decreased the average intersection-union ratio (IoU) of the target boundary by 0.12, indicating that context modeling improved target localization accuracy.

[0085] Experimental results show that the system of the present invention, by constructing core modules such as multi-scale feature adaptive extraction, cross-modal semantic fusion, context association enhancement, and dynamic weight closed-loop adjustment, achieves deep coupling and collaborative optimization among modules, significantly improving the accuracy, robustness, and real-time performance of target detection and intelligent interpretation of high-resolution remote sensing images. It has reached an advanced level on multiple public datasets and has important theoretical and practical value.

[0086] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make many possible variations and modifications to the technical solutions of the present invention, or modify them into equivalent embodiments, without departing from the scope of the technical solutions of the present invention. Therefore, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention should fall within the protection scope of the technical solutions of the present invention.

Claims

1. A high-resolution remote sensing image target detection and intelligent interpretation system, characterized in that ,comprising: a multi-scale feature adaptive extraction module, configured to receive an input high-resolution remote sensing image, perform feature extraction of different scales on the remote sensing image based on a hierarchical attention mechanism, and generate a multi-scale feature map group containing shallow detail features, middle texture features and deep semantic features; a cross-modal semantic fusion module, connected with the multi-scale feature adaptive extraction module, configured to receive the multi-scale feature map group and spectral information of the original remote sensing image, perform cross-modal fusion of spectral features and spatial features through an adaptive channel attention mechanism, generate a fusion feature map, and feed statistical characteristics of the fusion feature map back to the multi-scale feature adaptive extraction module to adjust weight distribution of feature extraction; a context association enhancement module, connected with the cross-modal semantic fusion module, configured to receive the fusion feature map, construct a global context relationship graph and a local context relationship graph, model spatial association between a target and a surrounding environment based on a graph convolution network, generate context enhanced features, and reversely adjust fusion weights of the cross-modal semantic fusion module according to a quality evaluation result of the context enhanced features; a dynamic weight closed-loop adjustment module, connected with the multi-scale feature adaptive extraction module, the cross-modal semantic fusion module and the context association enhancement module respectively, configured to receive intermediate features and final detection results output by each module, calculate contribution scores of each module, generate a dynamic weight coefficient group based on the contribution scores, and feed the dynamic weight coefficient group back to the multi-scale feature adaptive extraction module, the cross-modal semantic fusion module and the context association enhancement module respectively, to realize adaptive adjustment of parameters of each module; a target precise positioning and attribute extraction module, connected with the context association enhancement module and the dynamic weight closed-loop adjustment module, configured to receive the context enhanced features and the dynamic weight coefficient group, generate candidate target regions based on a weighted region proposal network, perform precise segmentation and class identification on the candidate target regions, extract spatial position, geometric shape, spectral characteristics and semantic categories of the target, generate a detection result containing target position coordinates, class labels and attribute descriptions, and feed an accuracy rate of the detection result back to the dynamic weight closed-loop adjustment module to update weight coefficients. 2.The high-resolution remote sensing image target detection and intelligent interpretation system of claim 1, characterized in that The multi-scale feature adaptive extraction module comprises: a feature pyramid construction unit, configured to perform down-sampling on the remote sensing image through a plurality of convolution layer groups, and generate feature levels of different resolutions; a hierarchical attention weighting unit, configured to apply spatial attention and channel attention to each feature level, and adaptively adjust weights of each level according to target scale distribution; a feature fusion unit, configured to perform up-sampling and splicing on weighted feature levels, and generate the multi-scale feature map group. 3.The high-resolution remote sensing image target detection and intelligent interpretation system of claim 1, wherein The cross-modal semantic fusion module comprises: a spectral feature extraction unit, configured to extract a spectral feature vector from a multi-spectral band of the original remote sensing image; an adaptive channel attention unit, configured to calculate channel correlation of the spectral feature and spatial features in the multi-scale feature map group, and generate channel attention weights. The weighted fusion unit is configured to perform weighted fusion on the spectral features and the spatial features based on the channel attention weights to generate the fused feature map. The fusion quality evaluation unit is configured to calculate information entropy and feature separability indexes of the fused feature map to generate a feedback signal to the multi-scale feature adaptive extraction module.

4. The high-resolution remote sensing image target detection and intelligent interpretation system of claim 1, wherein The context correlation enhancement module includes: The global context modeling unit is configured to construct long-range dependency relationships between all pixel points in the fused feature map based on a self-attention mechanism to generate a global context relationship graph. The local context modeling unit is configured to extract local neighborhood features of each target region based on a deformable convolution to generate a local context relationship graph. The graph convolution enhancement unit is configured to construct the global context relationship graph and the local context relationship graph into a graph structure, update node features through a graph convolution network, and generate the context-enhanced features. The quality evaluation feedback unit is configured to evaluate target boundary definition and class confidence of the context-enhanced features to generate a quality evaluation result feedback to the cross-modal semantic fusion module.

5. The high resolution remote sensing imagery target detection and intelligent interpretation system of claim 1, wherein The dynamic weight closed-loop adjustment module includes: The contribution degree calculation unit is configured to calculate contribution degree scores of each module based on correlations between output features of each module and the final detection result. The weight optimization unit is configured to optimize weight coefficients of each module using a gradient descent method according to the contribution degree scores. The weight distribution unit is configured to distribute the optimized dynamic weight coefficients to the multi-scale feature adaptive extraction module, the cross-modal semantic fusion module, and the context correlation enhancement module, respectively. The convergence judgment unit is configured to judge whether the dynamic weight coefficients converge, and if not, continue to iteratively adjust, and if so, fix the weight coefficients.

6. The high resolution remote sensing imagery target detection and intelligent interpretation system of claim 1, wherein The target precise positioning and attribute extraction module includes: The region proposal generation unit is configured to generate candidate target regions using a region proposal network based on the context-enhanced features and the dynamic weight coefficients. The target segmentation unit is configured to perform pixel-level segmentation on the candidate target regions to extract accurate contours of the targets. The class recognition unit is configured to perform semantic classification on the segmented targets to determine class labels of the targets. The attribute extraction unit is configured to extract spatial position coordinates, geometric shape parameters, spectral characteristics, and texture features of the targets to generate attribute descriptions. The result integration unit is configured to integrate target position, class, and attribute information to generate the detection result and calculate an accuracy rate feedback to the dynamic weight closed-loop adjustment module.

7. The high-resolution remote-sensing image target detection and intelligent interpretation system of claim 1, wherein In the multi-scale feature adaptive extraction module, the hierarchical attention mechanism includes a spatial attention branch and a channel attention branch, the spatial attention branch is configured to calculate importance weights of different spatial positions in the feature map, the channel attention branch is configured to calculate importance weights of different feature channels, and the outputs of the two branches are multiplied and fused to act on the original feature map.

8. The high-resolution remote-sensing image target detection and intelligent interpretation system of claim 1, wherein The spectral features in the cross-modal semantic fusion module include a normalized vegetation index, a normalized water index, and a normalized building index, and the adaptive channel attention mechanism dynamically adjusts the weights of the spectral features according to the sensitivity differences of the spectral features to different target categories.

9. The high-resolution remote sensing image target detection and intelligent interpretation system of claim 1, wherein In the dynamic weight closed-loop adjustment module, the calculation of the contribution score is based on the intersection over union between the output feature maps of each module and the real label and the cross-entropy loss of category prediction, and the weight coefficient of the module with higher contribution is larger.

10. The high resolution remote sensing imagery target detection and intelligent interpretation system of claim 1, wherein The system further comprises a target tracking module for associatively tracking the same target in continuous multiple frames of remote sensing images, so as to further improve the stability and accuracy of target detection through time sequence information.

Images