Radar image intelligent enhancement and identification method and system based on multi-model fusion

Abstract

The invention belongs to the technical field of image enhancement and recognition, and particularly relates to a radar image intelligent enhancement and recognition method and system based on multi-model fusion, and the method comprises the steps: carrying out the adaptive suppression of speckle noise of an original radar image; feature point detection is carried out, robust transformation matrix estimation and adaptive contrast enhancement are carried out, and a corrected and enhanced image is output; utilizing the generative adversarial network and multi-loss function collaborative constraint to obtain a texture reconstruction image; establishing an image-semantic double-flow network architecture, performing cross-modal attention fusion to obtain a fusion feature map, and outputting a target recognition result; performing time phase division on the texture reconstruction image, judging a change type, and outputting a change detection result; and outputting a processing report including the enhanced image, the target list and change analysis. According to the method, noise suppression, correction enhancement, texture reconstruction, target recognition and change detection are integrated, the defect of fragmentation processing in the traditional technology is overcome, and the overall processing performance and the actual application adaptability are improved.

Description

Technical Field

[0001] This invention belongs to the field of image enhancement and recognition technology, specifically relating to a method and system for intelligent enhancement and recognition of radar images based on multi-model fusion. Background Technology

[0002] Radar imagery, with its multi-dimensional detection advantages, can accurately capture high-resolution information about the Earth's surface, playing a crucial role in key areas such as military reconnaissance, disaster monitoring, and land resource exploration. It is a core data source for achieving accurate perception across multiple scenarios. However, due to the combined effects of electromagnetic wave coherent scattering characteristics, radar platform motion deviations, and interference from complex environments, raw radar images face unavoidable technical bottlenecks: speckle noise can obscure the texture details of targets such as equipment structures and building edges; geometric distortion can cause target position shifts and shape stretching; and in low signal-to-noise ratio scenarios, the textures of weakly scattering targets are even more blurred and difficult to discern. These problems directly limit the effectiveness of subsequent applications of radar imagery.

[0003] To address these issues, various targeted technologies and models have been proposed, achieving some success in local processing stages (such as single noise suppression and basic geometric correction). However, the overall solutions still have significant limitations: First, the processing flow is fragmented, focusing only on optimizing a single step without a collaborative design across the entire process. This can easily lead to a chain reaction of problems, such as excessive denoising resulting in the loss of textures needed for recognition and texture misalignment after correction. Second, denoising techniques have poor adaptability. Traditional fixed-kernel filtering methods can easily cause blurred image edges, while deep learning models rely on a large number of labeled samples and have weak generalization ability in new scenarios. Third, geometric correction and texture enhancement are disconnected. Texture quality is not improved synchronously after correction, and performing texture enhancement alone can amplify residual noise. Fourth, target recognition is not robust enough, relying solely on visual features for judgment without associating with target semantic information (such as size and scattering characteristics). This results in a high misjudgment rate for targets with similar textures, failing to meet the actual needs of intelligent radar image processing in complex scenarios. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of this invention is to provide a radar image intelligent enhancement and recognition method and system based on multi-model fusion, which integrates noise suppression, correction enhancement, texture reconstruction, target recognition and change detection, solves the drawbacks of fragmented processing in traditional technology, and improves the overall processing performance and adaptability to practical applications.

[0005] To achieve the above objectives, this invention provides a radar image intelligent enhancement and recognition method based on multi-model fusion, comprising the following steps: For the acquired raw radar images, multi-scale dynamic pyramid feature decomposition, channel-space collaborative attention and bidirectional residual fusion are used to adaptively suppress speckle noise and obtain denoised images. Feature point detection is performed on the denoised image. The detected feature points are combined with radar orbit parameters to perform robust transformation matrix estimation and adaptive contrast enhancement, and the corrected and enhanced image is output. For the corrected and enhanced image, the texture details lost after noise suppression are reconstructed by using generative adversarial networks and multi-loss function co-constraints to obtain the texture reconstructed image; A dual-stream image-semantic network architecture is established to acquire image features and semantic features of texture reconstructed images and perform cross-modal attention fusion to obtain a fused feature map. Multi-scale target detection is then performed on the fused feature map to output the target recognition result. The texture reconstruction image is divided into temporal phases, features are extracted using Siamese Transformer, feature maps of each phase are output and difference fusion is performed, and the change region mask is initially obtained through adaptive threshold segmentation to determine the change type and output the change detection result. The system integrates and enhances the corrected and enhanced image, target recognition results, and change detection results, and outputs a processing report that includes the enhanced image, target list, and change analysis.

[0006] As a preferred embodiment of the present invention, the process of obtaining the denoised image includes: The original radar image is a grayscale image, and the pixel value corresponds to the radar echo intensity. A multi-scale dynamic pyramid feature decomposition structure is constructed to decompose the input original radar image into a multi-resolution layer, resulting in a high-resolution layer that focuses on small-scale noise and a low-resolution layer that targets large-scale noise. In both high-resolution and low-resolution layers, channel-spatial collaborative attention (RC-SAM) is embedded. In high-resolution layers, channel attention prioritizes learning the sensitivity of channels with concentrated small-scale noise, while spatial attention accurately locates local areas with dense noise. In low-resolution layers, channel attention focuses on channels dominated by large-scale noise, while spatial attention locates large-scale noise regions, thus achieving targeted filtering of noise at different scales. For the high-resolution and low-resolution layers after RC-SAM processing, a bidirectional residual connection is used to transfer the texture features of the high-resolution layer to the low-resolution layer. Then, feature fusion is performed on the high-resolution and low-resolution layers to output the denoised image.

[0007] As a preferred embodiment of the present invention, the process of outputting the corrected and enhanced image is as follows: Feature point detection is performed on the denoised image, and stable feature points are screened using local entropy thresholding to eliminate pseudo feature points caused by noise interference. The detected feature points are combined with radar orbit parameters, the robust transformation matrix is ​​estimated using the RANSAC algorithm, the geometric distortion in the denoised image is fitted, and the geometric transformation is completed based on the robust transformation matrix to obtain a geometrically aligned image. For geometrically aligned images, the ACE algorithm is used to construct a dynamic window centered on each pixel, calculate the mean and variance of gray levels within the window, adaptively adjust the pixel gray levels, and output a corrected and enhanced image.

[0008] As a preferred embodiment of the present invention, the method of using local entropy threshold to screen stable feature points specifically involves extracting the gray-level distribution of the local neighborhood of each candidate feature point in the denoised image, quantifying the texture complexity of the region, and expressing it as follows: ; In the formula, H is the local entropy and L is the number of gray levels; It represents the percentage of pixels with grayscale value i within a local neighborhood. Set the local entropy threshold H C For candidate feature points, if H ≥ H C If H < H, then the candidate feature point is a stable feature point and should be retained. C If the candidate feature point is false, it should be removed.

[0009] As a preferred embodiment of the present invention, the process of obtaining the texture reconstruction image is as follows: The generator-discriminator architecture is set up. The generator adopts an improved U-Net structure, and sparse convolution is introduced in the encoder to focus on key texture features. The decoder gradually recovers high-resolution texture through transposed convolution and attention gating. The discriminator adopts the PatchGAN structure to judge the authenticity of the image on a pixel block basis, avoiding local texture distortion caused by global discrimination. The design incorporates a hybrid loss function that combines adversarial loss, texture-aware loss, and pixel-level loss. Based on the texture type of the input corrected and enhanced image, similar texture samples are dynamically selected from a high-resolution radar image sample library as supervisory data. The corrected and enhanced image is then input into the generator to generate an initial texture reconstruction image. The loss value is calculated using the hybrid loss function and backpropagated to the generator and discriminator. The model parameters are iteratively optimized until the loss converges, resulting in a Rad-GAN model with accurate texture reconstruction capabilities. The trained Rad-GAN model takes the corrected and enhanced image as input and outputs a textured reconstructed image.

[0010] As a preferred embodiment of the present invention, the adversarial loss adopts least squares adversarial loss, the texture-aware loss adopts perceptual loss based on pre-trained CNN, and the pixel-level loss adopts L1 loss. Weights are set for the adversarial loss, texture-aware loss and pixel-level loss respectively, and weighted fusion is performed to obtain the final hybrid loss.

[0011] As a preferred embodiment of the present invention, the process of outputting the target recognition result includes: A target semantic knowledge base is set up, and an image-semantic dual-stream network architecture is established, including parallel networks for image stream and semantic stream. The image stream adopts a CNN-Transformer hybrid structure. First, the low-level visual features of the image are reconstructed by extracting texture through ResNet, and then the global correlation between features is captured by the Transformer encoder to generate image features. The semantic stream converts the attributes in the target semantic knowledge base into semantic embedding vectors, and aligns them with the image features through the semantic encoder to obtain semantic features. The similarity matrix between image features and semantic features is calculated. Through attention weight allocation, cross-modal attention fusion of image features and semantic features is performed to obtain a fused feature map. At the same time, a dynamic weight update mechanism is introduced to adjust the contribution ratio of image stream and semantic stream in real time according to the recognition confidence. A feature pyramid network is used to perform multi-scale target detection on the fused feature map. Duplicate detection boxes are removed by non-maximum suppression, and the target recognition results are output.

[0012] As a preferred embodiment of the present invention, the process of outputting the change detection result includes: A parameter-sharing Siamese Transformer network is constructed to extract features from the texture reconstruction images at each time step. The self-attention mechanism of the Transformer encoder is used to capture local texture and global structural features. The corresponding feature points at each time step are aligned through the cross-attention mechanism to output the feature maps of each time step. The differences between the feature maps of each time phase are calculated using both Euclidean distance and cosine similarity. The difference results are then fused, and the change region mask is initially obtained through adaptive threshold segmentation. The change region mask is post-processed using a conditional random field to optimize the boundary of the change region. At the same time, the features of the change region are matched with a preset change type library, and the change type is determined by a classifier to output refined change detection results.

[0013] As a preferred embodiment of the present invention, the method for outputting the processing report is as follows: Establish the correlation between image region, target recognition, and change detection. Utilize an ensemble classifier to perform secondary determination of target type in the target recognition results and change type in the change detection. Finally, integrate and enhance the corrected and enhanced image, target recognition results, and change detection results, and output a processing report. The processing report includes: Enhanced image with target location and changed areas marked; A list of targets, including target type, target recognition confidence level, and target coordinates; Results of change analysis including differences in change type, area of ​​change, and time of change.

[0014] A radar image intelligent enhancement and recognition system based on multi-model fusion, including the following methods: Denoising module: For the acquired raw radar image, multi-scale dynamic pyramid feature decomposition, channel-space collaborative attention and bidirectional residual fusion are used to adaptively suppress speckle noise and obtain the denoised image; Correction and enhancement module: Performs feature point detection on the denoised image, combines the detected feature points with radar trajectory parameters, performs robust transformation matrix estimation and adaptive contrast enhancement, and outputs a corrected and enhanced image; Texture Reconstruction Module: For the corrected and enhanced image, the texture details lost after noise suppression are reconstructed by using a generative adversarial network and multiple loss functions in a collaborative manner, thus obtaining a texture reconstructed image; Target recognition module: Establish an image-semantic dual-stream network architecture, acquire image features and semantic features of texture reconstructed images, perform cross-modal attention fusion to obtain a fused feature map, perform multi-scale target detection on the fused feature map, and output the target recognition result; Change detection module: Divides the texture reconstruction image into temporal phases, extracts features through Siamese Transformer, outputs feature maps of each phase and performs difference fusion, obtains the change region mask through adaptive threshold segmentation, determines the change type, and outputs the change detection results; Integration and Enhancement Module: Integrates and enhances the corrected and enhanced image, target recognition results, and change detection results, and outputs a processing report containing the enhanced image, target list, and change analysis.

[0015] The beneficial effects of this invention are: This invention achieves precise matching between noise scale and filtering intensity by combining channel-space collaborative attention and dynamic resolution pyramid in multi-scale speckle noise suppression. Compared with traditional fixed filtering or single deep learning methods, it significantly improves the preservation of target textures (such as equipment structure textures and building edges). In geometric correction and enhancement, it integrates improved radar feature point detection (introducing local entropy threshold to screen stable feature points) and robust transformation matrix estimation (RANSAC algorithm with weights assigned to high-confidence feature points) to effectively eliminate spurious feature points and outliers, significantly reducing geometric errors. At the same time, it solves the problem of local over-darkness / over-brightness in the corrected image through adaptive contrast enhancement, avoiding information loss in subsequent separate processing. In texture reconstruction, relying on the generator-discriminator architecture, it combines multiple loss function constraints and dynamic sample adaptive matching to accurately repair texture details lost after noise suppression, solving the texture distortion problem that is prone to occur in traditional GANs and improving the reconstruction generalization in different scenarios.

[0016] In target recognition, this invention, based on image-semantic dual-stream network and cross-modal attention fusion, deeply associates visual features with target semantic attributes (size, scattering characteristics, etc.), significantly reducing the misjudgment rate of similar textured targets and significantly improving the robustness of recognition in complex scenes. In multi-temporal change detection, through the self-attention and cross-attention mechanism of Siamese Transformer, combined with Euclidean distance, cosine similarity dual-constraint difference calculation and CRF post-processing, it achieves high-precision localization of change areas and accurate determination of type (such as flooding, collapse, new construction), with processing efficiency meeting the needs of real-time monitoring.

[0017] The entire process of this invention revolves around a closed loop of processing, enhancement, identification, and verification. By integrating noise suppression, correction and enhancement, texture reconstruction, target recognition, and change detection through multi-source result association mapping and ensemble learning classification decision-making, it solves the drawbacks of fragmented processing in traditional technologies and improves overall processing performance and adaptability to practical applications. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the method steps of the present invention; Figure 2 This is a flowchart illustrating the principle of this invention; Figure 3 This is a system module diagram of the present invention. Detailed Implementation

[0019] The embodiments of the present invention will be further described below with reference to the accompanying drawings: Example 1: As Figure 1 and Figure 2 As shown, the intelligent enhancement and recognition method for radar images based on multi-model fusion includes the following steps: For the acquired raw radar images, multi-scale dynamic pyramid feature decomposition, channel-space collaborative attention and bidirectional residual fusion are used to adaptively suppress speckle noise and obtain denoised images. Feature point detection is performed on the denoised image. The detected feature points are combined with radar orbit parameters to perform robust transformation matrix estimation and adaptive contrast enhancement, and the corrected and enhanced image is output. For the corrected and enhanced image, the texture details lost after noise suppression are reconstructed by using generative adversarial networks and multi-loss function co-constraints to obtain the texture reconstructed image; A dual-stream image-semantic network architecture is established to acquire image features and semantic features of texture reconstructed images and perform cross-modal attention fusion to obtain a fused feature map. Multi-scale target detection is then performed on the fused feature map to output the target recognition result. The texture reconstruction image is divided into temporal phases, features are extracted using Siamese Transformer, feature maps of each phase are output and difference fusion is performed, and the change region mask is initially obtained through adaptive threshold segmentation to determine the change type and output the change detection result. The system integrates and enhances the corrected and enhanced image, target recognition results, and change detection results, and outputs a processing report that includes the enhanced image, target list, and change analysis.

[0020] The original radar image is a grayscale image, with pixel values ​​corresponding to radar echo intensity (including speckle noise and dynamic range imbalance). Acquiring the original radar image involves matching a suitable radar hardware system and data acquisition process to the specific application scenario (such as military reconnaissance, disaster monitoring, and land resource exploration): First, select the appropriate radar equipment based on the scenario requirements. For example, military reconnaissance can use airborne synthetic aperture radar (SAR) or ground-based phased array radar, while disaster monitoring and land resource exploration can use spaceborne SAR (such as the radar system carried by satellites like Sentinel-1) or airborne side-looking radar. Then, actively transmit electromagnetic waves towards the target area using the selected radar equipment, simultaneously receiving the echo signals reflected from the target. The equipment's built-in data acquisition module converts the amplitude, phase, and other analog signals of the echo signals into digital signals and records key metadata of the radar platform (such as azimuth, flight altitude, imaging mode, and resolution parameters). Finally, according to industry-standard radar data formats (such as the commonly used CEOS and HDF5 formats for SAR data), integrate and encapsulate the digital echo signals and metadata to form the original radar grayscale image, providing input data for subsequent multi-scale speckle noise adaptive suppression.

[0021] The process of obtaining the denoised image includes: A multi-scale dynamic pyramid feature decomposition structure (a multi-scale feature fusion model of pyramid structure) is constructed to decompose the input original radar image into multi-resolution layers, resulting in a high-resolution layer that focuses on small-scale noise (such as fine-grained spots) and a low-resolution layer that targets large-scale noise (such as regional block noise). In both high-resolution and low-resolution layers, channel-spatial collaborative attention (RC-SAM) is embedded. In high-resolution layers, channel attention prioritizes learning the sensitivity of channels with concentrated small-scale noise (e.g., strong scattering channels are more significantly affected by noise), while spatial attention precisely locates localized noise-dense regions. In low-resolution layers, channel attention focuses on channels dominated by large-scale noise, while spatial attention locates large-scale noise regions, enabling targeted filtering of noise at different scales. The filtering intensity is dynamically adjusted to avoid over-filtering of textured regions. For the high-resolution and low-resolution layers after RC-SAM processing, a bidirectional residual connection is used to transfer the texture features (such as edges and corners) of the high-resolution layer to the low-resolution layer, avoiding the loss of details caused by low-resolution processing. Then, feature fusion is performed on the high-resolution and low-resolution layers to output the denoised image.

[0022] The high-resolution layer refers to the pyramid layer with a resolution close to that of the original radar image. This layer can accurately preserve small-scale noise such as fine-grained spots in the original radar image, while completely preserving the texture details of targets such as equipment structure, building edges, and corners. The low-resolution layer is the pyramid layer obtained by downsampling (reducing the resolution) the original radar image. The image resolution of this layer is lower than that of the original radar image and the high-resolution layer. It can weaken the distribution characteristics of large-scale noise such as local detail interference and highlight regional block noise, which is convenient for efficiently processing large-scale noise areas and can reduce the amount of computation to improve processing efficiency.

[0023] To address the scale-dependent nature of speckle noise in radar images, adaptive noise suppression can be achieved at different resolutions while maximizing the preservation of target textures (such as equipment structure textures and building edges). By combining dynamic pyramids with bidirectional residual attention, precise matching between noise scale and filtering intensity is achieved, significantly improving texture preservation compared to traditional methods.

[0024] The process of outputting the corrected and enhanced image is as follows: Feature point detection is performed on the denoised image (including geometric distortion), and stable feature points are screened using local entropy thresholding to remove pseudo feature points caused by noise interference. The detected feature points are combined with radar orbit parameters (such as azimuth and altitude), and the robust transformation matrix (perspective transformation matrix, a 3×3 matrix describing the mapping relationship between distorted image pixels and real space pixel coordinates) is estimated using the Random Sample Consensus Algorithm (RANSAC). The geometric distortion in the denoised image is fitted, and the geometric transformation is completed based on the robust transformation matrix to obtain a geometrically aligned image. A weighting factor can be introduced to give higher weights to high-confidence feature points, effectively eliminating outliers caused by terrain undulations and better achieving geometric distortion correction. For geometrically aligned images, to address potential issues of local over-darkness / over-brightness, the ACE algorithm is employed to construct a dynamic window centered on each pixel, calculate the mean and variance of grayscale within the window, adaptively adjust pixel grayscale values, avoid noise amplification caused by traditional algorithms, enhance details in weak texture areas, and output a corrected and enhanced image.

[0025] The method of using local entropy thresholding to select stable feature points involves extracting the gray-level distribution of the local neighborhood for each candidate feature point in the denoised image, quantifying the texture complexity of the region, and representing it as follows: ; In the formula, H is the local entropy and L is the number of gray levels; It represents the percentage of pixels with grayscale value i within a local neighborhood. Set the local entropy threshold H C For candidate feature points, if H ≥ H C If H < H, then the candidate feature point is a stable feature point (such as a building vertex or road intersection), indicating that the region has rich texture and stable structure, and should be retained as a valid reference point for subsequent geometric correction. C If the region has a uniform texture (such as a uniform noise area or a flat, unstructured surface), then the candidate feature point is a pseudo-feature point and should be removed to avoid it interfering with the estimation of the subsequent transformation matrix.

[0026] By combining feature matching and physical modeling to achieve high-precision geometric correction, the contrast and texture clarity of the corrected image can be enhanced simultaneously. Combining feature point stability screening with physical parameter modeling improves the accuracy of geometric correction; simultaneous correction and enhancement avoids information loss caused by subsequent separate processing.

[0027] The process of obtaining a texture reconstruction image is as follows: The generator-discriminator architecture is set up. The generator adopts an improved U-Net structure, and sparse convolution is introduced in the encoder to focus on key texture features (such as bright spots of radar echoes of equipment). The decoder gradually recovers high-resolution textures through transposed convolution and attention gating. The discriminator adopts a PatchGAN structure to judge the authenticity of the image on a pixel block basis, avoiding local texture distortion caused by global discrimination. The design incorporates a hybrid loss function that integrates adversarial loss, texture-aware loss, and pixel-level loss. Based on the texture type (e.g., city, ocean, mountain) of the input corrected and enhanced image, similar texture samples are dynamically selected from a high-resolution radar image sample library (labeled with texture type) as supervisory data to improve adaptability to different scenes. The corrected and enhanced image is then input into the generator to generate an initial texture reconstruction image. The loss value is calculated using the hybrid loss function and backpropagated to the generator and discriminator to iteratively optimize the model parameters until the loss converges, resulting in a Rad-GAN model with accurate texture reconstruction capabilities (i.e., a generator-discriminator architecture model that incorporates sparse convolution). Using the trained Rad-GAN model, the texture reconstructed image is output as the input, with the correction and enhancement image (including texture blur) as the input.

[0028] The adversarial loss employs least-squares adversarial loss, the texture-aware loss uses a perceptual loss based on a pre-trained CNN, and the pixel-level loss uses L1 loss. Weights are assigned to the adversarial loss, texture-aware loss, and pixel-level loss, and then weighted and fused to obtain the final hybrid loss. The adversarial loss ensures the realism of the reconstructed texture, the texture-aware loss constrains the structural consistency between the reconstructed texture and the real sample, and the pixel-level loss reduces pixel errors and avoids texture blurring.

[0029] By employing adversarial learning and sparse feature constraints, texture details lost after noise suppression are reconstructed, improving image visual quality and the usability of features for subsequent recognition. Texture constraints and dynamic sample matching are introduced to address the texture distortion problem common in traditional GANs, enhancing the generalization ability of texture reconstruction across different scenarios.

[0030] The process of outputting target recognition results includes: A target semantic knowledge base is set up (including the attributes of targets such as equipment, buildings, and ships: size, scattering characteristics, typical shape, etc.). An image-semantic dual-stream network architecture is established, including image stream and semantic stream parallel networks. The image stream adopts a CNN-Transformer hybrid structure (CNN-Transformer hybrid model). First, the low-level visual features of the image are reconstructed by extracting texture through ResNet. Then, the Transformer encoder captures the global correlation between features (such as the spatial relationship between equipment and the surrounding environment) to generate image features. The semantic stream converts the attributes in the target semantic knowledge base into semantic embedding vectors. The semantic encoder aligns the vectors with the image features to obtain semantic features. The similarity matrix between image features and semantic features is calculated. Through attention weight allocation, the image regions that match the semantics are strengthened (such as the strong scattering region of the equipment) and the interference regions (such as background clutter) are weakened. Through attention weight allocation, cross-modal attention fusion of image features and semantic features is performed to obtain a fused feature map. At the same time, a dynamic weight update mechanism is introduced to adjust the contribution ratio of image stream and semantic stream in real time according to the recognition confidence. The Feature Pyramid Network (FPN) is used to perform multi-scale target detection on the fused feature map, which is suitable for targets of different sizes (such as small equipment and large ships). Duplicate detection boxes are removed by non-maximum suppression (NMS), and the target recognition results (including category, location and confidence) are output.

[0031] By fusing image visual features with target semantic features across modalities, the robustness of target recognition in complex scenes is improved, and the false positive rate for similar targets is reduced. Furthermore, by fusing visual and semantic features through cross-modal attention, the problem of false positives for similar targets is addressed, further enhancing the robustness of recognition in complex scenes.

[0032] The process of outputting change detection results includes: A parameter-sharing Siamese Transformer network is constructed to extract features from texture reconstruction images of various time phases (such as pre-disaster / post-disaster, different monitoring periods). The self-attention mechanism of the Transformer encoder is used to capture local texture and global structural features. The cross-attention mechanism is used to align the corresponding feature points of various time phases (such as fixed buildings and roads), reduce the interference of non-changing factors (such as imaging angle), and output feature maps of various time phases. The differences between feature maps of each time phase are calculated using a dual index of Euclidean distance and cosine similarity. Euclidean distance measures pixel-level feature differences (such as grayscale changes in flooded areas), while cosine similarity measures feature direction differences (such as texture structure changes caused by building collapse). The difference results are fused, and a preliminary mask of the changed region is obtained through adaptive threshold segmentation (the threshold is automatically determined based on the local features of the difference map to distinguish between changed and unchanged regions). The change region mask is post-processed using Conditional Random Field (CRF) (utilizing contextual information between pixels to optimize the continuity and accuracy of the boundary), and the change region boundary is optimized (e.g., eliminating isolated noise points). At the same time, the features of the change region are matched with a preset change type library (e.g., flooding, collapse, new construction), and the change type is determined by a classifier, outputting refined change detection results (including change region mask map and change type).

[0033] To address the feature differences in multi-temporal radar images, this method achieves high-precision localization and type determination of changing regions, meeting the needs of scenarios such as disaster monitoring and military reconnaissance. By employing a twin Transformer and dual-constraint difference calculation, the method improves the feature alignment accuracy of multi-temporal images, enabling precise localization and type determination of changing regions.

[0034] The method for outputting the processing report is as follows: Establish the correlation between image regions, target recognition, and change detection. For example, associate the equipment location in target recognition with the unchanged area in change detection to confirm target stability (e.g., a long-standing radar station has not changed). Associate newly built targets with changed areas to verify the accuracy of change types (e.g., the type of change in the area corresponding to the newly built building). Using an ensemble classifier, with target recognition confidence (e.g., the probability value of target category determination), change detection accuracy (e.g., pixel-level accuracy of changed areas), and image feature quality (e.g., the texture clarity of the corrected and enhanced image) as input features, perform secondary judgment on the target type in the target recognition results and the change type in the change detection, reducing the risk of misjudgment by a single module (e.g., when the target recognition confidence is low but the change detection accuracy is high, combine image texture features to reconfirm the correlation between the target and the change). Finally, integrate and enhance the corrected and enhanced image, target recognition results, and change detection results, and output a processing report.

[0035] By collaborating and integrating multi-dimensional information, structured reports that can be directly used for decision-making can be generated. Enhanced image annotation: Mark the target location (such as the coordinate frame of the equipment) and the changed areas (such as using different colors to mark newly built and collapsed areas) on the corrected and enhanced image; Target list generation: Organize information such as target categories, confidence levels, and coordinate locations to form a clear target list; Change analysis presents: statistical change type (new construction, collapse, movement, etc.), area of ​​change region, multi-temporal impact (such as the time difference between two consecutive radar images), and finally outputs a complete intelligent processing report including enhanced images, target list, and change analysis.

[0036] Example 2: Building upon Example 1, cross-scale correlation-enhanced channel-spatial collaborative attention (RC-CS-NAM) is embedded in both the high-resolution and low-resolution layers: First, a cross-scale noise correlation mapping between the high-resolution layer and the low-resolution layer is established. By calculating the gray-level correlation and spatial location correspondence of pixels between the two layers, the correspondence between small-scale noise clusters and large-scale correlated noise regions is marked (such as the small-scale speckle noise of a target echo in the high-resolution layer and its corresponding blurred noise region in the low-resolution layer). In the high-resolution layer, in addition to learning the channel sensitivity of small-scale noise concentration, channel attention also introduces the channel features of the low-resolution layer associated noise region as guiding weights to enhance the identification of cross-scale associated small noise; when locating local noise-dense areas, spatial attention combines the spatial range of the low-resolution layer associated noise region to avoid missing small noise points associated with large-scale noise. In the low-resolution layer, while channel attention focuses on the channel dominated by large-scale noise, it also refers to the channel response of small noise clusters associated with the high-resolution layer to optimize the distinction of large noise with blurred edges. When spatial attention locates a large noise area, it refines the boundary recognition of the large noise area by associating the spatial distribution of small noise clusters associated with the high-resolution layer, thereby achieving collaborative filtering of cross-scale associated noise. For the high-resolution and low-resolution layers after RC-CS-NAM processing, dynamic weighted bidirectional residual fusion is employed: Calculate two metrics in the two-layer feature: noise filtering integrity and texture detail retention (noise filtering integrity is quantified by the change in the pixel proportion of the noise region, and texture detail retention is quantified by the change in the edge gradient value). The residual fusion weights are dynamically allocated based on the weighted results of the two indicators: if the high-resolution layer has high texture detail retention but low noise filtering integrity, the fusion weights of the low-resolution layer features are increased to supplement the filtering effect; if the low-resolution layer has high noise filtering integrity but loses a lot of texture details, the fusion weights of the high-resolution layer features are increased to retain details. By using dynamic weighted residual connections, features from high-resolution and low-resolution layers are bidirectionally transferred and fused, ultimately outputting a denoised image.

[0037] By embedding RC-CS-NAM, the accuracy of noise filtering can be improved, while the loss rate of texture details can be reduced, providing a higher quality denoised image input for subsequent geometric correction feature point detection and texture reconstruction detail restoration.

[0038] Example 3: As Figure 3 As shown, the radar image intelligent enhancement and recognition system based on multi-model fusion is used to implement the method in Embodiment 1 or Embodiment 2, including: Denoising module: For the acquired raw radar image, multi-scale dynamic pyramid feature decomposition, channel-space collaborative attention and bidirectional residual fusion are used to adaptively suppress speckle noise and obtain the denoised image; Correction and enhancement module: Performs feature point detection on the denoised image, combines the detected feature points with radar trajectory parameters, performs robust transformation matrix estimation and adaptive contrast enhancement, and outputs a corrected and enhanced image; Texture Reconstruction Module: For the corrected and enhanced image, the texture details lost after noise suppression are reconstructed by using a generative adversarial network and multiple loss functions in a collaborative manner, thus obtaining a texture reconstructed image; Target recognition module: Establish an image-semantic dual-stream network architecture, acquire image features and semantic features of texture reconstructed images, perform cross-modal attention fusion to obtain a fused feature map, perform multi-scale target detection on the fused feature map, and output the target recognition result; Change detection module: Divides the texture reconstruction image into temporal phases, extracts features through Siamese Transformer, outputs feature maps of each phase and performs difference fusion, obtains the change region mask through adaptive threshold segmentation, determines the change type, and outputs the change detection results; Integration and Enhancement Module: Integrates and enhances the corrected and enhanced image, target recognition results, and change detection results, and outputs a processing report containing the enhanced image, target list, and change analysis.

Claims

1. A radar image intelligent enhancement and recognition method based on multi-model fusion, characterized in that... Includes the following steps: For the acquired raw radar images, multi-scale dynamic pyramid feature decomposition, channel-space collaborative attention and bidirectional residual fusion are used to adaptively suppress speckle noise and obtain denoised images. Feature point detection is performed on the denoised image. The detected feature points are combined with radar orbit parameters to perform robust transformation matrix estimation and adaptive contrast enhancement, and the corrected and enhanced image is output. For the corrected and enhanced image, the texture details lost after noise suppression are reconstructed by using generative adversarial networks and multi-loss function co-constraints to obtain the texture reconstructed image; A dual-stream image-semantic network architecture is established to acquire image features and semantic features of texture reconstructed images and perform cross-modal attention fusion to obtain a fused feature map. Multi-scale target detection is then performed on the fused feature map to output the target recognition result. The texture reconstruction image is divided into temporal phases, features are extracted using Siamese Transformer, feature maps of each phase are output and difference fusion is performed, and the change region mask is initially obtained through adaptive threshold segmentation to determine the change type and output the change detection result. The system integrates and enhances the corrected and enhanced image, target recognition results, and change detection results, and outputs a processing report that includes the enhanced image, target list, and change analysis.

2. The radar image intelligent enhancement and recognition method based on multi-model fusion according to claim 1, characterized in that, The process of obtaining the denoised image includes: The original radar image is a grayscale image, and the pixel value corresponds to the radar echo intensity. A multi-scale dynamic pyramid feature decomposition structure is constructed to decompose the input original radar image into a multi-resolution layer, resulting in a high-resolution layer that focuses on small-scale noise and a low-resolution layer that targets large-scale noise. In both high-resolution and low-resolution layers, channel-spatial collaborative attention (RC-SAM) is embedded. In high-resolution layers, channel attention prioritizes learning the sensitivity of channels with concentrated small-scale noise, while spatial attention accurately locates local areas with dense noise. In low-resolution layers, channel attention focuses on channels dominated by large-scale noise, while spatial attention locates large-scale noise regions, thus achieving targeted filtering of noise at different scales. For the high-resolution and low-resolution layers after RC-SAM processing, a bidirectional residual connection is used to transfer the texture features of the high-resolution layer to the low-resolution layer. Then, feature fusion is performed on the high-resolution and low-resolution layers to output the denoised image.

3. The intelligent radar image enhancement and recognition method based on multi-model fusion according to claim 1, characterized in that, The process of outputting the corrected and enhanced image is as follows: Feature point detection is performed on the denoised image, and stable feature points are screened using local entropy thresholding to eliminate pseudo feature points caused by noise interference. The detected feature points are combined with radar orbit parameters, the robust transformation matrix is ​​estimated using the RANSAC algorithm, the geometric distortion in the denoised image is fitted, and the geometric transformation is completed based on the robust transformation matrix to obtain a geometrically aligned image. For geometrically aligned images, the ACE algorithm is used to construct a dynamic window centered on each pixel, calculate the mean and variance of gray levels within the window, adaptively adjust the pixel gray levels, and output a corrected and enhanced image.

4. The intelligent radar image enhancement and recognition method based on multi-model fusion according to claim 3, characterized in that, The method of using local entropy thresholding to select stable feature points involves extracting the gray-level distribution of the local neighborhood for each candidate feature point in the denoised image, quantifying the texture complexity of the region, and representing it as follows: ； In the formula, H is the local entropy and L is the number of gray levels; It represents the percentage of pixels with grayscale value i within a local neighborhood. Set the local entropy threshold H C For candidate feature points, if H ≥ H C If H < H, then the candidate feature point is a stable feature point and should be retained. C If the candidate feature point is false, it should be removed.

5. The intelligent radar image enhancement and recognition method based on multi-model fusion according to claim 1, characterized in that, The process of obtaining a texture reconstruction image is as follows: The generator-discriminator architecture is set up. The generator adopts an improved U-Net structure, and sparse convolution is introduced in the encoder to focus on key texture features. The decoder gradually recovers high-resolution texture through transposed convolution and attention gating. The discriminator adopts the PatchGAN structure to judge the authenticity of the image on a pixel block basis, avoiding local texture distortion caused by global discrimination. The design incorporates a hybrid loss function that combines adversarial loss, texture-aware loss, and pixel-level loss. Based on the texture type of the input corrected and enhanced image, similar texture samples are dynamically selected from a high-resolution radar image sample library as supervisory data. The corrected and enhanced image is then input into the generator to generate an initial texture reconstruction image. The loss value is calculated using the hybrid loss function and backpropagated to the generator and discriminator. The model parameters are iteratively optimized until the loss converges, resulting in a Rad-GAN model with accurate texture reconstruction capabilities. The trained Rad-GAN model takes the corrected and enhanced image as input and outputs a textured reconstructed image.

6. The intelligent enhancement and recognition method for radar images based on multi-model fusion according to claim 5, characterized in that, The adversarial loss adopts least squares adversarial loss, the texture-aware loss adopts perceptual loss based on pre-trained CNN, and the pixel-level loss adopts L1 loss. Weights are set for the adversarial loss, texture-aware loss and pixel-level loss respectively, and weighted fusion is performed to obtain the final hybrid loss.

7. The intelligent radar image enhancement and recognition method based on multi-model fusion according to claim 1, characterized in that, The process of outputting target recognition results includes: A target semantic knowledge base is set up, and an image-semantic dual-stream network architecture is established, including parallel networks for image stream and semantic stream. The image stream adopts a CNN-Transformer hybrid structure. First, the low-level visual features of the image are reconstructed by extracting texture through ResNet, and then the global correlation between features is captured by the Transformer encoder to generate image features. The semantic stream converts the attributes in the target semantic knowledge base into semantic embedding vectors, and aligns them with the image features through the semantic encoder to obtain semantic features. The similarity matrix between image features and semantic features is calculated. Through attention weight allocation, cross-modal attention fusion of image features and semantic features is performed to obtain a fused feature map. At the same time, a dynamic weight update mechanism is introduced to adjust the contribution ratio of image stream and semantic stream in real time according to the recognition confidence. A feature pyramid network is used to perform multi-scale target detection on the fused feature map. Duplicate detection boxes are removed by non-maximum suppression, and the target recognition results are output.

8. The intelligent enhancement and recognition method for radar images based on multi-model fusion according to claim 1, characterized in that, The process of outputting change detection results includes: A parameter-sharing Siamese Transformer network is constructed to extract features from the texture reconstruction images at each time step. The self-attention mechanism of the Transformer encoder is used to capture local texture and global structural features. The corresponding feature points at each time step are aligned through the cross-attention mechanism, and the feature maps of each time step are output. The differences between the feature maps of each time phase are calculated using both Euclidean distance and cosine similarity. The difference results are then fused, and the change region mask is initially obtained through adaptive threshold segmentation. The change region mask is post-processed using a conditional random field to optimize the boundary of the change region. At the same time, the features of the change region are matched with a preset change type library, and the change type is determined by a classifier to output refined change detection results.

9. The intelligent enhancement and recognition method for radar images based on multi-model fusion according to claim 1, characterized in that, The method for outputting the processing report is as follows: Establish the correlation between image region, target recognition, and change detection. Utilize an ensemble classifier to perform secondary determination of target type in the target recognition results and change type in the change detection. Finally, integrate and enhance the corrected and enhanced image, target recognition results, and change detection results, and output a processing report. The processing report includes: Enhanced image with target location and changed areas marked; A list of targets, including target type, target recognition confidence level, and target coordinates; Results of change analysis including differences in change type, area of ​​change, and time of change.

10. A radar image intelligent enhancement and recognition system based on multi-model fusion, used to implement the method described in any one of claims 1-9, characterized in that, include: Denoising module: For the acquired raw radar image, multi-scale dynamic pyramid feature decomposition, channel-space collaborative attention and bidirectional residual fusion are used to adaptively suppress speckle noise and obtain the denoised image; Correction and enhancement module: Performs feature point detection on the denoised image, combines the detected feature points with radar trajectory parameters, performs robust transformation matrix estimation and adaptive contrast enhancement, and outputs a corrected and enhanced image; Texture Reconstruction Module: For the corrected and enhanced image, the texture details lost after noise suppression are reconstructed by using a generative adversarial network and multiple loss functions in a collaborative manner, thus obtaining a texture reconstructed image; Target recognition module: Establish an image-semantic dual-stream network architecture, acquire image features and semantic features of texture reconstructed images, perform cross-modal attention fusion to obtain a fused feature map, perform multi-scale target detection on the fused feature map, and output the target recognition result; Change detection module: Divides the texture reconstruction image into temporal phases, extracts features through Siamese Transformer, outputs feature maps of each phase and performs difference fusion, obtains the change region mask through adaptive threshold segmentation, determines the change type, and outputs the change detection results; Integration and Enhancement Module: Integrates and enhances the corrected and enhanced image, target recognition results, and change detection results, and outputs a processing report containing the enhanced image, target list, and change analysis.

Images