Method and system for detecting boundary enhancement based on visual base model feature adaptation

Abstract

The application belongs to the field of remote sensing change detection, and discloses a detection boundary enhancement method and system based on visual basic model feature adaptation. The application performs adaptive processing on the features extracted from the visual basic model, so that the features are converted from general semantic representation to remote sensing change detection task sensitive representation. While maintaining strong semantic discrimination ability, the application effectively suppresses non-change interference such as illumination difference, seasonal change, sensor imaging difference, etc., improves the stability of change area determination from the source, and reduces false changes and missed detection phenomena. The application introduces a double-time multi-scale feature pyramid structure, so that large-scale features are responsible for the overall consistency determination of the change area, and small-scale features focus on small structures and complex edges, and the change area determination is stable and fine in space level.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing change detection, specifically involving a detection boundary enhancement method and system based on visual basic model feature adaptation. Background Technology

[0002] Remote sensing change detection involves comparing and analyzing remote sensing images of the same area acquired at different times to automatically identify changed areas and their spatial distribution, and further determine the type of change when necessary. It is a key foundational technology in land resource monitoring, urban expansion assessment, disaster assessment, ecological environment supervision, and verification of illegal construction and project progress. With the widespread use of high-resolution remote sensing imagery in engineering applications, change detection results not only need high accuracy in determining whether a change has occurred, but also need to meet the needs of boundary-sensitive applications such as vector mapping, area statistics, and detailed inspections in terms of boundary location, contour integrity, and geometric consistency of changed areas. Therefore, in engineering practice, change detection is often considered a pixel-level segmentation problem with boundary quality as its core. However, remote sensing images generally exhibit differences in imaging conditions, geometric registration errors, large spans in ground feature scale, and complex background textures. Especially in scenarios with small targets, slender structures, and regular geometric boundaries, the boundaries of changed areas are easily disturbed, resulting in problems such as jagged edges, holes, breaks, or shifts.

[0003] Existing research has led to a relatively rich technical system for remote sensing change detection. From early traditional pixel-level change detection and shallow learning methods to object-oriented change detection methods, and more recently, the widely used deep learning end-to-end models and feature transfer and adaptation methods for visual foundational models, all revolve around the core process of bi-temporal difference modeling, change discrimination, and change map generation. Traditional methods typically rely on image difference, ratio, or change vector methods to construct difference maps, combined with threshold segmentation, statistical decision, or clustering to achieve change detection. While the process is clear and the implementation cost is low, it is highly sensitive to thresholds, feature construction, and scene conditions, making it difficult to maintain stable boundary quality in complex backgrounds and cross-condition applications. Object-oriented change detection suppresses pixel-level noise through prior segmentation, improving regional coherence to some extent, but its effectiveness depends on the segmentation scale and object matching strategy, and its ability to characterize thin structures and complex boundaries remains limited.

[0004] With the development of convolutional neural networks, end-to-end change detection models based on Siamese structures have gradually become mainstream. These models significantly improve the overall recognition accuracy of changed regions through multi-scale feature fusion and deep supervision strategies. However, due to the limited receptive field of convolutional operators and the loss of detail caused by multiple downsampling, problems such as blurred boundaries, broken elongated structures, or local adhesion still easily occur in high-resolution scenes. Transformer-based change detection methods enhance global relationship modeling capabilities through self-attention mechanisms, demonstrating outstanding performance in complex background noise suppression and large-scale structural consistency. However, they have high computational costs, are highly dependent on the scale of training data and annotation quality, and may still exhibit overly smooth boundaries and unstable details when lacking design for boundary constraints. Recent developments in visual foundation models have provided new technical approaches for change detection. By transferring general segmentation and boundary priors and combining them with lightweight adaptation modules, they are expected to improve segmentation generalization capabilities in complex scenes. However, in remote sensing change detection, specialized designs are still needed for modeling dual-temporal differences and fine-grained boundary preservation.

[0005] In conclusion, although existing remote sensing change detection technologies have made significant progress in the overall identification of changed areas, they still generally suffer from problems such as obvious degradation of changed area boundaries, difficulty in maintaining the integrity of thin structures, and susceptibility to false edges and false changes under complex texture conditions in high-resolution complex scenes. These issues make it difficult to simultaneously meet the dual requirements of engineering applications for accuracy in change determination and stability of boundary quality. Summary of the Invention

[0006] The purpose of this invention is to overcome the problem that it is difficult to simultaneously ensure the stability of the determination of changing regions and the fine integrity of the boundary quality in high-resolution complex scenes, and to provide a detection boundary enhancement method and system based on visual basic model feature adaptation.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a detection boundary enhancement method based on feature adaptation of a visual fundamental model, comprising the following steps: Acquire dual-temporal remote sensing images, preprocess the dual-temporal remote sensing images, construct a pixel-level change ground truth mask, match the pixel-level change ground truth mask with the dual-temporal remote sensing images, and obtain data samples and training labels. Feature extraction is performed on the data samples, and the extracted features are adapted to obtain a dual-temporal multi-scale feature pyramid. Directional frequency decomposition and adaptive weighted enhancement are performed on the dual-temporal multi-scale feature pyramid to obtain features after directional frequency attention enhancement; Channel selection and feature fusion are performed on the features after directional frequency attention enhancement to obtain a pyramid coupling layer; A spatial attention mechanism is used to perform spatial attention weighting and decoding boundary enhancement on the pyramid coupling layer to obtain the final spatial attention map. Based on the final spatial attention map, a total loss function is constructed, and a detection boundary enhancement model based on the adaptation of visual base model features is built according to the total loss function.

[0008] A further improvement of this invention lies in the following method for acquiring dual-temporal remote sensing images, preprocessing the dual-temporal remote sensing images, constructing a pixel-level change ground truth mask, and matching the pixel-level change ground truth mask with the dual-temporal remote sensing images to obtain data samples and training labels: Acquire dual-temporal remote sensing images of the same geographic area and register the dual-temporal remote sensing images; The registered dual-temporal remote sensing images are cropped to the required size to obtain several sub-blocks; Normalize all sub-blocks; Construct a pixel-level change ground truth mask and match the pixel-level change ground truth mask with the normalized sub-blocks; The matched sub-blocks are used as data samples, and the pixel-level change ground truth mask after matching is used as training labels.

[0009] A further improvement of this invention lies in the following method for extracting features from data samples and adapting the extracted features to obtain a dual-temporal multi-scale feature pyramid: Obtain the pre-trained FastSAM encoder; The data samples are processed by a pre-trained FastSAM encoder to extract multi-scale feature pyramids, resulting in four independent feature maps at different scales in each dual-temporal remote sensing image. An independent adaptation module is designed for each scale, and channel projection is performed using convolution to obtain the adapted dual-temporal multi-scale feature pyramid.

[0010] A further improvement of this invention lies in the following method for performing directional frequency decomposition and adaptive weighted enhancement on the dual-temporal multi-scale feature pyramid to obtain the features after directional frequency attention enhancement: Obtain the adapted dual-temporal multi-scale feature pyramid, analyze the adapted dual-temporal multi-scale feature pyramid, and obtain the feature map; Horizontal pooling is performed on the feature map to obtain low-frequency components. High-frequency components are obtained based on the feature map and low-frequency components. The low-frequency components and high-frequency components are then weighted and fused to obtain the horizontally processed features. The horizontally processed features are subjected to vertical frequency decomposition to obtain low-frequency components and high-frequency components. The low-frequency components and high-frequency components are then weighted and fused to obtain the frequency-processed features. The feature map and the frequency-processed features are fused with learnable scaling parameters to obtain bidirectional frequencies. An adaptive attention mechanism is used to process bidirectional frequencies and capture context at different scales. The contexts at different scales are added element by element, and then aggregated through convolution to obtain the features enhanced by directional frequency attention.

[0011] A further improvement of this invention lies in the following method for performing channel selection and feature fusion on the features after directional frequency attention enhancement to obtain the pyramid coupling layer: The features after directional frequency attention enhancement are obtained, and the channels of the features after directional frequency attention enhancement are divided into two groups. The first group is then processed by convolution stacking. Calculate the channel attention weights for the two sets of features separately; The two sets of features are fused with CSB through cross-group feature exchange to obtain two fused features. The two fused features are weighted by channel attention weights and then compressed through convolution to obtain compressed features; The features enhanced by directional frequency attention are sequentially passed through three cascaded CSB modules to obtain channel dependencies; Two FFB modules are used to progressively fuse the channel dependencies of different CSB stage features to obtain information at different levels of refinement. Based on information at different refinement levels, residual connection compression features are applied, and pyramid layers are applied to dual-temporal inputs respectively, followed by coupling through splicing to obtain a pyramid coupling layer.

[0012] A further improvement of this invention lies in employing a spatial attention mechanism to perform spatial attention weighting and decoding boundary enhancement on the pyramid coupling layer, resulting in the final spatial attention map. The specific method is as follows: Three hierarchical decoding blocks are used to progressively upsample the pyramid coupling layer to obtain dual-phase data; The two temporal phases are decoded separately using independent decoders, and then concatenated after convolutional compression. The splicing results are used to generate a spatial attention map through a bottleneck structure; The spatial attention maps are stacked and refined in depth, and then compressed by convolution; Spatial attention weights are modulated on the spatial attention map after convolution; The modulated spatial attention map is projected onto a single-channel variation probability map, and then bilinear interpolation is used to sample back to the original input resolution. The spatial attention map at the original input resolution is binarized and vectorized to obtain the final spatial attention map.

[0013] A further improvement of this invention lies in the following method for constructing a total loss function based on the final spatial attention map, and then constructing a detection boundary enhancement model based on visual base model feature adaptation according to the total loss function: Obtain the final spatial attention map and ground truth labels, and construct the binary cross-entropy loss function; Obtain intermediate predictions from both time phases and construct a potential similarity loss function; The total loss function is obtained by adding the binary cross-entropy loss function and the similarity loss function with equal weights. Based on the total loss function, the AdamW optimizer is used for model training and iteration to obtain a detection boundary enhancement model based on the adaptation of visual base model features.

[0014] Secondly, the present invention provides a detection boundary enhancement system based on feature adaptation of a visual fundamental model, comprising: The data preprocessing module is used to acquire dual-temporal remote sensing images, preprocess the dual-temporal remote sensing images, construct a pixel-level change ground truth mask, match the pixel-level change ground truth mask with the dual-temporal remote sensing images, and obtain data samples and training labels. The feature extraction module is used to extract features from data samples and adapt the extracted features to obtain a dual-temporal multi-scale feature pyramid. The feature enhancement module is used to perform directional frequency decomposition and adaptive weighted enhancement on the dual-temporal multi-scale feature pyramid to obtain features after directional frequency attention enhancement. The pyramid coupling module is used to perform channel selection and feature fusion on the features after directional frequency attention enhancement to obtain the pyramid coupling layer. The spatial attention weighting module is used to perform spatial attention weighting and decoding boundary enhancement on the pyramid coupling layer using a spatial attention mechanism to obtain the final spatial attention map. The model training module is used to construct the total loss function based on the final spatial attention map, and to construct a detection boundary enhancement model based on the visual base model feature adaptation according to the total loss function.

[0015] A further improvement of this invention is that the function of the data preprocessing module is implemented through the following method: Acquire dual-temporal remote sensing images of the same geographic area and register the dual-temporal remote sensing images; The registered dual-temporal remote sensing images are cropped to the required size to obtain several sub-blocks; Normalize all sub-blocks; Construct a pixel-level change ground truth mask and match the pixel-level change ground truth mask with the normalized sub-blocks; The matched sub-blocks are used as data samples, and the pixel-level change ground truth mask after matching is used as training labels.

[0016] A further improvement of this invention is that the function of the feature extraction module is implemented through the following method: Obtain the pre-trained FastSAM encoder; The data samples are processed by a pre-trained FastSAM encoder to extract multi-scale feature pyramids, resulting in four independent feature maps at different scales in each dual-temporal remote sensing image. An independent adaptation module is designed for each scale, and channel projection is performed using convolution to obtain the adapted dual-temporal multi-scale feature pyramid.

[0017] A further improvement of this invention is that the function of the feature enhancement module is implemented through the following method: Obtain the adapted dual-temporal multi-scale feature pyramid, analyze the adapted dual-temporal multi-scale feature pyramid, and obtain the feature map; Horizontal pooling is performed on the feature map to obtain low-frequency components. High-frequency components are obtained based on the feature map and low-frequency components. The low-frequency components and high-frequency components are then weighted and fused to obtain the horizontally processed features. The horizontally processed features are subjected to vertical frequency decomposition to obtain low-frequency components and high-frequency components. The low-frequency components and high-frequency components are then weighted and fused to obtain the frequency-processed features. The feature map and the frequency-processed features are fused with learnable scaling parameters to obtain bidirectional frequencies. An adaptive attention mechanism is used to process bidirectional frequencies and capture context at different scales. The contexts at different scales are added element by element, and then aggregated through convolution to obtain the features enhanced by directional frequency attention.

[0018] A further improvement of this invention is that the function of the pyramid coupling module is implemented through the following method: The features after directional frequency attention enhancement are obtained, and the channels of the features after directional frequency attention enhancement are divided into two groups. The first group is then processed by convolution stacking. Calculate the channel attention weights for the two sets of features separately; The two sets of features are fused with CSB through cross-group feature exchange to obtain two fused features. The two fused features are weighted by channel attention weights and then compressed through convolution to obtain compressed features; The features enhanced by directional frequency attention are sequentially passed through three cascaded CSB modules to obtain channel dependencies; Two FFB modules are used to progressively fuse the channel dependencies of different CSB stage features to obtain information at different levels of refinement. Based on information at different refinement levels, residual connection compression features are applied, and pyramid layers are applied to dual-temporal inputs respectively, followed by coupling through splicing to obtain a pyramid coupling layer.

[0019] A further improvement of this invention is that the function of the spatial attention weighting module is implemented through the following method: Three hierarchical decoding blocks are used to progressively upsample the pyramid coupling layer to obtain dual-phase data; The two temporal phases are decoded separately using independent decoders, and then concatenated after convolutional compression. The splicing results are used to generate a spatial attention map through a bottleneck structure; The spatial attention maps are stacked and refined in depth, and then compressed by convolution; Spatial attention weights are modulated on the spatial attention map after convolution; The modulated spatial attention map is projected onto a single-channel variation probability map, and then bilinear interpolation is used to sample back to the original input resolution. The spatial attention map at the original input resolution is binarized and vectorized to obtain the final spatial attention map.

[0020] A further improvement of this invention is that the function of the model training module is implemented through the following method: Obtain the final spatial attention map and ground truth labels, and construct the binary cross-entropy loss function; Obtain intermediate predictions from both time phases and construct a potential similarity loss function; The total loss function is obtained by adding the binary cross-entropy loss function and the similarity loss function with equal weights. Based on the total loss function, the AdamW optimizer is used for model training and iteration to obtain a detection boundary enhancement model based on the adaptation of visual base model features.

[0021] Thirdly, the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of a detection boundary enhancement method based on visual basic model feature adaptation.

[0022] Fourthly, the present invention provides a storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of a detection boundary enhancement method based on feature adaptation of a visual basic model.

[0023] Compared with the prior art, the present invention has the following beneficial effects: This invention adapts features extracted from a visual baseline model, transforming it from a general semantic representation to a sensitive representation for remote sensing change detection tasks. While maintaining strong semantic discriminative ability, it effectively suppresses non-change interference such as illumination differences, seasonal variations, and sensor imaging differences, fundamentally improving the stability of change region identification and reducing false changes and missed detections. This invention introduces a dual-temporal, multi-scale feature pyramid structure, where large-scale features are responsible for determining the overall consistency of change regions, while small-scale features focus on fine structures and complex edges, achieving a balance between stable change region identification and detailed depiction at the spatial level. Through directional frequency decomposition and adaptive weighted enhancement mechanisms, this invention specifically strengthens high-frequency edge responses and change contour features with directional consistency, making roads, building edges, and narrow targets stand out in complex backgrounds, thus significantly improving boundary blurring, breakage, and jaggedness issues. Furthermore, the channel selection and feature fusion process within the pyramid coupling layer effectively filters and complements cross-scale and cross-temporal information, avoiding redundant feature interference and giving higher weights to channels strongly correlated with change boundaries, helping to maintain boundary continuity and integrity. Finally, a decoding and boundary enhancement strategy based on spatial attention mechanisms enables the model to pay refined attention to the edge positions of changed regions during the output stage, achieving accurate localization and enhanced representation of boundary pixels. Furthermore, a joint loss function constrains both the changed region determination and boundary quality, allowing them to be optimized collaboratively during training. In summary, this invention effectively improves the robustness and consistency of changed region identification in high-resolution remote sensing change detection without relying on additional manual rules or complex post-processing. It also significantly enhances the clarity, continuity, and geometric integrity of changed boundaries, making it suitable for practical remote sensing applications with complex land cover types, large scale differences, and strong background interference. Attached Figure Description

[0024] Figure 1 This is a flowchart of the present invention; Figure 2 This is a system diagram of the present invention; Figure 3 This is a schematic diagram of the Adaptive Attention Block (AAB). Figure 4 This is a schematic diagram of the pyramid coupling layer (CSB / FFB). Figure 5 This is a flowchart illustrating Example 9; Figure 6 This is a schematic diagram of the system in Example 10. Detailed Implementation

[0025] To further understand the content of this invention, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments are merely illustrative and not limiting of the invention.

[0026] Example 1: See Figure 1 The detection boundary enhancement method based on visual base model feature adaptation includes the following steps: S1. Acquire dual-temporal remote sensing images, preprocess the dual-temporal remote sensing images, construct a pixel-level change ground truth mask, match the pixel-level change ground truth mask with the dual-temporal remote sensing images, and obtain data samples and training labels.

[0027] S2 extracts features from the data samples and adapts the extracted features to obtain a dual-temporal multi-scale feature pyramid.

[0028] S3 performs directional frequency decomposition and adaptive weighted enhancement on the dual-temporal multi-scale feature pyramid to obtain features after directional frequency attention enhancement.

[0029] S4 performs channel selection and feature fusion on the features after directional frequency attention enhancement to obtain the pyramid coupling layer.

[0030] S5 employs a spatial attention mechanism to perform spatial attention weighting and decoding boundary enhancement on the pyramid coupling layer, resulting in the final spatial attention map.

[0031] S6. Based on the final spatial attention map, construct the total loss function, and construct the detection boundary enhancement model based on the feature adaptation of the visual base model according to the total loss function.

[0032] Example 2: A detection boundary enhancement system based on visual base model feature adaptation includes: The data preprocessing module is used to acquire dual-temporal remote sensing images, preprocess the dual-temporal remote sensing images, construct a pixel-level change ground truth mask, match the pixel-level change ground truth mask with the dual-temporal remote sensing images, and obtain data samples and training labels.

[0033] The feature extraction module is used to extract features from the data samples and adapt the extracted features to obtain a dual-temporal multi-scale feature pyramid.

[0034] The feature enhancement module is used to perform directional frequency decomposition and adaptive weighted enhancement on the dual-temporal multi-scale feature pyramid to obtain features after directional frequency attention enhancement.

[0035] The pyramid coupling module is used to perform channel selection and feature fusion on the features after directional frequency attention enhancement to obtain the pyramid coupling layer.

[0036] The spatial attention weighting module is used to perform spatial attention weighting and decoding boundary enhancement on the pyramid coupling layer using a spatial attention mechanism to obtain the final spatial attention map.

[0037] The model training module is used to construct the total loss function based on the final spatial attention map, and to construct a detection boundary enhancement model based on the visual base model feature adaptation according to the total loss function.

[0038] Example 3: This embodiment transforms the original dual-temporal remote sensing images into normalized sample pairs. Through spatial registration, slicing, and normalization, it ensures a strict pixel-level correspondence between the two temporal images, providing stable and reliable data input for subsequent feature extraction. The specific methods for acquiring the dual-temporal remote sensing images, preprocessing them, constructing a pixel-level change ground truth mask, and matching the pixel-level change ground truth mask with the dual-temporal remote sensing images to obtain data samples and training labels are as follows: Step 1: Acquire dual-temporal remote sensing images of the same geographic area. (time )and (time The image is in RGB three-channel format, among which... , and These represent the height and width of the image, respectively.

[0039] Step two, for and Perform spatial registration to ensure pixel-level correspondence (skip if already registered); crop the large-format image into several sub-blocks at a fixed size (256×256 or 1024×1024 pixels) to meet the memory requirements for network training and inference.

[0040] Step 3: Normalize the sliced ​​images by scaling the pixel values ​​from 0 to 255 to 0 to 1 (divided by 255) to stabilize network training and improve the comparability between different images.

[0041] Step 4: Prepare a pixel-level ground truth mask. ,in Indicates the change in pixels, Represents unchanged pixels, and the corresponding One-to-one matching of sample slices.

[0042] Step 5: Output the data sample pairs required for training / inference. and training labels (During the training phase only), record information such as sample index and slice position for subsequent batch loading and result reassembly.

[0043] Example 4: This embodiment extracts a multi-scale feature pyramid using a frozen FastSAM encoder and employs a lightweight adaptation module to map pre-trained features to the change detection task space. The specific method for extracting features from data samples and adapting the extracted features to obtain the dual-temporal multi-scale feature pyramid is as follows: Step 1: Load the pre-trained FastSAM encoder weights (pre-trained on the SA-1B dataset, containing 11 million images and 1 billion masks), denoted as ... It remains frozen during training and does not participate in backpropagation updates.

[0044] Step 2: The preprocessed dual-temporal images are then processed separately. , The input is a FastSAM encoder, which independently extracts multi-scale feature pyramids. For each time phase... The encoder outputs four feature maps at different scales, as shown in the following formula: (1) in Indicates phase In the Features at each scale, scale index Corresponding to downsampling step size The original feature channel dimensions are respectively .

[0045] Step 3, for each scale Design an independent adapter module Channel projection is performed using 1×1 convolution, as shown in the following formula: (2) in This represents a 1×1 convolution weight matrix. This represents the convolution operation. This is for batch normalization. To correct the activation function of the linear unit. Channel compression configuration: stride 4 / 8 / 16 / 32 respectively from Compress to aisle.

[0046] Step 4: Output the adapted dual-temporal multi-scale feature pyramid and This information is then used by the subsequent directional frequency attention enhancement module. The adapted features retain the general visual representation capabilities of FastSAM pre-training and are adapted to the feature space of the remote sensing change detection task through a learnable projection layer.

[0047] Example 5: This embodiment uses Adaptive Attention Blocks (AAB) to perform directional frequency decomposition and adaptive weighting on the adapted features, suppressing high-frequency texture noise in remote sensing images while retaining useful directional high-frequency information such as building edges, thus solving the problem of texture-induced pseudo-changes. The specific method for performing directional frequency decomposition and adaptive weighting enhancement on the dual-temporal multi-scale feature pyramid to obtain the features after directional frequency attention enhancement is as follows: Step 1: Receive feature maps from the feature adaptation module. ,in For the number of channels, , For the height and width of the space.

[0048] Step 2, input features Horizontal pooling is performed to obtain low-frequency components. High-frequency components are obtained through subtraction. Then, learnable weights are applied for weighted fusion, as shown in the following formula: (3) in For learnable parameters, This represents element-wise multiplication. The high-frequency weights are incremented by 1 to ensure the residual connection form, which is beneficial for gradient flow.

[0049] Step 3: Process the horizontally processed features Perform frequency decomposition in the vertical direction and Then, a weighted fusion is performed, using the following formula: (4) in These are learnable parameters used to adaptively adjust the importance of low-frequency and high-frequency components in the vertical direction.

[0050] Step 4: Convert the original input features Features after frequency processing The formula for fusing through learnable scaling parameters is as follows: (5) in These are learnable parameters that control the contributions of the original features and frequency modulation features, respectively. They are set during initialization. , To ensure stability in the early stages of training.

[0051] Step 5, implement three AAB variants to capture context at different scales: (1) AAB_Global uses adaptive average pooling to capture global context; (2) AAB_Local_k7 uses local pooling with a kernel size of 7; (3) AAB_Local_k11 uses local pooling with a kernel size of 11.

[0052] Step 6: Add the outputs of the three AAB variants element-wise, and then aggregate them using a 1×1 convolution: .

[0053] Step 7, see Figure 3 The output features, enhanced by directional frequency attention, are used by subsequent pyramid coupling layers. Through directional frequency decomposition and adaptive weighting, the AAB module effectively suppresses texture noise while preserving building edge information.

[0054] Example 6: See Figure 4 This embodiment achieves effective coupling of cross-temporal features and hierarchical fusion of multi-scale information through a pyramid coupling layer composed of a channel selection block (CSB) and a feature fusion block (FFB), thus solving the problem of insufficient cross-temporal coupling. The specific method for performing channel selection and feature fusion on the features after directional frequency attention enhancement to obtain the pyramid coupling layer is as follows: Step 1: Receive the feature map output from the AAB module. Divide the channel into two groups The first group is processed by stacking three 3×3 convolution layers: .

[0055] Step 2: Channel Attention Mechanism. Calculate the channel attention weights for each of the two sets of features, using the following formula: (6) in For global average pooling (which reduces the spatial dimension) Compress to , Compress the channel by 8 times (dimensionality reduction). Restore channel dimension (upgrade dimension). This is the Sigmoid activation function.

[0056] Step 3: Perform cross-group feature exchange and Then, through learnable projection fusion: .

[0057] Step four, for the two features to be fused Application Channel Attention Weighting and Then concatenate and compress using a 1×1 convolution: .

[0058] Step 5: Process the input features Three CSB modules are cascaded sequentially: , , We will gradually refine the interaction patterns between channels and capture increasingly complex channel dependencies.

[0059] Step 6: Progressively fuse features from different CSB stages using two FFB modules: Merge the third and second levels The fusion result is combined with the first level, aggregating information from different levels of refinement.

[0060] Step 7, Apply residual join Maintain gradient flow. For dual-temporal inputs, apply pyramid layers separately and then couple them using a stitching method: ,in and The two temporal phases are respectively output after passing through the pyramid layer, and a dual-temporal feature correlation is established.

[0061] Example 7: This embodiment uses a Spatial Attention Block (SAB) to generate a spatial attention map that highlights regions of significant change. It then combines a progressive upsampling decoder and a Residual Refinement Module (ResCD) to achieve clear boundary mask reconstruction, thus addressing the boundary erosion and fracture problem of slender structures. Specifically, the method for using a spatial attention mechanism to assign spatial attention weights and enhance the decoding boundary of the pyramid coupling layer to obtain the final spatial attention map is as follows: Step 1 involves progressive upsampling using three hierarchical decoding blocks (Dec2, Dec1, Dec0), with a channel configuration of 160→80→40→64. Each decoding block includes transposed convolution (spatial upsampling), skip connections (fusion of encoder features), convolutional thinning, and AAB-Pyramid attention enhancement, with resolution gradually restored from stride 16→8→4→2.

[0062] Step two: For the dual-phase input, obtain the values ​​using independent decoders. (in (Batch size), each branch generates 8-channel intermediate predictions through a 1×1 convolutional segmentation head. Then splice them together For use by SAB.

[0063] Step 3: Receive the spliced ​​dual-temporal prediction A spatial attention map is generated using the bottleneck structure, as shown in the following formula: (7) in, Compress 16 channels to 4 channels (compression factor) (After approval and normalization) (Momentum = 0.95) and After activation, Restored to 16 channels. Using the Sigmoid activation function, the attention weights are constrained to... scope.

[0064] Step 4: Concatenate the outputs of the dual-branch decoder. Deep refinement is performed using a stack of 6 ResBlocks (each ResBlock contains two 3×3 convolutions and residual connections), and then projected from 128 channels to 16 channels using 1×1 convolutions. .

[0065] Step 5: Apply spatial attention weights to the projected features for modulation, as shown in the following formula: (8) in This represents element-wise multiplication. This attention-weighted modulation ensures that the network focuses on regions of significant change while suppressing irrelevant changes and background noise, significantly improving the accuracy and boundary quality of change detection.

[0066] Step 6: Project the 16-channel features onto the single-channel variation probability map using a final 1×1 convolution. (in logits form), then upsampled to the original input resolution using bilinear interpolation. .

[0067] Step 7: During the training phase, output logits for loss calculation. During the inference phase, apply a sigmoid function to the logits to obtain the probability. Then, thresholding (threshold=0.5) is performed to generate a binary transformation map. (Where 1 represents change and 0 represents no change). Optionally, morphological processing (opening / closing operations to remove noise and fill holes) and contour extraction are performed on the binary image, and it is vectorized into a polygon / polyline format to form a GIS boundary product.

[0068] Example 8: This embodiment combines pixel-level supervision and feature-level consistency constraints for change detection using a composite loss function, ensuring that the model can accurately identify changing pixels while also learning temporally consistent discriminative feature representations. Specifically, the method for constructing a total loss function based on the final spatial attention map, and then building a detection boundary enhancement model based on visual base model feature adaptation using this total loss function, is as follows: Step 1: Use binary cross-entropy with logits as the primary supervisory loss. Given the predicted change graph... (logits form without Sigmoid) and truth labels (Where 0 represents no change and 1 represents change), the loss is defined as follows: (9) in, For the Sigmoid function, map logits to probability space, Summing over all pixel positions. For changing pixels ( The loss is This encourages the model to output high probabilities; for unchanged pixels ( The loss is This encourages the model to output low probabilities. This loss directly optimizes the ability to detect pixel-level changes.

[0069] Step 2: Receive the intermediate prediction of the two temporal phases. (8-channel feature logits), applied temperature scaling and Softmax normalization (Temperature parameters) For each spatial location Calculate the two-phase eigenvectors cosine similarity According to the truth label Calculate pixel-level loss: If (Unchanged) Encourage feature alignment; if (Change) then Allows for feature divergence. Batch average yield ,in This represents the total number of pixels.

[0070] Step 3: Add the binary cross-entropy loss and the potential similarity loss with equal weights, as shown in the following formula: (10) This composite loss function provides both direct pixel-level supervision ( ), and through feature-level constraints ( Regularized model learning ensures that the model not only accurately identifies changes, but also learns a robust internal representation that is consistent with time.

[0071] Step 4: Train the model using the AdamW optimizer (a weight decay variant of Adam), with the following parameters configured: initial learning rate. Momentum parameters , The weight decays to To prevent overfitting, a polynomial decay learning rate schedule is used. ,in The initial learning rate, The maximum number of training iterations is 0.9, with a decay exponent of 0.9.

[0072] Step 5: Set the batch size to 16 and train for 50,000 iterations on an NVIDIA RTX 4090 GPU (24GB VRAM). The input slices are 256×256 pixels. Each iteration includes: (1) Forward propagation, with input bi-temporal image pairs. truth value Calculate and predict and intermediate prediction (2) Loss calculation, calculation and Summation yields (3) Backpropagation, calculate the gradient of the loss with respect to the model parameters; (4) Parameter update, update the model parameters according to the gradient through the AdamW optimizer (note that the FastSAM encoder parameters are kept frozen, and only the parameters of the adapter module, AAB, CSB, FFB, SAB, decoder and ResCD are updated).

[0073] Step 6: Periodically evaluate model performance on the validation set (e.g., every 1000 iterations), calculate metrics such as average F1 score (mF1), average intersection-union ratio (mIoU), precision, recall, and overall accuracy (OA), and save the weights of the best-performing model.

[0074] Step 7: Load the optimal model weights, input the two-temporal image pairs to be detected, perform forward propagation, and output a binary change map or change probability map for practical application scenarios (such as urban monitoring, disaster assessment, land use change analysis, etc.).

[0075] Example 9: See Figure 5This embodiment addresses issues such as "texture-induced pseudo-changes, insufficient cross-temporal coupling, and erosion and fracture of slender structural boundaries" in high-resolution urban building change detection. It achieves fine extraction and boundary enhancement of the contours of changed areas through a combination of techniques including "frozen visual basic model encoder + lightweight feature adaptation + directional frequency attention enhancement + pyramid coupling fusion + spatial attention boundary sharpening decoding".

[0076] Step 1: Input registered dual-temporal remote sensing images of the same area. , Data is processed by slicing and normalizing; during the training phase, ground truth masks corresponding to the changes are prepared. .

[0077] Step two, respectively , Input the frozen FastSAM (Fast Segment Anything Model) encoder to extract multi-scale feature pyramids (downsampling strides of 4, 8, 16, and 32); then perform channel alignment and compression on the features at each scale (1×1 convolution + batch normalization (BN) + ReLU activation) to map the high-dimensional channels of the base model to the number of channels required for the change detection task.

[0078] Step 3: Adaptive Attention Block (AAB) is used to decompose and weight the features in the horizontal and vertical directions into low-frequency and high-frequency components, thereby suppressing texture noise and preserving the edges of the building orientation.

[0079] Step four involves forming a hierarchical pyramid coupling layer using a Channel Selection Block (CSB) and a Feature Fusion Block (FFB) to achieve cross-temporal and multi-scale fusion.

[0080] Step 5: Spatial attention weights are generated by the Spatial Attention Block (SAB) and assigned to regions with significant changes; combined with the multi-level upsampling decoding and residual refinement module ResCD (Residual Change Detection), a change mask with clear boundaries is output.

[0081] Step 6: End-to-end training is performed using a composite loss function of binary cross-entropy loss and potential similarity loss, with the AdamW optimizer and a multinomial learning rate decay strategy.

[0082] This invention has been validated through comparative experiments on publicly available remote sensing change detection benchmark datasets. The results show that this method achieves optimal or tied-optimal performance in both evaluation groups: in the first evaluation group, Pre=95.92, Rec=95.44, OA=99.17, mF1=95.66, and mIoU=91.68; in the second evaluation group, Pre=97.97, Rec=97.37, OA=99.69, mF1=97.66, and mIoU=95.49. Compared with methods such as FC-Siam, SNUNet, BIT, ChangeFormer, CTD-Former, CGNet, EADTeer, SAM-CD, and CFNet, this method demonstrates greater stability in overall accuracy and comprehensive indices (mF1, mIoU), proving the feasibility of this scheme and its good engineering application effects.

[0083] Table 1 Comparison of Levir-CD datasets

[0084] Table 2 Comparison of WHU-CD datasets

[0085] Table 1 shows the quantitative comparison results (%) of this method on the LEVIR-CD dataset, and Table 2 shows the quantitative comparison results (%) of this method on the WHU-CD dataset.

[0086] This invention provides clearer boundary changes and more complete contours. Compared to change detection methods that rely solely on conventional feature fusion, this invention introduces an adaptive attention mechanism (AAB) in the feature enhancement stage. This mechanism decomposes and adaptively weights low-frequency / high-frequency components in the horizontal and vertical directions, suppressing false edge interference caused by textures while enhancing the directional edge response of targets such as buildings. Combined with spatial attention weighting and residual refinement boundary sharpening decoding (SAB + ResCD), the output change mask has sharper edges, more continuous elongated structures, and higher contour fit, thereby improving the mapping and usability of the change results. This invention produces fewer false changes and is more robust (especially for areas with complex textures). This invention suppresses high-frequency texture noise through AAB and explicitly weights areas with significant changes through SAB in the backend, reducing the misleading effect of high-frequency areas such as roof textures and road textures on change detection. This reduces erroneous outputs such as "halo-like false changes" and "edge noise fragments," making change detection results more stable and reliable in complex urban scenes.

[0087] In this invention, the frozen visual base model encoder can be replaced with SAM, MobileSAM, or other visual base models with general representation capabilities. The feature adaptation module (1×1 convolution + normalization + activation) can also be replaced with depthwise separable convolution, linear projection, or different normalization methods without changing the overall process. The directional frequency attention block (AAB), pyramid coupling layer (CSB / FFB), and spatial attention boundary sharpening decoding (SAB+ResCD) can be reduced, the number of cascaded layers adjusted, or replaced with similar attention / fusion / residual refinement structures according to computational and accuracy requirements to achieve equivalent boundary enhancement effects. In addition to building change detection, this method can also be used for extracting changes in land features such as roads, water bodies, and farmland, as well as for applications requiring high-quality change boundary output, such as post-disaster damage assessment and urban renewal monitoring.

[0088] Example 10: See Figure 6 The present invention also provides an electronic device 100 based on a detection boundary enhancement method adapted to visual basic model features; the electronic device 100 includes a memory 101, at least one processor 102, a computer program 103 stored in the memory 101 and executable on the at least one processor 102, and at least one communication bus 104.

[0089] The memory 101 can be used to store the computer program 103. The processor 102 implements the steps of the detection boundary enhancement method based on visual basic model feature adaptation described in Embodiment 1 by running or executing the computer program stored in the memory 101 and calling the data stored in the memory 101. The memory 101 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device 100 (such as audio data), etc. In addition, the memory 101 may include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other non-volatile solid-state storage device.

[0090] The at least one processor 102 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor 102 may be a microprocessor or any conventional processor. The processor 102 is the control center of the electronic device 100, connecting various parts of the electronic device 100 via various interfaces and lines.

[0091] The memory 101 in the electronic device 100 stores multiple instructions to implement a detection boundary enhancement method based on visual base model feature adaptation, and the processor 102 can execute the multiple instructions to achieve the following: Acquire dual-temporal remote sensing images, preprocess the dual-temporal remote sensing images, construct a pixel-level change ground truth mask, match the pixel-level change ground truth mask with the dual-temporal remote sensing images, and obtain data samples and training labels. Feature extraction is performed on the data samples, and the extracted features are adapted to obtain a dual-temporal multi-scale feature pyramid. Directional frequency decomposition and adaptive weighted enhancement are performed on the dual-temporal multi-scale feature pyramid to obtain features after directional frequency attention enhancement; Channel selection and feature fusion are performed on the features after directional frequency attention enhancement to obtain a pyramid coupling layer; A spatial attention mechanism is used to perform spatial attention weighting and decoding boundary enhancement on the pyramid coupling layer to obtain the final spatial attention map. Based on the final spatial attention map, a total loss function is constructed, and a detection boundary enhancement model based on the adaptation of visual base model features is built according to the total loss function.

[0092] Example 11: If the modules / units integrated in the electronic device 100 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, and a read-only memory (ROM).

[0093] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0094] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0095] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0096] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A detection boundary enhancement method based on visual basic model feature adaptation, characterized in that, Includes the following steps: Acquire dual-temporal remote sensing images, preprocess the dual-temporal remote sensing images, construct a pixel-level change ground truth mask, match the pixel-level change ground truth mask with the dual-temporal remote sensing images, and obtain data samples and training labels. Feature extraction is performed on the data samples, and the extracted features are adapted to obtain a dual-temporal multi-scale feature pyramid. Directional frequency decomposition and adaptive weighted enhancement are performed on the dual-temporal multi-scale feature pyramid to obtain features after directional frequency attention enhancement; Channel selection and feature fusion are performed on the features after directional frequency attention enhancement to obtain a pyramid coupling layer; A spatial attention mechanism is used to perform spatial attention weighting and decoding boundary enhancement on the pyramid coupling layer to obtain the final spatial attention map. Based on the final spatial attention map, a total loss function is constructed, and a detection boundary enhancement model based on the adaptation of visual base model features is built according to the total loss function.

2. The detection boundary enhancement method based on visual basic model feature adaptation according to claim 1, characterized in that, The specific method for acquiring dual-temporal remote sensing images, preprocessing the images, constructing a pixel-level change ground truth mask, and matching the mask with the images to obtain data samples and training labels is as follows: Acquire dual-temporal remote sensing images of the same geographic area and register the dual-temporal remote sensing images; The registered dual-temporal remote sensing images are cropped to the required size to obtain several sub-blocks; Normalize all sub-blocks; Construct a pixel-level change ground truth mask and match the pixel-level change ground truth mask with the normalized sub-blocks; The matched sub-blocks are used as data samples, and the pixel-level change ground truth mask after matching is used as training labels.

3. The detection boundary enhancement method based on visual basic model feature adaptation according to claim 1, characterized in that, The specific method for extracting features from data samples and adapting the extracted features to obtain a dual-temporal multi-scale feature pyramid is as follows: Obtain the pre-trained FastSAM encoder; The data samples are processed by a pre-trained FastSAM encoder to extract multi-scale feature pyramids, resulting in four independent feature maps at different scales in each dual-temporal remote sensing image. An independent adaptation module is designed for each scale, and channel projection is performed using convolution to obtain the adapted dual-temporal multi-scale feature pyramid.

4. The detection boundary enhancement method based on visual basic model feature adaptation according to claim 1, characterized in that, The specific method for performing directional frequency decomposition and adaptive weighted enhancement on a dual-temporal multi-scale feature pyramid to obtain the features after directional frequency attention enhancement is as follows: Obtain the adapted dual-temporal multi-scale feature pyramid, analyze the adapted dual-temporal multi-scale feature pyramid, and obtain the feature map; Horizontal pooling is performed on the feature map to obtain low-frequency components. High-frequency components are obtained based on the feature map and low-frequency components. The low-frequency components and high-frequency components are then weighted and fused to obtain the horizontally processed features. The horizontally processed features are subjected to vertical frequency decomposition to obtain low-frequency components and high-frequency components. The low-frequency components and high-frequency components are then weighted and fused to obtain the frequency-processed features. The feature map and the frequency-processed features are fused with learnable scaling parameters to obtain bidirectional frequencies. An adaptive attention mechanism is used to process bidirectional frequencies and capture context at different scales. The contexts at different scales are added element by element, and then aggregated through convolution to obtain the features enhanced by directional frequency attention.

5. The detection boundary enhancement method based on visual basic model feature adaptation according to claim 1, characterized in that, The specific method for performing channel selection and feature fusion on the features after directional frequency attention enhancement to obtain the pyramid coupling layer is as follows: The features after directional frequency attention enhancement are obtained, and the channels of the features after directional frequency attention enhancement are divided into two groups. The first group is then processed by convolution stacking. Calculate the channel attention weights for the two sets of features separately; The two sets of features are fused with CSB through cross-group feature exchange to obtain two fused features. The two fused features are weighted by channel attention weights and then compressed through convolution to obtain compressed features; The features enhanced by directional frequency attention are sequentially passed through three cascaded CSB modules to obtain channel dependencies; Two FFB modules are used to progressively fuse the channel dependencies of different CSB stage features to obtain information at different levels of refinement. Based on information at different refinement levels, residual connection compression features are applied, and pyramid layers are applied to dual-temporal inputs respectively, followed by coupling through splicing to obtain a pyramid coupling layer.

6. The detection boundary enhancement method based on visual basic model feature adaptation according to claim 1, characterized in that, The specific method for applying spatial attention weighting and decoding boundary enhancement to the pyramid coupling layer using a spatial attention mechanism to obtain the final spatial attention map is as follows: Three hierarchical decoding blocks are used to progressively upsample the pyramid coupling layer to obtain dual-phase data; The two temporal phases are decoded separately using independent decoders, and then concatenated after convolutional compression. The splicing results are used to generate a spatial attention map through a bottleneck structure; The spatial attention maps are stacked and refined in depth, and then compressed by convolution; Spatial attention weights are modulated on the spatial attention map after convolution; The modulated spatial attention map is projected onto a single-channel variation probability map, and then bilinear interpolation is used to sample back to the original input resolution. The spatial attention map at the original input resolution is binarized and vectorized to obtain the final spatial attention map.

7. The detection boundary enhancement method based on visual basic model feature adaptation according to claim 1, characterized in that, Based on the final spatial attention map, a total loss function is constructed. The specific method for constructing a detection boundary enhancement model based on visual base model feature adaptation according to the total loss function is as follows: Obtain the final spatial attention map and ground truth labels, and construct the binary cross-entropy loss function; Obtain intermediate predictions from both time phases and construct a potential similarity loss function; The total loss function is obtained by adding the binary cross-entropy loss function and the similarity loss function with equal weights. Based on the total loss function, the AdamW optimizer is used for model training and iteration to obtain a detection boundary enhancement model based on the adaptation of visual base model features.

8. A detection boundary enhancement system based on visual basic model feature adaptation, characterized in that, include: The data preprocessing module is used to acquire dual-temporal remote sensing images, preprocess the dual-temporal remote sensing images, construct a pixel-level change ground truth mask, match the pixel-level change ground truth mask with the dual-temporal remote sensing images, and obtain data samples and training labels. The feature extraction module is used to extract features from data samples and adapt the extracted features to obtain a dual-temporal multi-scale feature pyramid. The feature enhancement module is used to perform directional frequency decomposition and adaptive weighted enhancement on the dual-temporal multi-scale feature pyramid to obtain features after directional frequency attention enhancement. The pyramid coupling module is used to perform channel selection and feature fusion on the features after directional frequency attention enhancement to obtain the pyramid coupling layer. The spatial attention weighting module is used to perform spatial attention weighting and decoding boundary enhancement on the pyramid coupling layer using a spatial attention mechanism to obtain the final spatial attention map. The model training module is used to construct the total loss function based on the final spatial attention map, and to construct a detection boundary enhancement model based on the visual base model feature adaptation according to the total loss function.

9. The detection boundary enhancement system based on visual fundamental model feature adaptation according to claim 8, characterized in that, The data preprocessing module's functionality is implemented through the following methods: Acquire dual-temporal remote sensing images of the same geographic area and register the dual-temporal remote sensing images; The registered dual-temporal remote sensing images are cropped to the required size to obtain several sub-blocks; Normalize all sub-blocks; Construct a pixel-level change ground truth mask and match the pixel-level change ground truth mask with the normalized sub-blocks; The matched sub-blocks are used as data samples, and the pixel-level change ground truth mask after matching is used as training labels.

10. The detection boundary enhancement system based on visual fundamental model feature adaptation according to claim 8, characterized in that, The feature extraction module's functionality is implemented using the following methods: Obtain the pre-trained FastSAM encoder; The data samples are processed by a pre-trained FastSAM encoder to extract multi-scale feature pyramids, resulting in four independent feature maps at different scales in each dual-temporal remote sensing image. An independent adaptation module is designed for each scale, and channel projection is performed using convolution to obtain the adapted dual-temporal multi-scale feature pyramid.

11. The detection boundary enhancement system based on visual fundamental model feature adaptation according to claim 8, characterized in that, The feature enhancement module's functionality is implemented through the following methods: Obtain the adapted dual-temporal multi-scale feature pyramid, analyze the adapted dual-temporal multi-scale feature pyramid, and obtain the feature map; Horizontal pooling is performed on the feature map to obtain low-frequency components. High-frequency components are obtained based on the feature map and low-frequency components. The low-frequency components and high-frequency components are then weighted and fused to obtain the horizontally processed features. The horizontally processed features are subjected to vertical frequency decomposition to obtain low-frequency components and high-frequency components. The low-frequency components and high-frequency components are then weighted and fused to obtain the frequency-processed features. The feature map and the frequency-processed features are fused with learnable scaling parameters to obtain bidirectional frequencies. An adaptive attention mechanism is used to process bidirectional frequencies and capture context at different scales. The contexts at different scales are added element by element, and then aggregated through convolution to obtain the features enhanced by directional frequency attention.

12. The detection boundary enhancement system based on visual fundamental model feature adaptation according to claim 8, characterized in that, The functionality of the pyramid coupling module is implemented through the following methods: The features after directional frequency attention enhancement are obtained, and the channels of the features after directional frequency attention enhancement are divided into two groups. The first group is then processed by convolution stacking. Calculate the channel attention weights for the two sets of features separately; The two sets of features are fused with CSB through cross-group feature exchange to obtain two fused features. The two fused features are weighted by channel attention weights and then compressed through convolution to obtain compressed features; The features enhanced by directional frequency attention are sequentially passed through three cascaded CSB modules to obtain channel dependencies; Two FFB modules are used to progressively fuse the channel dependencies of different CSB stage features to obtain information at different levels of refinement. Based on information at different refinement levels, residual connection compression features are applied, and pyramid layers are applied to dual-temporal inputs respectively, followed by coupling through splicing to obtain a pyramid coupling layer.

13. The detection boundary enhancement system based on visual fundamental model feature adaptation according to claim 8, characterized in that, The functionality of the spatial attention weighting module is implemented through the following methods: Three hierarchical decoding blocks are used to progressively upsample the pyramid coupling layer to obtain dual-phase data; The two temporal phases are decoded separately using independent decoders, and then concatenated after convolutional compression. The splicing results are used to generate a spatial attention map through a bottleneck structure; The spatial attention maps are stacked and refined in depth, and then compressed by convolution; Spatial attention weights are modulated on the spatial attention map after convolution; The modulated spatial attention map is projected onto a single-channel variation probability map, and then bilinear interpolation is used to sample back to the original input resolution. The spatial attention map at the original input resolution is binarized and vectorized to obtain the final spatial attention map.

14. The detection boundary enhancement system based on visual fundamental model feature adaptation according to claim 8, characterized in that, The functionality of the model training module is implemented through the following methods: Obtain the final spatial attention map and ground truth labels, and construct the binary cross-entropy loss function; Obtain intermediate predictions from both time phases and construct a potential similarity loss function; The total loss function is obtained by adding the binary cross-entropy loss function and the similarity loss function with equal weights. Based on the total loss function, the AdamW optimizer is used for model training and iteration to obtain a detection boundary enhancement model based on the adaptation of visual base model features.

15. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the detection boundary enhancement method based on visual basic model feature adaptation as described in any one of claims 1 to 7.

16. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the detection boundary enhancement method based on visual basic model feature adaptation as described in any one of claims 1 to 7.

Images