Unmanned aerial vehicle remote sensing image change detection method and system based on hierarchical decision and dynamic local registration
By employing a hierarchical decision-making and dynamic local registration approach, the problems caused by ground texture and parallax in UAV ultra-high resolution imagery are solved, enabling high-precision, low-cost, and rapid change detection, which is suitable for emergency response at the edge of UAVs.
Patent Information
- Application Number
- CN202511784150.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-01
- Publication Date
- 2026-03-03
AI Technical Summary
Existing change detection methods cannot simultaneously address semantic false alarms caused by high-frequency ground textures and local geometric false detections caused by parallax when processing UAV ultra-high resolution imagery, and deep learning methods lack zero-shot generalization capability.
A hierarchical decision-making and dynamic local registration approach is adopted, including data preprocessing, global guidance and local execution modules. Through global registration, radiometric correction, physical attention mechanism and dynamic local reregistration, ground object interference and parallax error are suppressed. Combined with traditional features and robust statistical methods, high-precision change detection is achieved.
It significantly reduces false alarm rate, improves detection accuracy and computational efficiency, has strong scene adaptability and low-cost deployment capability, and is suitable for rapid emergency response at the edge of UAVs.
Smart Images
Figure CN121600316A_ABST
Abstract
Description
Technical fields:
[0001] This invention relates to the field of remote sensing image processing technology, and in particular to a method and system for detecting changes in ultra-high resolution images from unmanned aerial vehicles (UAVs). Background technology:
[0002] Change detection is one of the core tasks of remote sensing image analysis, with significant application value in fields such as land resource monitoring, illegal building identification, disaster emergency assessment, and refined urban management. In recent years, with the rapid development of UAV technology, acquiring ultra-high-resolution imagery at the centimeter and even millimeter levels has become commonplace. High-resolution imagery provides unprecedented detail of ground features, bringing new opportunities for refined change detection, but also posing serious challenges to the accuracy and robustness of traditional change detection algorithms.
[0003] Currently, existing change detection methods are mainly divided into three categories, but they all have significant drawbacks when processing UAV ultra-high resolution imagery: The first category: methods based on pixel algebra (such as pixel difference). These methods are simple, but they have extremely stringent requirements for image registration accuracy and are highly sensitive to changes in illumination. When UAVs acquire images at different times, differences in their flight attitude, altitude, and illumination conditions are inevitable. These differences generate a large number of false changes in areas where no real changes have occurred (such as roads and rooftops), leading to a very high false alarm rate. The second category: methods based on traditional features (such as SIFT, SURF, etc.). These methods are somewhat robust to illumination, but when applied to ultra-high resolution imagery, they expose two fundamental flaws: 1) The imagery captures excessive details of ground texture. For example, seasonal vegetation changes and water ripples, details that can be ignored in traditional low-to-medium resolution imagery, are identified as a massive number of change feature points in ultra-high resolution, resulting in dense, meaningless false alarms that severely obscure the true change signals; 2) These methods typically rely on global homography transformations for registration. However, the severe parallax problem caused by low-altitude drone flight makes it impossible for a single global matrix to perfectly align all ground features. For example, the displacement of the top and bottom of a tall building is inconsistent across different images. This local registration error inevitably leads to a large number of geometric false detections at the contour edges of all elevation objects (such as buildings and bridges). The third category consists of deep learning-based methods (such as Siamese Network and U-Net). These methods perform well on specific datasets, but their performance heavily relies on massive amounts of pixel-level precisely labeled training samples. For drone imagery acquired from different regions, seasons, and sensors, it is often necessary to re-collect and label data and retrain the model. This results in poor generalization ability and high labeling and computational costs, making it unsuitable for the urgent need for rapid, zero-sample change detection of newly emerging scenes.
[0004] In summary, existing change detection technologies generally fail to simultaneously address semantic false alarms caused by high-frequency ground textures (vegetation, water bodies) and local geometric false detections caused by parallax when dealing with UAV ultra-high-resolution imagery. Furthermore, deep learning methods lack zero-shot generalization capability. Therefore, there is an urgent need for a novel change detection method specifically designed for UAV ultra-high-resolution imagery that requires no training, actively suppresses ground interference, and overcomes local registration errors. Summary of the Invention:
[0005] This invention aims to provide a method and system for detecting changes in remote sensing images based on hierarchical decision-making and dynamic local registration, to solve the problems of high false alarm and high false detection rates caused by the inability of existing technologies to handle parallax and ground object interference. The system includes: a data preprocessing module, a global guidance module, a local execution module, and a result fusion and output module. The overall flowchart is shown below. Figure 1 As shown. It mainly includes the following steps:
[0006] P0: Image preprocessing: Acquire source images of the same region at different time phases, and preprocess the image pairs to provide a standardized data foundation for subsequent analysis; this step includes:
[0007] P01: Global Registration: An image registration algorithm is used to spatially align images from different time periods to reduce overall geometric deviations caused by differences in shooting angle and time.
[0008] P02: Radiometric Correction: A radiometric normalization method is used to correct the brightness and contrast of the image pair in order to reduce the interference of factors such as changes in illumination on change detection.
[0009] S1: Global Reconnaissance and Guidance: Perform large-scale feature analysis on the processed image to generate a Region of Interest (ROI) instruction map for the execution layer; this specific step includes:
[0010] S11: Large-scale difference calculation: Using a large-scale feature description method, the difference degree of image pairs is calculated on sparse grid points to obtain an initial difference response map.
[0011] S12: Large-scale threshold calculation: A robust statistical threshold calculation method is used to calculate a stable baseline threshold for the original difference map that is not affected by noise suppression.
[0012] S13: Physical Attention Guidance: Introduce the Physical Attention Mechanism (PAM) to weight the initial difference response map to suppress the interference response of land features (such as vegetation and water bodies) that are not related to the change, and obtain a noise-suppressed guidance map.
[0013] S14: Generate ROI instruction map: Using the baseline threshold calculated in S12, filter the instruction map generated in S13 to generate an instruction map of regions of interest (ROI) with high, medium and low confidence levels, and send the instructions to the execution layer.
[0014] S2: Local Precise Execution: Receives and parses the instruction graph of the ROI from the guidance layer, performs fine-grained operations within its defined area to confirm changes; this step specifically includes:
[0015] S21: Traverse ROI regions: Traverse each connected region in the ROI instruction graph.
[0016] S22: Dynamic Local Reregistration: For each candidate region, feature matching and transformation model calculation are performed again to achieve high-precision local registration of the region, so as to further reduce the residuals and disparities that may be left in the region by global registration (P01).
[0017] S23: Fine-scale detection at small and medium scales: Change detection is performed within the finely registered region using small and medium-scale features.
[0018] S24: Dynamic threshold adjustment: such as Figure 3 As shown, the system receives a weight map from the guidance layer and dynamically adjusts the sensitivity threshold for detecting small-scale change points based on the confidence level (i.e., weight) assigned to the region in the guidance layer.
[0019] S3: Result Fusion and Output: Fusion of detection results from all regions, output of the final change map, and transmission of the results to the display and storage module.
[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0021] 1. Fundamentally improves the accuracy and reliability of change detection: By introducing a "dynamic local fine registration" mechanism, after locking the suspected area at the global guidance layer, the disparity and registration residual within that area are specifically reduced at the local execution layer, geometrically mitigating false detections caused by these factors. Simultaneously, the Physical Attention (PAM) mechanism effectively suppresses interference responses from high-frequency ground features such as vegetation and water bodies, significantly reducing the false alarm rate at the semantic level.
[0022] 2. Achieved intelligent optimization of computing resources and significant improvement in efficiency: Through the “guidance-execution” hierarchical decision-making framework, the computationally intensive fine processing (local registration, small-scale detection) is strictly limited to the high-value suspected areas defined by the guidance layer. This avoids the waste of resources caused by the indiscriminate calculation of the entire map in traditional methods, and enables the processing of a larger range of data with limited computing power, which greatly improves detection efficiency.
[0023] 3. It has strong scene adaptability and cost advantages: The entire process is based on traditional features and robust statistical methods, without relying on a large amount of labeled data for model training. This fundamentally overcomes the bottleneck problems of poor generalization ability and high data cost of deep learning models in new scenarios.
[0024] 4. Achieves rapid deployment and high engineering applicability at the edge: This method is based on traditional feature analysis, and its computational load is much lower than that of deep learning inference. It is easy to deploy to edge terminals such as drones. While ensuring accuracy, it provides a solution that can be deployed immediately without training, providing highly reliable technical support for tasks such as emergency response and inspection of unknown areas.
[0025] 5. The system architecture boasts advantages of scientific division of labor and synergistic benefits: The "guidance-execution" framework proposed in this invention is not a simple process concatenation, but rather a highly efficient collaborative framework. This architecture decouples global coarse screening from local fine screening at the system design level, allowing the guidance layer and execution layer to be optimized independently and perform their respective functions. This not only achieves the resource optimization efficiency improvement mentioned in point two, but also brings deeper advantages such as enhanced system reliability, functional modularity, and strong scalability. Attached Figure Description
[0027] Figure 1 This is a general framework diagram of the change detection system proposed in this invention;
[0028] Figure 2 A flowchart illustrating the generation of Region of Interest (ROI) instructions in the global guidance layer of this invention;
[0029] Figure 3 This is a flowchart of the dynamic local reregistration and fine detection process at the local execution layer of this invention;
[0030] Figure 4 This is a schematic diagram illustrating the generation principle of the Physical Attention Mechanism (PAM) of this invention. Detailed Implementation
[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The described embodiments are only some embodiments of this invention, not all embodiments.
[0033] like Figure 1 The diagram shows the overall framework of a change detection system proposed in this invention. The method in this embodiment includes the following steps:
[0034] P0: Image preprocessing module. For example... Figure 1As shown, this module receives T1 and T2 images as input. The module includes two sub-steps:
[0035] P01: The SIFT feature matching algorithm is used to calculate the global homography matrix between T1 and T2 images, and the T2 image is transformed to make it coarsely aligned with the T1 image in space.
[0036] P02: Preferably, the Reinhard color transfer algorithm is used to perform color and illumination correction on the T2 image, based on the T1 image. A specific implementation example is as follows:
[0037] 1. Convert T1 and T2 images from BGR color space to CIELAB color space.
[0038] 2. Calculate the mean values of T1 and T2 images in the L, a, and b channels. , ) and standard deviation ( , ).
[0039] 3. Channel values for T2 images The corrected channel values are obtained by applying the following transformation. :
[0040] ( ) + (1)
[0041] 4. Convert the corrected T2 image back to BGR space.
[0042] This module outputs the "registered and corrected image pair" to step S1.
[0043] S1: Global guidance module. For example... Figure 1 As shown, this module receives preprocessed image pairs, and its core task is to generate a ROI instruction graph and weight graph to guide the execution layer. Figure 2 As shown, Figure 2 Here is a detailed flowchart of the internal process of step S1:
[0044] S11: As Figure 2 As shown, the image pair after registration and correction from P0 is received, and preferably, SIFT descriptors of the T1 and T2 images are calculated at sparse grid points p. The "large-scale difference map" D(p) is defined as the L2 norm (Euclidean distance) of the two:
[0045] (2)
[0046] S12: As Figure 2 As shown, this "large-scale difference map" is directly used to calculate the "stability baseline threshold". Preferably, this threshold... Calculated using "Median Absolute Deviation (MAD)":
[0047] 1. Let D be the set of all non-zero difference values D(p) in S11.
[0048] 2. Calculate the median of D. .
[0049] 3. Calculate the absolute deviation of the median. .
[0050] 4. Stable baseline threshold Defined as:
[0051] (3)
[0052] Where k is an empirical constant, preferably, the value of k is between 2.5 and 3.5.
[0053] S13: As Figure 2 As shown, the T1 image was used to "generate a Physical Attention Map (PAM)". Figure 4 As shown, the PAM is preferably generated in the following manner:
[0054] Vegetation Analysis Module: Extracts vegetation areas by analyzing hue in the HSV color space and generates a vegetation noise map. .
[0055] Water Analysis Module: Extracts water ripples and generates a water noise map by analyzing the blue channel and the Laplacian response. .
[0056] General Texture Analysis Module: Extracts other high-frequency textures by calculating local entropy, generating a general high-frequency texture map. .
[0057] The "weighted fusion module" preferably generates the final PAM(p) by weighted summation:
[0058] (4)
[0059] in, For preset weights, and .
[0060] S14: As Figure 2 As shown, the "Calculation Guidance Diagram" Preferably, this is achieved through inhibitory multiplication:
[0061] (5)
[0062] Where σ is the inhibition factor, and then S12 is used to calculate... right Perform multi-level threshold segmentation to generate a weighted graph and an ROI instruction graph, and send them to the S2 module.
[0063] The stable baseline threshold in step S12 is used to perform multi-level segmentation (such as high, medium and low threshold control) of the guidance map in step S13 into multi-level confidence regions. The segmentation results finally generate a weight map and an ROI instruction map, which are then sent to module S2.
[0064] S2: Local execution module. For example... Figure 1 As shown, it receives the ROI instruction graph and weight graph from module S1 and performs a precision operation. Figure 3 As shown, Figure 3 Here is a detailed flowchart of the internal process of step S2:
[0065] S21: As Figure 3 As shown, each connected region in the ROI instruction graph is traversed. Loops S22 to S24 are executed for each region.
[0066] S22: As Figure 3 As shown, for the current candidate region, preferably, an affine transformation model is used for high-precision local registration. By re-matching feature points within this small region, a 6-DOF affine matrix A and a translation vector t are calculated:
[0067] (6)
[0068] This reduces local registration errors caused by parallax.
[0069] S23: Calculate small-scale feature differences on a denser grid in the region after fine registration in S22.
[0070] S24: As Figure 3 As shown, the weighted graph passed in by S1 plays a role here. The local threshold at this point... Defined as the baseline threshold for this region (e.g., the median of small-scale differences in S23) and a modulation function The product of:
[0071] (7)
[0072] Where f(w) is a monotonically decreasing function of weight w(p), a preferred linear example is as follows:
[0073] (8)
[0074] Where w(p) is the weight (from 0 to 1) of the point from S1. It is the multiplier for the low-confidence region (e.g., 1.2). It is the multiplier for the high confidence region (e.g., 0.7).
[0075] The output of this loop is a "set of change points".
[0076] S3: Result fusion and output module. For example... Figure 1 As shown, this module receives all the "regional change point sets" output by S2. Preferably, the DBSCAN clustering algorithm is used to post-process all change point sets to remove outlier noise points and cluster the change points into blocks. Finally, the "final change map" is output, completing the detection.
[0077] It should be noted that the specific algorithms, parameters, and implementation methods described above are merely preferred embodiments to aid in understanding the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made based on the core ideas of the present invention should fall within the scope of protection of the present invention.
Claims
1. A method for detecting changes in remote sensing images based on hierarchical decision-making, characterized in that, Includes the following steps: (a) Calculate the large-scale difference map between the time phase 1 (T1) and time phase 2 (T2) images; (b) Generate a Physical Attention Map (PAM) for suppressing ground object interference based on the T1 image; (c) Combine the large-scale difference map and the PAM to generate one or more region of interest (ROI) instruction maps; (d) Perform dynamic local snap registration only within the area defined by the ROI instruction map to reduce local disparity and global registration residuals locally; (e) Perform fine-scale detection at small and medium scales within the area precisely registered in (d) and output the change results.
2. The method as described in claim 1, characterized in that, Step (c) in the global guidance process further includes: (c1) Calculate a stable baseline threshold from the large-scale difference map (a); (c2) Apply the PAM(b) to the large-scale difference map (a) to obtain the guidance map; (c3) The guidance map (c2) is segmented using the stable baseline threshold (c1) to generate the ROI instruction map.
3. The method as described in claim 1, characterized in that, The global guidance step (c) further includes generating a weight graph; Furthermore, step (e) further includes: dynamically adjusting the sensitivity of the fine detection at the small and medium scales using the weight map.
4. As described in claim 1, characterized in that, The generation of the Physical Attention Map (PAM) (b) includes at least the following steps: (b1) Vegetation analysis; (b2) Water analysis; (b3) General texture analysis.
5. The method as described in claim 4, characterized in that, The water analysis further includes identifying water ripples through color features and texture features such as Laplacian or local entropy.
6. The method as described in claim 1, characterized in that, Prior to step (a), the method further includes (p) an image preprocessing step, which includes at least: (p1) Global registration of image pairs is performed using a feature matching algorithm; (p2) Color migration or radiation correction.
7. A remote sensing image change system based on hierarchical decision-making, characterized in that, include: A global guidance module is configured as follows: (a) Calculate the large-scale difference map between the time phase 1 (T1) and time phase 2 (T2) images; (b) Generate a Physical Attention Map (PAM) for suppressing ground object interference; (c) Combine the difference map and PAM to generate one or more region of interest (ROI) instruction maps; A local execution module is configured as follows: (d) Perform local fine registration only within the area defined by the ROI instruction map to reduce parallax; (e) Perform fine-scale detection at small and medium scales within the finely registered region to output the changes.
8. The system as described in claim 7, characterized in that, When executing step (c), the global guidance module is further configured to: (c1) Calculate the stability baseline threshold from the large-scale difference map (a); (c2) Apply the PAM(b) to the large-scale difference map (a) to obtain the guidance map; (c3) The guidance map (c2) is segmented using the stable baseline threshold (c1) to generate the ROI instruction map.
9. The system as described in claim 7, characterized in that, The global guidance module is further configured to generate a weight graph; Furthermore, the local execution module is further configured to dynamically adjust the sensitivity of the fine detection (e) at the small and medium scale using the weight map.
10. The system as described in claim 7, characterized in that, It further includes a Physical Attention Map (PAM) generation module, which is configured to: (b1) Perform vegetation analysis; (b2) Perform water analysis; (b3) Perform general texture analysis.