Remote sensing small target detecting and positioning method based on color features and region registration
By adopting efficient methods based on color features and region registration in the satellite-based remote sensing system, the problem of large-scale data processing and accuracy improvement in remote sensing small target recognition technology is solved, and efficient single pixel size target detection and precise geographical location are achieved. Determine, suitable for spacecraft environments with limited computing power.
Patent Information
- Application Number
- CN202410811884.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-21
- Publication Date
- 2024-11-12
AI Technical Summary
The existing satellite-borne remote sensing small object recognition technology has difficulties in processing large-scale data and improving image registration accuracy, especially real-time requirements and computing efficiency problems. It is time-consuming and labor-intensive to obtain labeled data sets, and the image registration is illuminated. , the change in view angle and deformation have a great impact, making it difficult to improve the accuracy.
Using an efficient method based on color features and region registration, the remote sensing image is converted to the HSV color space by preprocessing, the small target is detected by using the optimized K-mean clustering algorithm, and the latitude and longitude positioning is obtained in combination with ORB registration technology to adapt to the computing power and storage limited In the spacecraft environment, efficient image processing technology is used to detect and accurately determine the geographic location of a single pixel-sized target.
It realizes efficient detection and precise geographical location determination in a satellite-based remote sensing system, with high accuracy of detection results, suitable for spacecraft environments with limited computing capabilities and storage, simplifying algorithms, reducing computing volume, and improving real-time performance.
Smart Images

Figure CN118941762A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the intersecting field of computer vision, remote sensing technology and geographic information system, and in particular to a remote sensing small target detection and positioning method based on color features and regional registration. Background Art
[0002] Modern satellite-borne remote sensing technology has entered a stage of high maturity and widespread application. Remote sensing image small target recognition technology refers to the use of remote sensing sensors carried by satellites or other space platforms to detect and identify small targets on the earth's surface, such as vehicles, buildings, radiation sources, etc., by analyzing remote sensing image data. This technology has important applications in military reconnaissance, security monitoring, traffic management and other fields.
[0003] With the development of remote sensing technology, the resolution of satellite sensors has been continuously improved, and more detailed surface information can be obtained. This provides a better basis for the recognition of small targets. In addition, with the improvement of computing power and the development of machine learning algorithms, it has become possible to use remote sensing images for automatic small target recognition. At present, some important research progress has been made in satellite-borne remote sensing small target recognition technology.
[0004] Through feature extraction, researchers use image processing and computer vision techniques to extract features from remote sensing images, such as texture, shape, color, etc., for the recognition and classification of small targets. Researchers have developed various target detection algorithms, including traditional feature-based methods and deep learning-based methods. Deep learning-based methods, such as convolutional neural networks (CNNs) and target detection networks (such as Faster R-CNN, YOLO, etc.), have achieved significant performance improvements in small target recognition.
[0005] Remote sensing images are usually represented by pixel coordinates, while longitude and latitude coordinates are the representation of geographic information. By registering remote sensing images with offline maps, the correspondence between remote sensing images and geographic space can be realized locally, thereby realizing the geographic positioning and object extraction of remote sensing images. Image registration technology has made some important research progress. Researchers extract feature points from images and match them between different images to achieve image alignment. Common feature points include corner points, edges, spots, etc. It is also possible to directly use the pixel information of the image for registration, and achieve alignment by minimizing the difference between images. Common methods include least squares, mutual information, phase correlation, etc. In addition, by selecting some control points with known positions in the image and geographic information and matching them, the image and geographic coordinates can be aligned. Common methods include manual selection of control points and automatic extraction of control points.
[0006] Although some progress has been made in satellite-borne remote sensing small target recognition technology, there are still some challenges and difficulties. With the continuous increase of remote sensing data, it has become more difficult to process and analyze large-scale remote sensing image data. At the same time, real-time requirements also pose challenges to the computational efficiency of the algorithm. Obtaining large-scale annotated data sets is very important for training deep learning models, but annotating remote sensing image data requires a lot of time and human resources. In image registration, remote sensing images may be affected by different lighting conditions, changes in perspective, deformation and other factors, which will lead to differences between images and increase the difficulty of registration. For the longitude and latitude positioning of small targets, high-precision registration results are required. How to improve the accuracy of registration and reduce registration errors is a challenge. Summary of the invention
[0007] To address the deficiencies of the prior art, the present invention proposes a remote sensing small target detection and positioning method based on color features and regional registration, aiming to improve the ability to detect and accurately determine the geographic location of targets of a single pixel size in the earth observation system through efficient image processing technology.
[0008] The present invention adopts the following technical solutions:
[0009] A remote sensing small target detection and positioning method based on color features and regional registration. Considering the limited environmental conditions on flight platforms such as spacecraft and the limited computing power and storage space of the deployment platform, the method does not use complex deep learning methods, but uses more efficient algorithms, which mainly include small target detection based on color features and longitude and latitude positioning based on image registration. The specific steps of the method include:
[0010] Step 1: After preprocessing the input remote sensing image, convert it into the HSV color space for color extraction.
[0011] Step 2: Use the optimized K-means clustering algorithm to find small targets and calculate their pixel coordinates in the image.
[0012] Step 3: Generate an offline map file based on the approximate area where the target is located. Use the background to call the browser to obtain map information and synthesize the map image.
[0013] Step 4: Perform ORB (Oriented FAST and Rotated BRIEF) registration on the input image and the map image, and simulate mouse clicks based on the calculated coordinates to parse and obtain the longitude and latitude information.
[0014] Preferably, in step 1, the input remote sensing image is first preprocessed to improve brightness, enhance contrast and remove noise according to actual conditions. Then, it is converted to a more distinguishable HSV color space. A mask is created according to the target feature color and applied to the input image to extract the image area within the specified color range.
[0015] Preferably, in step 2, the initial center selection strategy of the K-means clustering algorithm is optimized, and the distance between the data points is considered when selecting the initial center point, so that the distance between the initial center points is more uniform, thereby avoiding the clustering effect caused by the initial center point being too concentrated or dispersed. In addition, the image data is normalized and feature scaled before calculating the Euclidean distance. First, the result after feature color extraction is reshaped into a two-dimensional array, in which each row represents a pixel, each pixel is composed of the values of three channels, and then the data type of the pixel value is converted to float32. After defining the convergence criterion and the number of clusters, the K-means clustering algorithm is executed. According to the label of the pixel, the color feature of each pixel is obtained from the cluster center, and the one-dimensional array of the color feature is reshaped into the same shape as the original image, and converted into a binary image. The image is contour searched to draw the detection frame, and the pixel coordinates of the target center point in the image are calculated.
[0016] Preferably, in step 3, a map object HTML file is created according to the approximate longitude and latitude of the target area and offline map data. The external library and style sheet introduced in the tag as an external link resource are replaced with local files to achieve offline. Then, Selenium is used to create a WebDriver object of a Chrome browser and set it to borderless mode. The Chrome background is called to open the local map file, and the tile information of the image displayed in the center of the window is found and obtained. Finally, the required tile arrangement is calculated according to the size, and the local map tiles are spliced into a large map image.
[0017] Preferably, in step 4, after the input image and the map image are converted into grayscale images, they are subjected to ORB registration, and the image resolution is reduced by adopting a pyramid downsampling strategy to reduce the amount of calculation. After the matching results are screened by using a threshold, the mapping relationship between the two images is obtained, thereby calculating the pixel coordinates of the target in the map image. Finally, a longitude and latitude pop-up window control is added to the local map file, and Chrome is called in the background to open the map file, and the corresponding coordinates are simulated by clicking the mouse. The pop-up window element containing the longitude and latitude information is found according to the web page source code, and the longitude and latitude values are parsed using conditional judgment and string operations.
[0018] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:
[0019] 1. The present invention does not use complex deep learning methods, but uses more efficient algorithms. The technology is simple and efficient, and can effectively adapt to application scenarios where environmental conditions on flight platforms such as spacecraft are limited and the computing power and storage space of the deployment platform are limited;
[0020] 2. The present invention uses efficient image processing technology to effectively implement the detection and precise geographic location determination of targets of a single pixel size in the earth observation system. The detection results and positioning results have good visual effects and high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1 It is an overall flow chart of the remote sensing small target detection and positioning method based on color features and regional registration of the present invention;
[0022] Figure 2 is an input image of the present invention;
[0023] Figure 3 is the preprocessed input image of the present invention;
[0024] Figure 4 is an input image in the HSV color space of the present invention;
[0025] Figure 5 A binary image of the clustering result of the present invention;
[0026] Figure 6 This is a small target detection result diagram of the present invention;
[0027] Figure 7 This is a detailed magnified test result diagram;
[0028] Figure 8 The local map data of the present invention;
[0029] Fig. 9 is the registration result between the input image and the map image of the present invention;
[0030] Fig.10 The latitude and longitude positioning result on the map image of the present invention;
[0031] Fig.11 Analysis of the longitude and latitude window magnified for the details of the present invention;
[0032] Fig.12 For comparison between the present invention and manual annotation on Google Maps;
[0033] Fig.13 It is the latitude and longitude result manually marked on the Google map of the present invention. DETAILED DESCRIPTION
[0034] The present invention is described below based on specific embodiments, but the present invention is not limited to the embodiments. In the detailed description of the present invention below, some specific details are described in detail. For those skilled in the art, the present invention can be fully understood without the description of these details.
[0035] A remote sensing small target detection and positioning method based on color features and regional registration, the overall flow chart of the method is as follows Figure 1 As shown, the following steps are included:
[0036] Step 1: After preprocessing the input remote sensing image, convert it into the HSV color space for color extraction.
[0037] Step 2: Use the optimized K-means clustering algorithm to find small targets and calculate their pixel coordinates in the image.
[0038] Step 3: Generate an offline map file based on the approximate area where the target is located. Use the background to call the browser to obtain map information and synthesize the map image.
[0039] Step 4: Perform ORB registration on the input image and the map image, simulate mouse clicks based on the calculated coordinates, and parse the longitude and latitude information.
[0040] Preferably, in step 1, the input remote sensing image is first preprocessed to improve brightness, enhance contrast and remove noise according to actual conditions. Figure 2 As shown, the preprocessed image is Figure 3 As shown. Then convert it to the more distinguishable HSV color space, as Figure 4 Then a mask is created based on the target feature color and applied to the input image to extract the image area within the specified color range.
[0041] Preferably, in step 2, the initial center selection strategy of the K-means clustering algorithm is optimized, and the distance between data points is considered when selecting the initial center point, so that the distance between the initial center points is more uniform, thereby avoiding the clustering effect affected by the initial center point being too concentrated or dispersed. In addition, the image data is normalized and feature scaled before calculating the Euclidean distance. First, the result after feature color extraction is reshaped into a two-dimensional array, in which each row represents a pixel, and each pixel consists of the values of three channels, and then the data type of the pixel value is converted to float32 type. After defining the convergence criterion and the number of clusters, the K-means clustering algorithm is executed. According to the pixel label, the color feature of each pixel is obtained from the cluster center, and the one-dimensional array of the color feature is reshaped into the same shape as the original image, and converted into a binary image, such as Figure 5The binary image is contour searched to draw the detection frame and the pixel coordinates of the target center point in the image are calculated. The detection results of small targets are shown in Figure 6 As shown, Figure 7 A magnified image of the test results.
[0042] Preferably, in step 3, a map object HTML file is created based on the latitude and longitude of the target area and the offline map data. Figure 8 As shown. Replace the external libraries and style sheets introduced in the tag with external link resources with local files to achieve offline. Then use Selenium to create a WebDriver object of the Chrome browser and set it to borderless mode. Call the Chrome background to open the local map file, find and obtain the tile information of the image displayed in the center of the window, and then obtain the tile arrangement pattern of the entire map. In order to accurately align the input image with the map, it is necessary to obtain a map image as large as possible. Use 3 times the longest side of the input image as the size of the map image. Finally, calculate the required tile arrangement according to the size, and splice the local map tiles into a large map image.
[0043] Preferably, in step 4, after the input image and the map image are converted into grayscale images, they are subjected to ORB registration, and a pyramid downsampling strategy is used to reduce the image resolution and the amount of calculation. After the matching results are screened using a threshold, the mapping relationship between the two images is obtained, such as Fig. 9 As shown, the pixel coordinates of the target in the map image are calculated. Add a latitude and longitude pop-up window control to the local map file, call Chrome in the background to open the map file, and create an ActionChains object to simulate the mouse click operation, such as Fig.10 As shown. Use the move_by_offset method to move the mouse to the coordinate point of the small target in the map image, use the click method to simulate the mouse click operation, and call perform to execute the action. Finally, find the pop-up window element containing longitude and latitude information according to the web page source code, obtain the text content, split it into multiple lines according to the line break, use conditional judgment and string operation to parse the longitude and latitude values, as shown below. Fig.11 shown.
[0044] The experiment was verified by manual annotation on Google Maps, such as Fig.12 shown. Fig.13 The longitude and latitude results obtained by manually marking the small target have a small error with the longitude and latitude results obtained in the experiment. And the target is searched on Google Maps based on the longitude and latitude results obtained in the experiment, and the error with the original target is within 4 pixels at the same resolution.
[0045] Finally, it should be noted that the specific embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to this embodiment, a person skilled in the art should understand that the technical solutions described in the foregoing specific embodiments may still be modified, or some or all of the technical features may be replaced by equivalents, but these modifications and replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A remote sensing small target detection and positioning method based on color features and regional registration, characterized in that: The method does not use a complex deep learning method, but uses a more efficient algorithm, which mainly includes small target detection based on color features and longitude and latitude positioning based on image registration, including the following steps: Step 1: After preprocessing the input remote sensing image, convert it into the HSV color space for color extraction; Step 2: Use the optimized K-means clustering algorithm to find small targets and calculate their pixel coordinates in the image; Step 3: Generate an offline map file based on the approximate area where the target is located, and obtain map information by calling the browser in the background to synthesize the map image; Step 4: Perform ORB registration on the input image and the map image, simulate mouse clicks based on the calculated coordinates, and parse the longitude and latitude information.
2. The remote sensing small target detection and positioning method based on color features and regional registration according to claim 1 is characterized in that: In step 1, firstly, the input remote sensing image is preprocessed to improve brightness, enhance contrast and remove noise according to actual conditions; secondly, the input image is converted to a more distinguishing HSV color space, and a mask is created according to the target feature color, which is applied to the input image to extract the image area within the specified color range.
3. The remote sensing small target detection and positioning method based on color features and regional registration according to claim 1 is characterized in that: In the step 2, the initial center selection strategy of the K-means clustering algorithm is optimized, and the distance between data points is considered when selecting the initial center point, so that the distance between the initial center points is more uniform, thereby avoiding the clustering effect affected by the initial center point being too concentrated or dispersed; in addition, the image data is normalized and feature scaled before calculating the Euclidean distance; first, the result after feature color extraction is reshaped into a two-dimensional array, in which each row represents a pixel, and each pixel is composed of values of three channels, and then the data type of the pixel value is converted to float32; after defining the convergence criterion and the number of clusters, the K-means clustering algorithm is executed; according to the pixel label, the color feature of each pixel is obtained from the cluster center, and the one-dimensional array of the color feature is reshaped into the same shape as the original image, and converted into a binary image; the image is contour searched to draw a detection frame, and the pixel coordinates of the target center point in the image are calculated.
4. The remote sensing small target detection and positioning method based on color features and regional registration according to claim 1 is characterized in that: In the step 3, a map object HTML file is created according to the approximate longitude and latitude of the target area and the offline map data; the external library and style sheet introduced in the tag as external link resources are replaced with local files to achieve offline operation; then Selenium is used to create a WebDriver object of a Chrome browser and set it to borderless mode; the Chrome background is called to open the local map file, and the tile information of the image displayed in the center of the window is found and obtained; finally, the required tile arrangement is calculated according to the size, and the local map tiles are spliced into a large map image.
5. The remote sensing small target detection and positioning method based on color features and regional registration according to claim 1 is characterized in that: In the step 4, after the input image and the map image are converted into grayscale images, they are ORB registered, and a pyramid downsampling strategy is used to reduce the image resolution and reduce the amount of calculation; After using the threshold to filter the matching results, the mapping relationship between the two images is obtained, so as to calculate the pixel coordinates of the target in the map image; finally, a longitude and latitude pop-up window control is added to the local map file, and Chrome is called in the background to open the map file and simulate a mouse click on the corresponding coordinates; the pop-up window element containing the longitude and latitude information is found according to the source code of the web page, and the longitude and latitude values are parsed using conditional judgment and string operations.
Citation Information
Patent Citations
Bayonet vehicle positioning method based on Gaussian color distance
CN105447448A
Optical remote sensing image target detection method based on FPGA heterogeneous deep learning
CN110516560A
Self-adaptive target segmentation method in remote sensing scene based on spectral characteristics
CN112017158A
Method and system for collaborative multi-satellite remote sensing
US20170235996A1
