Target recognition method, device, equipment and medium based on remote sensing image

Abstract

The present invention provides a target recognition method, device, equipment and medium based on remote sensing images, including: determining multiple segmentation masks according to a mask image corresponding to a target remote sensing image; and determining the spots corresponding to each of the multiple segmentation masks from the mask image; screening the multiple spots according to the first aspect ratio and the first area corresponding to each of the multiple spots to determine a launch pad image; determining a vehicle status image according to the launch pad image and a segmentation threshold, combined with the spot shape characteristics; determining a target recognition result corresponding to the target remote sensing image according to the launch pad image and the vehicle status image. The multiple segmentation masks can solve the problem of fine segmentation accuracy and poor precision of the traditional segmentation model, so as to obtain a launch pad image with better universality; determining the vehicle status image according to the segmentation threshold and the spot shape characteristics, and then combining the launch pad image, can extract deep-level target intelligence semantics, and obtain a recognition result with high accuracy and richness.

Description

Technical Field

[0001] The present invention relates to the technical field of image data processing, and in particular to a target recognition method, device, equipment and medium based on remote sensing images. Background Art

[0002] With the development of data collection platforms such as satellites, drones, and ground survey vehicles, the data sources for target detection and recognition are becoming increasingly abundant. Remote sensing images have broad application prospects in target recognition due to their advantages such as large coverage, periodic observation, and long-distance shooting. They provide important data support and decision-making basis for urban planning, environmental monitoring, resource management, and disaster management. At present, target recognition based on remote sensing images mainly includes: 1. Typical objects such as buildings, roads, and water bodies; 2. Typical facilities such as airports, ports, chemical plants, and schools; 3. Typical states such as aircraft, ships, and vehicles; 4. Typical disasters such as floods, droughts, landslides, and fires. The recognition content of the entire process is single, lacking an overall understanding of the system, and the recognition results are mainly location and range information, lacking deep semantic information.

[0003] The target recognition methods based on remote sensing images mainly include image processing methods, machine learning methods and deep learning methods. Image processing methods mainly rely on manually designed feature extraction, including edge detection, morphological operations, histogram analysis and texture analysis. These methods are intuitive, fast, and require low computing resources, but they are difficult to effectively separate and identify complex targets and backgrounds. Machine learning methods rely on statistical models and feature engineering, including support vector machines, decision trees, random forests, and K-means clustering. These methods are highly flexible, but have large data requirements and high model complexity. In recent years, deep learning methods have performed well in target recognition based on remote sensing images, especially convolutional neural networks (CNNs), which have the advantages of being able to process complex patterns and having strong generalization capabilities, but they also have disadvantages such as large data requirements, high computing resource consumption, and difficult model interpretation.

[0004] In other words, the existing target recognition methods based on remote sensing images have certain limitations, resulting in the lack of deep semantic information in the final recognition results, that is, the accuracy of the recognition results is low. Summary of the invention

[0005] The present invention provides a target recognition method, device, equipment and medium based on remote sensing images, which are used to solve the problem that the existing target recognition methods based on remote sensing images have certain limitations, resulting in a lack of deep semantic information in the final recognition results, that is, the defect of low accuracy of the recognition results. Through multiple segmentation masks, the problems of fine segmentation accuracy and poor accuracy of traditional segmentation models can be effectively solved to obtain a segmentation result with better universality (that is, a launch site image); then, according to the segmentation threshold and combined with the shape characteristics of the pattern, a vehicle status image with higher accuracy is determined from the launch site image, and then combined with the above-mentioned launch site image, deep target intelligence semantics can be extracted to obtain a target recognition result with higher accuracy and richness.

[0006] The present invention provides a target recognition method based on remote sensing images, comprising the following steps.

[0007] According to the mask image corresponding to the target remote sensing image, a plurality of segmentation masks are determined; and from the mask image, the image spots corresponding to each of the plurality of segmentation masks are determined.

[0008] The plurality of image spots are screened according to the first aspect ratios and the first areas respectively corresponding to the plurality of image spots to determine the emission field image.

[0009] The vehicle status image is determined according to the emission field flat image and the segmentation threshold value in combination with the pattern shape characteristics.

[0010] A target recognition result corresponding to the target remote sensing image is determined according to the launch site image and the vehicle status image.

[0011] According to a target recognition method based on remote sensing images provided by the present invention, the multiple spots are screened according to the first aspect ratios and first areas corresponding to each of the multiple spots to determine the launch pad image, including: for each spot, determining the first aspect ratio of the spot according to the outer contour vertices of the spot; and determining the first area of ​​the spot according to the projection coordinates of the outer contour vertices; retaining the spots corresponding to the first aspect ratio that meet the first screening condition and the first area that meet the second screening condition among the multiple spots, to obtain a first image containing launch pad features; wherein the first screening condition and the second screening condition are determined based on preset launch pad features in expert knowledge; and using a multi-scale mask classification framework to extract and enhance features of the first image to determine the launch pad image.

[0012] According to a target recognition method based on remote sensing images provided by the present invention, the vehicle status image is determined according to the launch site flat image and the segmentation threshold in combination with the pattern shape characteristics, including: determining the grayscale image of the launch site flat image, and the grayscale values ​​corresponding to all pixels in the grayscale image; determining the grayscale mean and the grayscale standard deviation according to all the grayscale values; determining the segmentation threshold according to the grayscale mean and the grayscale standard deviation; retaining the pixels corresponding to the pixel values ​​less than the segmentation threshold among all the pixels, and obtaining the vehicle status image in combination with the pattern shape characteristics.

[0013] According to a target recognition method based on remote sensing images provided by the present invention, the target recognition result corresponding to the target remote sensing image is determined according to the launch pad image and the vehicle status image, including: determining the shape and configuration of the launch pad according to the launch pad image; determining the mounting mode of the vehicle status according to the vehicle status image; and determining the target recognition result corresponding to the target remote sensing image according to the shape, the configuration and the mounting mode.

[0014] According to a target recognition method based on remote sensing images provided by the present invention, the first aspect ratio of the image spot is determined according to the outer contour vertices of the image spot, including: determining the center point of the image spot according to the outer contour vertices of the image spot; rotating the image spot around the center point multiple times to obtain multiple rotated images, and determining the envelope rectangles corresponding to each of the multiple rotated images; determining a circumscribed rectangle according to the target envelope rectangle with the smallest area among the multiple envelope rectangles; and determining the first aspect ratio of the image spot according to the circumscribed rectangle.

[0015] According to a target recognition method based on remote sensing images provided by the present invention, the multi-scale mask classification framework includes: sparse convolution, and the optimal layer combination in the neural network architecture search model; the multi-scale mask classification framework is used to extract and enhance the features of the first image to determine the launch site flat image, including: using the sparse convolution to perform feature encoding on the first image to obtain a second image; using the optimal layer combination to extract and enhance the features of the first image to obtain a third image; and fusing the second image and the third image to obtain the launch site flat image.

[0016] According to a target recognition method based on remote sensing images provided by the present invention, determining the mounting mode of the vehicle state according to the vehicle state image includes: when the second area of ​​the vehicle state image satisfies the third filtering condition, determining the second aspect ratio of the vehicle state image; wherein the third filtering condition is determined based on a preset vehicle state feature in expert knowledge; when the second area satisfies the fourth filtering condition, or the second aspect ratio satisfies the fifth filtering condition, determining that there is a tractor mounted as the mounting mode of the vehicle state; wherein the fourth filtering condition and the fifth filtering condition are determined based on the preset vehicle state feature; when the second area does not satisfy the fourth filtering condition, and the second aspect ratio does not satisfy the fifth filtering condition, determining that there is no tractor mounted as the mounting mode.

[0017] The present invention also provides a target recognition device based on remote sensing images, comprising the following modules.

[0018] The image determination module is used to determine multiple segmentation masks according to the mask image corresponding to the target remote sensing image; and determine the spots corresponding to each of the multiple segmentation masks from the mask image; screen the multiple spots according to the first aspect ratio and the first area corresponding to each of the multiple spots to determine the launch site flat image; determine the vehicle status image according to the launch site flat image and the segmentation threshold combined with the spot shape characteristics.

[0019] The target recognition module is used to determine the target recognition result corresponding to the target remote sensing image according to the launch site image and the vehicle status image.

[0020] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, any of the above-mentioned target recognition methods based on remote sensing images is implemented.

[0021] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the method for target recognition based on remote sensing images as described in any one of the above is implemented.

[0022] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements any of the above-mentioned target recognition methods based on remote sensing images.

[0023] The target recognition method, device, equipment and medium based on remote sensing images provided by the present invention determine multiple segmentation masks according to the mask image corresponding to the target remote sensing image; and determine the spots corresponding to each of the multiple segmentation masks from the mask image; screen the multiple spots according to the first aspect ratio and the first area corresponding to each of the multiple spots to determine the launch pad image; determine the vehicle status image according to the launch pad image and the segmentation threshold, combined with the spot shape characteristics; determine the target recognition result corresponding to the target remote sensing image according to the launch pad image and the vehicle status image. The method can effectively solve the problem of fine segmentation accuracy and poor accuracy of the traditional segmentation model through multiple segmentation masks, so as to obtain a segmentation result with better universality (i.e., the launch pad image); then determine the vehicle status image with higher accuracy from the launch pad image according to the segmentation threshold and the spot shape characteristics, and then combine the above-mentioned launch pad image to extract the deep-level target intelligence semantics and obtain a target recognition result with higher accuracy and richness. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0025] Figure 1 It is a flow chart of the target recognition method based on remote sensing images provided by the present invention.

[0026] Figure 2 It is a schematic diagram of a missile position provided by the present invention.

[0027] Figure 3 It is a schematic diagram of the remote sensing image sample provided by the present invention.

[0028] Figure 4 It is a schematic diagram of multiple pattern spots provided by the present invention.

[0029] Figure 5 It is a schematic diagram of a grayscale image of the emission field flat image provided by the present invention.

[0030] Figure 6 It is a schematic diagram of the structure of a target recognition device based on remote sensing images provided by the present invention.

[0031] Figure 7 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] It should be noted that the execution subject involved in the embodiment of the present invention can be a target recognition device based on remote sensing images, or it can be an electronic device. Optionally, the electronic device can include: a computer, a mobile terminal, a wearable device, etc.

[0034] The embodiment of the present invention is further described below by taking an electronic device as an example.

[0035] Figure 1 is a flow chart of a target recognition method based on remote sensing images provided by the present invention, such as Figure 1 As shown, the method includes the following steps 101 to 104.

[0036] Step 101: Determine a plurality of segmentation masks according to a mask image corresponding to a target remote sensing image; and determine, from the mask image, the image spots corresponding to each of the plurality of segmentation masks.

[0037] Among them, the target remote sensing image is processed and has more accurate information about the earth's surface.

[0038] A mask image is a binary image (usually composed of 0 and 1) that indicates which areas are of interest (value 1) and which areas are of no interest (value 0). In remote sensing image processing, mask images are often used to extract specific areas or filter out unnecessary information.

[0039] Segmentation mask refers to the mask used to segment the mask image into different regions or objects. In remote sensing image processing, segmentation mask can represent different types of objects, target objects, etc.

[0040] A patch usually refers to a continuous area with specific attributes identified by methods such as image segmentation. These areas appear as a collection of pixels with similar features (such as color, texture, shape, etc.) on the image.

[0041] After determining the target remote sensing image, the electronic device can first determine the mask image corresponding to the target remote sensing image, which can reduce the subsequent calculation amount, and then analyze the mask image to obtain multiple segmentation masks. These segmentation masks can accurately reflect the boundaries and shapes of different objects in the target remote sensing image, thereby improving the accuracy of information extraction; then, based on the multiple segmentation masks, the electronic device determines the spots corresponding to each of the multiple segmentation masks from the above mask image, so as to prepare for the subsequent screening of the multiple images to obtain a launch field image with higher accuracy.

[0042] Step 102: Screen the plurality of image spots according to the first aspect ratios and the first areas corresponding to the plurality of image spots, and determine a launch field image.

[0043] The first aspect ratio is used to characterize the shape of the corresponding pattern spot; and the first area is used to characterize the size of the corresponding pattern spot.

[0044] A launch pad image refers to an image containing launch pad features.

[0045] In the process of determining the launch pad image, the electronic device can first determine the first aspect ratio and the first area corresponding to each of the multiple spots, and then screen the multiple spots based on all the first aspect ratios and all the first areas to filter out unnecessary features (such as non-launch pad features) and obtain a launch pad image with higher accuracy. The entire process can more accurately match the specific features of the launch pad by calculating the aspect ratio and area of ​​the spots. The launch pad usually has a relatively uniform aspect ratio and area range. This process can effectively exclude spots whose aspect ratio or area does not match the launch pad, thereby improving the accuracy of subsequent target recognition.

[0046] Step 103: Determine the vehicle status image based on the launch site image and the segmentation threshold in combination with the pattern shape characteristics.

[0047] The vehicle status image refers to an image containing vehicle status features corresponding to the missile launcher.

[0048] Optionally, the pattern shape characteristics may include characteristic parameters such as aspect ratio and area of ​​the pattern.

[0049] Since the launch site image also contains vehicle status features, and there are differences in the pixel values ​​of the site features and the vehicle status features, the electronic device uses a segmentation threshold to segment the launch site image, and then judges the vehicle status based on the segmentation result according to the pattern shape features to obtain a vehicle status image with high accuracy. The whole process does not require the vehicle status features to be identified separately, and the vehicle status image can be obtained quickly and accurately by using the segmentation threshold and pattern shape features.

[0050] Step 104: Determine the target recognition result corresponding to the target remote sensing image according to the launch site image and the vehicle status image.

[0051] Among them, the target recognition result usually includes the missile position features corresponding to the target remote sensing image, the launch pad features contained in the launch pad image, and the vehicle status features contained in the vehicle status image.

[0052] Specifically, the target identification results include: the location and range of the target; the location, range and number of key components of the target; and the location and status of high-value sub-targets.

[0053] Among them, the target refers to the missile position; the key component of the target refers to the launch pad; and the sub-target refers to the above-mentioned missile launch vehicle.

[0054] In an embodiment of the present invention, multiple segmentation masks are determined according to a mask image corresponding to a target remote sensing image; and spots corresponding to each of the multiple segmentation masks are determined from the mask image; multiple spots are screened according to the first aspect ratio and the first area corresponding to each of the multiple spots to determine a launch pad image; a vehicle status image is determined according to the launch pad image and a segmentation threshold, combined with the spot shape features; and a target recognition result corresponding to the target remote sensing image is determined according to the launch pad image and the vehicle status image. This method can effectively solve the problem of fine segmentation accuracy and poor accuracy of the traditional segmentation model through multiple segmentation masks, so as to obtain a segmentation result with better universality (i.e., the launch pad image); then, according to the segmentation threshold, combined with the spot shape features, a vehicle status image with higher accuracy is determined from the launch pad image, and then combined with the above-mentioned launch pad image, deep-level target intelligence semantics can be extracted to obtain a target recognition result with higher accuracy and richness.

[0055] In order to better understand the embodiments of the present invention, the target recognition method based on remote sensing images is further described below.

[0056] Optionally, before step 101, the method may also include: the electronic device acquires an initial remote sensing image covering the mission area; the electronic device uses a small domain model to perform a wide-area search and recognition on the initial remote sensing image to obtain at least one position detection range; the electronic device crops the initial remote sensing image according to the at least one position detection range to obtain at least one cropped image; the electronic device determines a randomly determined cropped image from the at least one cropped image as the target remote sensing image.

[0057] The initial remote sensing image is an unprocessed Earth surface information collected by satellite or aircraft, and the initial remote sensing image may include at least one missile position (referred to as position). R C×H×WRepresents a three-dimensional array (or tensor), C represents the number of color channels of the initial remote sensing image; H represents the height of the initial remote sensing image; W represents the width of the initial remote sensing image.

[0058] The domain small model is a transformer-enhanced feature extractor that includes two feature alignment heads, namely feature classification alignment head and feature regression alignment head, which provide a better balance between learning task interactions and task-specific features, and are more flexible in learning alignment through task alignment predictors.

[0059] In the process of determining the target remote sensing image, the electronic device can first obtain a large-scale remote sensing image covering the mission area, that is, an initial remote sensing image; in order to realize the extraction and classification of position features in the initial remote sensing image, the initial remote sensing image can be input into the domain model, and the regression and classification of position features can be realized through the two feature alignment heads in the domain model to obtain at least one position detection range; then, the initial remote sensing image is cropped one by one according to the at least one position detection range to obtain the cropped images corresponding to the at least one position detection range, and then a randomly determined cropped image from all the cropped images is determined as the target remote sensing image. The whole process is aimed at realizing a fast and accurate search for positions from the initial remote sensing image and judging the category and model of each position.

[0060] Among them, at least one of the missile positions (such as Figure 2 (as shown) is the deployment location of the air defense missile system. Specifically, the missile position is mainly composed of the launch area (i.e., the launch position), the guidance area (i.e., the guidance position), the technical support position (i.e., the technical support area) and the camp. For the missile position, the launch pad is a key component of the missile position. It is a ground project and facility built for missile launches, and is also an important support for the parking and launching of missile launchers; the missile launcher is a high-value sub-target of the missile position. It is movable and plays an important role in intercepting the launched missiles. It is the main equipment for the air defense missile system. Based on this, according to the launch pad and the missile launcher, the combat readiness level, mobility, and even mission scope of the launch position can be analyzed and judged. In addition, the shape, number, and size of the launch pads in different missile positions vary depending on the country and the form of position construction.

[0061] It should be noted that the above-mentioned air defense missile systems are all all-weather, large-airspace, multi-channel systems, mainly used for regional air defense and key point air defense. The task of this air defense missile system is to provide various types of air defense protection to protect important institutions and mobile forces from various types of enemy air weapons threats from low to high altitudes.

[0062] Optionally, before step 102, the method may further include: the electronic device performs a remote sensing image sample (such as Figure 3 As shown in the figure, the electronic device performs analysis and identification detection to obtain preset launch site characteristics and preset vehicle status characteristics; the electronic device generates expert knowledge based on the preset launch site characteristics and the preset vehicle status characteristics.

[0063] Among them, the remote sensing image samples contain launch site samples and missile launch vehicle samples. The missile launch vehicle samples are usually parked in the launch site samples.

[0064] When analyzing and identifying remote sensing image samples, the electronic equipment determines the shape and configuration of the launch site samples as preset launch site features, and determines the mounting method of the missile launch vehicle samples as preset vehicle state features, thereby generating expert knowledge.

[0065] Optionally, the shape of the launch pad sample is a square or a rectangle; the configuration of the launch pad sample is a separate configuration or a connected configuration.

[0066] In addition, the materials of launch site samples usually include cement concrete, asphalt concrete and hard soil.

[0067] Among them, in the remote sensing image samples, the launch site samples corresponding to cement concrete appear white, the launch site samples corresponding to asphalt concrete appear gray and / or black, and the launch site samples corresponding to hard soil appear grayish white, which can be easily confused with building roofs and hardened ground.

[0068] Exemplarily, the length of a launch pad sample configured individually is 11 meters to 30 meters, and the width is 7 meters to 21 meters; the length of a launch pad sample configured in a connected manner is 70 meters to 90 meters, and the width is approximately 30 to 50 meters.

[0069] Optionally, the missile launcher vehicle sample can be mounted in two ways: with a tractor (i.e. with a front vehicle) and without a tractor (i.e. without a front vehicle).

[0070] Among them, the missile launcher samples mounted on tractors have the ability to quickly maneuver and transfer, and can be maneuvered at any time according to mission or combat needs; the missile launcher samples without tractors are usually responsible for normal air defense combat readiness duty tasks in key areas of the area and do not need to be quickly maneuvered according to the mission.

[0071] Optionally, the method may also include: the electronic device analyzes and identifies the remote sensing image samples to obtain preset missile status characteristics; the electronic device generates expert knowledge based on the preset missile status characteristics, preset launch site characteristics and preset vehicle status characteristics.

[0072] Among them, the remote sensing image samples also contain missile samples, and the missile samples are usually placed in the missile launcher vehicle samples.

[0073] The electronic equipment analyzes and identifies the remote sensing image samples, determines the missile status corresponding to the missile sample as the preset missile status characteristics, and then combines the preset launch site characteristics and preset vehicle status characteristics to generate expert knowledge with a high degree of completeness.

[0074] Optionally, the missile state corresponding to the missile sample is a launch standby state or a horizontal state.

[0075] It should be noted that due to the presence or absence of the vehicle head and the state of the missile, the shadow of the missile launcher vehicle sample also presents different shapes in the remote sensing image samples taken at different time points. Figure 3 From the above, we can see that the missile launcher samples in the first and second rows are mounted without a tractor, while the missile launcher samples in the third and fourth rows are mounted with a tractor. The missile launcher samples with a front are longer, and when there is a shadow, the shadow area of ​​the missile launcher samples with a front is also larger. Based on this, the use of length and aspect ratio is helpful to distinguish the vehicle status of the missile launcher samples.

[0076] Optionally, the electronic device determines a plurality of segmentation masks according to a mask image corresponding to the target remote sensing image, which may include: the electronic device determines the mask image corresponding to the target remote sensing image; the electronic device inputs the mask image into a Segment Anything Model (SAM) to obtain a plurality of segmentation masks output by the Segment Anything Model. The SAM may include an image encoder, a hint encoder, and a mask decoder.

[0077] In the process of determining multiple segmentation masks, the electronic device may first use the position detection range corresponding to the target remote sensing image as the mask area, that is, the mask image; then, use the image encoder to encode the mask image to obtain a high-dimensional, dense feature representation, that is, the image embedding feature; then use the prompt encoder to receive the manually input encoding information, which includes: point information, frame information, and text information; then, fuse the image embedding feature and the encoding information to obtain fused information, and use the mask decoder to decode the fused information to obtain multiple segmentation masks. The whole process realizes the segmentation of the mask image, so that the electronic device can subsequently determine the target image segmentation vector data (such as Figure 4 As shown), each target image segmentation vector data is a spot.

[0078] In some embodiments, the electronic device screens a plurality of spots according to first aspect ratios and first areas corresponding to each of the plurality of spots to determine a launch pad image, which may include: for each spot, the electronic device determines the first aspect ratio of the spot according to the outer contour vertices of the spot; and determines the first area of ​​the spot according to the projection coordinates of the outer contour vertices; the electronic device retains the spots corresponding to the first aspect ratio that meet the first screening condition and the first area that meet the second screening condition among the plurality of spots, to obtain a first image containing launch pad features; wherein the first screening condition and the second screening condition are determined based on preset launch pad features in expert knowledge; the electronic device uses a multi-scale mask classification framework to extract and enhance features of the first image to determine the launch pad image.

[0079] The first screening condition is used to characterize the range of aspect ratios of the launch pad; the second screening condition is used to characterize the area size of the launch pad.

[0080] In the process of determining the launch pad image, the electronic device can determine the first aspect ratio and the first area corresponding to each of the multiple spots. Specifically, for each spot, the first aspect ratio of the spot can be determined according to the outer contour vertices of the spot; then the projection coordinates of the outer contour vertices are determined, and the first area of ​​the spot is determined according to the projection coordinates; then, all the spots are screened and filtered according to all the first aspect ratios and all the first areas. At this time, the spots corresponding to the first aspect ratio that meets the first filtering condition and the first area that meets the second filtering condition among the multiple spots can be retained to obtain a first image containing launch pad features, and the spots corresponding to the first aspect ratio that does not meet the first filtering condition and / or the first area that does not meet the second filtering condition are eliminated; then, a multi-scale mask classification framework is used to extract and enhance features of the first image, weaken the roofs of buildings and hardened ground that are easily confused with the launch pad, so as to obtain a launch pad image with higher accuracy.

[0081] It should be noted that, for the kth pattern spot among the multiple pattern spots, there is no limitation on the timing of the electronic device determining the first aspect ratio of the kth pattern spot and the electronic device determining the first area of ​​the kth pattern spot.

[0082] In addition, the projection coordinates of the outer contour vertices of the kth spot can be used as {A1(x1,y1),A2(x2,y2),…,A i (x i ,y i ),…,A n (x n ,y n )} means that n≥3, and n represents the total number of projection coordinates.

[0083] Optionally, the first area of ​​the kth spot may be calculated using the area calculation formula:

[0084] Calculated.

[0085] in, Indicates the first area of ​​the kth spot, in square meters; Represents the first constant, usually taking the value as 1 / 2; , .

[0086] Optionally, the first screening condition is a first aspect ratio threshold. Exemplarily, the value of the first aspect ratio threshold is 1.67.

[0087] The first aspect ratio that meets the first screening condition refers to a first aspect ratio that is less than or equal to the first aspect ratio threshold.

[0088] Optionally, the second screening condition is a first preset area range. Exemplarily, the value of the first preset area range is [350, 1100].

[0089] The first area satisfying the second screening condition refers to the first area within the first preset area range.

[0090] In some embodiments, the electronic device determines a first aspect ratio of the pattern spot based on the outer contour vertices of the pattern spot, which may include: the electronic device determines a center point of the pattern spot based on the outer contour vertices of the pattern spot; the electronic device rotates the pattern spot around the center point multiple times to obtain multiple rotated patterns, and determines envelope rectangles corresponding to each of the multiple rotated patterns; the electronic device determines a circumscribed rectangle based on a target envelope rectangle with the smallest area among the multiple envelope rectangles; the electronic device determines the first aspect ratio of the pattern spot based on the circumscribed rectangle.

[0091] In the process of determining the first aspect ratio of the kth spot, the electronic device can first determine the outer contour vertex of the spot, and then determine the center point of the spot; then rotate the spot around the center point multiple times to obtain multiple rotated spots, that is, the number of rotated spots is the same as the number of rotations; then determine the envelope rectangles corresponding to the multiple rotated spots, that is, the number of envelope rectangles is also the same as the number of rotations; then determine the area of ​​all envelope rectangles, and determine the minimum area from all areas, and then determine the target envelope rectangle corresponding to the minimum area from all envelope rectangles; then, the electronic device determines the circumscribed rectangle corresponding to the spot according to the target envelope rectangle, and then determines the first aspect ratio of the spot. The whole process determines the outer contour vertex of the spot and finds the center point, and then rotates multiple times to explore the shape of the spot in different directions, so that the most compact form of the spot under different viewing angles can be captured. In addition, by selecting the minimum area in the envelope rectangle, it is ensured that the calculated aspect ratio best represents the actual shape of the spot, thereby improving the accuracy of the aspect ratio calculation.

[0092] It should be noted that the outer contour vertices of the kth spot can be {B1(x1,y1),B2(x2,y2),…,B i (x i ,y i ),…,B n (x n ,y n )}. The number of vertices in the outer contour vertex is the same as the coordinate data in the above-mentioned projection coordinates.

[0093] Optionally, the center point of the kth spot can use the center point formula: , Calculated.

[0094] in, The horizontal coordinate of the center point; Indicates the vertical coordinate of the center point.

[0095] Optionally, the electronic device rotates the pattern spot multiple times around the center point to obtain multiple rotated pattern spots, which may include: the electronic device rotates the pattern spot multiple times around the center point according to a preset angle to obtain multiple rotated pattern spots.

[0096] Among them, any vertex in the rotated spot corresponding to the kth spot can be rotated using the rotation formula: b =(x a -x c )×cos(θ)-(y a -y c )×sin(θ)+x c ,y b =(xa -x c )×sin(θ)-(y a -y c )×cos(θ)+y c Calculated.

[0097] (x a ,y a ) represents the vertex coordinates of any vertex in the outer contour; (x b ,y b ) represents the vertex coordinates corresponding to the rotated pattern, that is, the vertex coordinates of any vertex in the outer contour after rotation; (x c ,y c ) represents the coordinates of the center point; θ represents the preset angle.

[0098] At this time, the outer contour vertices of the rotated spot corresponding to the kth spot can be {C1(x1,y1),C2(x2,y2),…,C i (x i ,y i ),…,C n (x n ,y n )}express.

[0099] For example, when the number of rotations is 180, for the kth pattern spot, the electronic device may determine 180 rotated pattern spots, and then determine 180 envelope rectangles to obtain a rectangle set {E1, E2, ..., E i ,…,E 180}, E i Represents the i-th envelope rectangle among the 180 envelope rectangles.

[0100] Combine {C1(x1,y1),C2(x2,y2),…,C i (x i ,y i ),…,C n (x n ,y n )}, the vertex coordinates of the envelope rectangle corresponding to the rotated spot obtained by rotating the kth spot once are:

[0101] D1(Min(x1,x2,…,x i ,…,x n ),Min(y1,y2,…,y i ,…,y n ));

[0102] D2(Max(x1,x2,…,x i ,…,x n),Min(y1,y2,…,y i ,…,y n ));

[0103] D3(Max(x1,x2,…,x i ,…,x n ),Max(y1,y2,…,y i ,…,y n ));

[0104] D4(Min(x1,x2,…,x i ,…,x n ),Max(y1,y2,…,y i ,…,y n )).

[0105] Among them, D1 represents the coordinates of the first vertex among the four vertices; D2 represents the coordinates of the second vertex among the four vertices; D3 represents the coordinates of the third vertex among the four vertices; and D4 represents the coordinates of the fourth vertex among the four vertices.

[0106] It should be noted that, since the electronic device has determined the four vertices of each envelope rectangle, for each envelope rectangle, the electronic device accurately determines the area corresponding to the envelope rectangle according to the four vertices of the envelope rectangle. When the number of envelope rectangles is 180, the number of areas is also 180.

[0107] Optionally, the electronic device determines the circumscribed rectangle according to the target envelope rectangle with the smallest area among multiple envelope rectangles, which may include: the electronic device determines the center point of the target envelope matrix, and rotates the target envelope matrix in reverse about the center point by a preset angle to obtain the circumscribed rectangle.

[0108] This center point-based rotation method avoids the offset or distortion of the target envelope matrix during the rotation process, thereby ensuring the accuracy of the rotated target envelope matrix, that is, ensuring the accuracy of the circumscribed rectangle.

[0109] Optionally, the electronic device determines the first aspect ratio of the image spot based on the circumscribed rectangle, which may include: the electronic device determines the two-dimensional size of the circumscribed rectangle based on four vertices of the circumscribed rectangle, the two-dimensional size including: a first distance and a second distance; the electronic device determines the length and width of the circumscribed rectangle based on the first distance and the second distance; the electronic device determines the first aspect ratio of the image spot based on the length and the width.

[0110] Among them, the four vertices of the circumscribed rectangle can be represented by D'1(x1,y1), D'2(x2,y2), D'1(x3,y3) and D'2(x4,y4).

[0111] Based on this, the first distance can be calculated using D'1(x1, y1) and D'2(x2, y2) and represented by d1; the second distance can be calculated using D'2(x2, y2) and D'1(x3, y3) and represented by d2.

[0112] Taking the first distance d1 as an example, the first distance d1 can be calculated using the distance formula:

[0113] Calculated.

[0114] At this time, the first aspect ratio of the kth spot can be calculated using the ratio formula: R1=Max(d1,d2) / Min(d1,d2).

[0115] Wherein, R1 represents the first aspect ratio of the kth spot.

[0116] In some embodiments, a multi-scale mask classification framework may include: sparse convolution, and an optimal layer combination in a neural architecture search (NAS) model; an electronic device uses a multi-scale mask classification framework to extract and enhance features of a first image to determine a launch plateau image, which may include: an electronic device uses sparse convolution to perform feature encoding on the first image to obtain a second image; the electronic device uses an optimal layer combination to extract and enhance features on the first image to obtain a third image; and the electronic device fuses the second image and the third image to obtain a launch plateau image.

[0117] Among them, sparse convolution only performs convolution operations on non-zero (or active) data when processing sparse data (such as the first image), thereby significantly reducing the amount of calculation, which is crucial to improving data processing speed and reducing energy consumption.

[0118] The optimal layer combination in the NAS model enables the NAS model to automatically find the optimal neural network architecture for a given task (such as extracting and enhancing launch field features), which can effectively reduce the complexity and time cost of manual design.

[0119] In the process of determining the launch field flat image, the electronic device can use sparse convolution to perform feature encoding on the first image to quickly and accurately obtain the second image; and use the optimal layer combination in the NAS model to perform feature extraction and feature enhancement on the first image again to obtain a more effective feature representation and enhance feature competition between channels, that is, to obtain a third image with higher accuracy, and then fuse it with the above-mentioned second image to obtain a launch field flat image with higher accuracy, thereby realizing accurate extraction of the launch field flat features.

[0120] It should be noted that the timing of the electronic device determining the second image and the electronic device determining the third image is not limited.

[0121] Optionally, before inputting the first image into the multi-scale mask classification framework, the method may further include: the electronic device scaling the first image to obtain a scaled first image.

[0122] Exemplarily, the size of the scaled first image is 256×256.

[0123] The whole process mainly uses the existing transfer weights to initialize the parameters of the designed multi-scale mask classification framework to improve the network's feature representation of the missile launcher target.

[0124] In some embodiments, the electronic device determines the vehicle status image based on the emission field flat image and the segmentation threshold, combined with the pattern shape characteristics, which may include: the electronic device determines the grayscale image of the emission field flat image, and the grayscale values ​​corresponding to all pixels in the grayscale image; the electronic device determines the grayscale mean and the grayscale standard deviation based on all the grayscale values; the electronic device determines the segmentation threshold based on the grayscale mean and the grayscale standard deviation; the electronic device retains the pixels corresponding to the pixel values ​​less than the segmentation threshold among all the pixels, and obtains the vehicle status image in combination with the pattern shape characteristics.

[0125] In the process of determining the vehicle status image, the electronic device may first determine the grayscale histogram (abbreviated as grayscale image) of the emission field image (e.g. Figure 5 As shown in FIG. 1 ), the grayscale image represents the number of pixels contained in each grayscale level in the launch pad image. Specifically, the grayscale image can be generated by counting the grayscale value of each pixel in the launch pad image. Since the launch pad in the missile position is usually a cement floor with a high spectral grayscale value, while the grayscale value of the missile launch vehicle and the shadow of the missile launch vehicle is low, the segmentation threshold can be determined according to the grayscale values ​​corresponding to all pixels in the grayscale image in combination with the grayscale image features of the missile launch vehicle and the pad background in the launch pad. Specifically, the electronic device determines the grayscale mean and grayscale standard deviation according to all grayscale values, and then determines the segmentation threshold. In order to extract the vehicle state feature, the electronic device can perform pixel screening on the grayscale image according to the segmentation threshold, retain the pixels corresponding to the pixel values ​​less than the segmentation threshold among all pixels, obtain the segmentation result, and then perform vehicle state recognition on the segmentation result in combination with the pattern shape feature to obtain a vehicle state image with high accuracy.

[0126] It should be noted that the i-th grayscale value among the m grayscale values ​​corresponding to the grayscale image can be represented by Pi, m≥2, and m represents the total number of grayscale values.

[0127] Optionally, the grayscale mean can be calculated using the first formula: Calculated.

[0128] Optionally, the grayscale standard deviation can be calculated using the second formula: Calculated.

[0129] Optionally, the segmentation threshold may be calculated using the third formula: ST=M+b×SD.

[0130] Among them, M represents the grayscale mean; represents the grayscale standard deviation; ST represents the segmentation threshold; b represents the second constant, which is usually 1 / 2.

[0131] It should be noted that the above process uses the global thresholding method (GlobalThresholding) in the binary segmentation method to segment the image within the extracted launch pad range (i.e., the launch pad image). The global thresholding method is usually used to segment the launch pad image into two parts: foreground and background. It is mainly based on the characteristics of the image grayscale histogram and selects a global threshold (i.e., the segmentation threshold) to divide the pixels in the launch pad image into two categories. Through binary segmentation, the missile launch vehicle and the missile launch vehicle shadow can be segmented as a whole, so that the shadow is also used as part of the feature. Through the overall change characteristics of the missile launch vehicle and the shadow, under the guidance of expert knowledge, the appropriate shape parameters are selected to determine the vehicle status.

[0132] In some embodiments, the electronic device determines the target recognition result corresponding to the target remote sensing image based on the launch pad image and the vehicle status image, which may include: the electronic device determines the shape and configuration of the launch pad based on the launch pad image; the electronic device determines the mounting method of the vehicle status based on the vehicle status image; the electronic device determines the target recognition result corresponding to the target remote sensing image based on the shape, configuration and mounting method.

[0133] After determining the launch pad image and the vehicle status image, the electronic device can analyze the launch pad image to obtain the shape and configuration of the launch pad; and analyze the vehicle status image to obtain the mounting method of the vehicle status; and then determine the target recognition result corresponding to the target remote sensing image based on all the data, namely the shape, configuration and mounting method.

[0134] That is to say, the above target identification result is a multi-level target intelligence semantics, which may include: the location and range of the target; the location, range and quantity of key components of the target; and the location and status of high-value sub-targets.

[0135] In some embodiments, the electronic device determines the mounting mode of the vehicle status based on the vehicle status image, which may include: when the second area of ​​the vehicle status image satisfies the third filtering condition, the electronic device determines the second aspect ratio of the vehicle status image; wherein the third filtering condition is determined based on preset vehicle status characteristics in expert knowledge; when the second area satisfies the fourth filtering condition, or the second aspect ratio satisfies the fifth filtering condition, the electronic device determines that there is a tractor mounted as the mounting mode of the vehicle status; wherein the fourth filtering condition and the fifth filtering condition are determined based on preset vehicle status characteristics; when the second area does not satisfy the fourth filtering condition, and the second aspect ratio does not satisfy the fifth filtering condition, the electronic device determines that there is no tractor mounted as the mounting mode.

[0136] Among them, the third screening condition is used to characterize the area size of the missile launcher; the fourth screening condition is used to characterize the area size of the missile launcher; and the fifth screening condition is used to characterize the length-to-width ratio range of the missile launcher.

[0137] In the process of determining the mounting mode of the vehicle state, the electronic device can first extract the characteristics of the missile launcher vehicle based on the third filtering condition. Specifically, the second area of ​​the vehicle state image is first determined, and when the second area meets the third filtering condition, it means that the size of the missile launcher in the vehicle state image meets the preset vehicle state characteristics. At this time, the vehicle state of the missile launcher can be judged. Specifically, the second aspect ratio corresponding to the vehicle state image is determined: if the second area meets the fourth filtering condition and the second aspect ratio meets the fifth filtering condition, then the mounting mode of the vehicle state is determined as having a tractor mounted; if the second area does not meet the fourth filtering condition or the second aspect ratio does not meet the fifth filtering condition, then the mounting mode is determined as having no tractor mounted.

[0138] Optionally, the second area may also be calculated using the above area calculation formula.

[0139] Optionally, a process for the electronic device to determine the second aspect ratio is similar to a process for the electronic device to determine the first aspect ratio, and is not described in detail herein.

[0140] Optionally, the third screening condition is a second preset area range. Exemplarily, the value of the second preset area range is [30, 120].

[0141] The second area satisfying the third screening condition means that the second area is within a second preset area range.

[0142] Optionally, the fourth screening condition is an area threshold. Exemplarily, the value of the area threshold is 80.

[0143] Among them, the second area meets the fourth screening condition means that the second area is greater than or equal to the above area threshold; the second area does not meet the fourth screening condition means that the second area is smaller than the above area threshold.

[0144] Optionally, the fifth screening condition is a second aspect ratio threshold. Exemplarily, the value of the second aspect ratio threshold is 0.3.

[0145] The second aspect ratio satisfies the fifth screening condition, which means that the second aspect ratio is less than or equal to the second aspect ratio threshold; the second aspect ratio does not satisfy the fifth screening condition, which means that the second aspect ratio is greater than the second aspect ratio threshold.

[0146] For example, the mounting method of the vehicle state can be determined by the following formula:

[0147] Sure.

[0148] in, Indicates the mounting mode; represents the second area; Indicates the secondary aspect ratio.

[0149] The target recognition device based on remote sensing images provided by the present invention is described below. The target recognition device based on remote sensing images described below and the target recognition method based on remote sensing images described above can be referred to each other.

[0150] Figure 6 is a schematic diagram of the structure of the target recognition device based on remote sensing images provided by the present invention, such as Figure 6 As shown, the device includes the following image determination module 601 and target recognition module 602.

[0151] The image determination module 601 is used to determine a plurality of segmentation masks according to a mask image corresponding to a target remote sensing image; and determine, from the mask image, the spots corresponding to each of the plurality of segmentation masks; screen the plurality of spots according to the first aspect ratio and the first area corresponding to each of the plurality of spots to determine a launch pad image; and determine a vehicle status image according to the launch pad image and the segmentation threshold value in combination with the spot shape characteristics;

[0152] The target recognition module 602 is used to determine the target recognition result corresponding to the target remote sensing image according to the launch site image and the vehicle status image.

[0153] Optionally, the image determination module 601 is specifically used to determine, for each spot, a first aspect ratio of the spot based on the outer contour vertices of the spot; and determine a first area of ​​the spot based on the projection coordinates of the outer contour vertices; retain the spots among the multiple spots that correspond to the first aspect ratio that satisfies the first filtering condition and the first area that satisfies the second filtering condition, to obtain a first image containing launch field features; wherein the first filtering condition and the second filtering condition are determined based on preset launch field features in expert knowledge; and use a multi-scale mask classification framework to extract and enhance features of the first image to determine the launch field image.

[0154] Optionally, the image determination module 601 is specifically used to determine the grayscale image of the emission field flat image and the grayscale values ​​corresponding to all pixels in the grayscale image; determine the grayscale mean and the grayscale standard deviation based on all the grayscale values; determine the segmentation threshold based on the grayscale mean and the grayscale standard deviation; retain the pixels corresponding to the pixel values ​​less than the segmentation threshold among all the pixels, and obtain the vehicle status image in combination with the pattern shape characteristics.

[0155] Optionally, the target recognition module 602 is specifically used to determine the shape and configuration of the launch pad based on the launch pad image; determine the mounting mode of the vehicle status based on the vehicle status image; and determine the target recognition result corresponding to the target remote sensing image based on the shape, the configuration and the mounting mode.

[0156] Optionally, the image determination module 601 is specifically used to determine the center point of the spot based on the outer contour vertices of the spot; rotate the spot around the center point multiple times to obtain multiple rotated spots, and determine the envelope rectangles corresponding to each of the multiple rotated spots; determine the circumscribed rectangle based on the target envelope rectangle with the smallest area among the multiple envelope rectangles; and determine the first aspect ratio of the spot based on the circumscribed rectangle.

[0157] Optionally, the multi-scale mask classification framework includes: sparse convolution, and an optimal layer combination in a neural network architecture search model; an image determination module 601, specifically used to use the sparse convolution to perform feature encoding on the first image to obtain a second image; use the optimal layer combination to perform feature extraction and enhancement on the first image to obtain a third image; and fuse the second image and the third image to obtain the emission field flat image.

[0158] Optionally, the target recognition module 602 is specifically used to determine the second aspect ratio of the vehicle status image when the second area of ​​the vehicle status image satisfies a third filtering condition; wherein the third filtering condition is determined based on preset vehicle status characteristics in expert knowledge; when the second area satisfies a fourth filtering condition, or the second aspect ratio satisfies a fifth filtering condition, determining that there is a tractor mounted as the mounting mode of the vehicle status; wherein the fourth filtering condition and the fifth filtering condition are determined based on the preset vehicle status characteristics; when the second area does not satisfy the fourth filtering condition, and the second aspect ratio does not satisfy the fifth filtering condition, determining that there is no tractor mounted as the mounting mode.

[0159] Figure 7 is a schematic diagram of the structure of the electronic device provided by the present invention, such as Figure 7 As shown, the electronic device may include: a processor 710, a communication interface 720, a memory 730 and a communication bus 740, wherein the processor 710, the communication interface 720 and the memory 730 communicate with each other through the communication bus 740. The processor 710 may call the logic instructions in the memory 730 to execute a target recognition method based on a remote sensing image, the method comprising: determining a plurality of segmentation masks according to a mask image corresponding to a target remote sensing image; and determining from the mask image the spots corresponding to each of the plurality of segmentation masks; screening the plurality of spots according to the first aspect ratio and the first area corresponding to each of the plurality of spots to determine a launch pad image; determining a vehicle status image according to the launch pad image and the segmentation threshold and in combination with the spot shape feature; and determining a target recognition result corresponding to the target remote sensing image according to the launch pad image and the vehicle status image.

[0160] In addition, the logic instructions in the above-mentioned memory 730 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present invention can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0161] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the target recognition method based on remote sensing images provided by the above-mentioned methods, and the method includes: determining multiple segmentation masks according to a mask image corresponding to the target remote sensing image; and determining from the mask image the spots corresponding to each of the multiple segmentation masks; screening the multiple spots according to the first aspect ratio and the first area corresponding to each of the multiple spots to determine a launch site image; determining a vehicle status image according to the launch site image and the segmentation threshold, combined with the spot shape characteristics; and determining a target recognition result corresponding to the target remote sensing image according to the launch site image and the vehicle status image.

[0162] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the target recognition method based on remote sensing images provided by the above-mentioned methods, the method comprising: determining multiple segmentation masks according to a mask image corresponding to the target remote sensing image; and determining, from the mask image, the spots corresponding to each of the multiple segmentation masks; screening the multiple spots according to the first aspect ratio and the first area corresponding to each of the multiple spots to determine a launch pad image; determining a vehicle status image according to the launch pad image and the segmentation threshold, combined with the spot shape characteristics; and determining a target recognition result corresponding to the target remote sensing image according to the launch pad image and the vehicle status image.

[0163] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0164] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0165] Finally, it should be noted that the above embodiments 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 the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A target recognition method based on remote sensing images, characterized in that: include: Determine a plurality of segmentation masks according to a mask image corresponding to the target remote sensing image; and determining, from the mask image, the spots corresponding to each of the plurality of segmentation masks; Screening the plurality of image spots according to first aspect ratios and first areas corresponding to the plurality of image spots to determine a launch field image; Determine the vehicle status image according to the launch site image and the segmentation threshold value in combination with the shape characteristics of the image spots; Determining a target recognition result corresponding to the target remote sensing image according to the launch site image and the vehicle status image; The step of screening the plurality of image spots according to the first aspect ratios and the first areas respectively corresponding to the plurality of image spots to determine the emission field image comprises: For each image spot, a first aspect ratio of the image spot is determined according to the outer contour vertices of the image spot; and a first area of ​​the image spot is determined according to the projection coordinates of the outer contour vertices; the first aspect ratio is the aspect ratio corresponding to the circumscribed rectangle determined according to the target envelope rectangle with the smallest area among the multiple envelope rectangles obtained by rotating the image spot around the center point multiple times; the center point is determined according to the outer contour vertices of the image spot; Retaining, among the plurality of image spots, image spots corresponding to the first aspect ratio that satisfies the first screening condition and the first area that satisfies the second screening condition, to obtain a first image containing launch pad features; wherein the first screening condition and the second screening condition are determined based on preset launch pad features in expert knowledge; A multi-scale mask classification framework is used to extract and enhance features of the first image to determine the launch field plateau image.

2. The target recognition method based on remote sensing images according to claim 1, characterized in that: The determining of the vehicle status image according to the emission field image and the segmentation threshold in combination with the spot shape feature comprises: Determine a grayscale image of the emission field plateau image and grayscale values ​​corresponding to all pixels in the grayscale image; According to all gray values, determine the gray mean and gray standard deviation; Determining the segmentation threshold according to the grayscale mean and the grayscale standard deviation; The pixels corresponding to the pixel values ​​less than the segmentation threshold value among all the pixels are retained, and the vehicle status image is obtained by combining the pattern shape features.

3. The target recognition method based on remote sensing images according to claim 1 or 2, characterized in that: The step of determining a target recognition result corresponding to the target remote sensing image according to the launch site image and the vehicle status image includes: Determining the shape and configuration of the launch pad according to the launch pad image; Determining a mounting mode of the vehicle state according to the vehicle state image; A target recognition result corresponding to the target remote sensing image is determined according to the shape, the configuration mode and the mounting mode.

4. The target recognition method based on remote sensing images according to claim 1, characterized in that: The determining, according to the outer contour vertices of the pattern spot, a first aspect ratio of the pattern spot comprises: Determine the center point of the image spot according to the outer contour vertices of the image spot; Rotate the pattern spot multiple times around the center point to obtain multiple rotated pattern spots, and determine envelope rectangles corresponding to each of the multiple rotated pattern spots; Determine a circumscribed rectangle according to a target envelope rectangle with the smallest area among multiple envelope rectangles; A first aspect ratio of the image spot is determined according to the circumscribed rectangle.

5. The target recognition method based on remote sensing images according to claim 1, characterized in that: The multi-scale mask classification framework includes: sparse convolution, and an optimal layer combination in a neural network architecture search model; the multi-scale mask classification framework is used to extract and enhance features of the first image to determine the launch field image, including: Using the sparse convolution, feature encoding is performed on the first image to obtain a second image; Using the optimal layer combination, extracting and enhancing features of the first image to obtain a third image; The second image and the third image are fused to obtain the launch site image.

6. The target recognition method based on remote sensing images according to claim 3 is characterized in that: The step of determining the mounting mode of the vehicle state according to the vehicle state image includes: When the second area of ​​the vehicle state image satisfies a third screening condition, determining a second aspect ratio of the vehicle state image; wherein the third screening condition is determined based on a preset vehicle state feature in expert knowledge; When the second area satisfies the fourth screening condition, or the second aspect ratio satisfies the fifth screening condition, the tractor mounting is determined as the mounting mode of the vehicle state; wherein the fourth screening condition and the fifth screening condition are determined based on the preset vehicle state characteristics; When the second area does not satisfy the fourth screening condition and the second aspect ratio does not satisfy the fifth screening condition, no tractor mounting is determined as the mounting mode.

7. A device for implementing the target recognition method based on remote sensing images as claimed in any one of claims 1 to 6, characterized in that: include: An image determination module, used for determining a plurality of segmentation masks according to a mask image corresponding to a target remote sensing image; and determining, from the mask image, the spots corresponding to the multiple segmentation masks; screening the multiple spots according to the first aspect ratios and the first areas corresponding to the multiple spots to determine a launch pad image; and determining a vehicle status image according to the launch pad image and the segmentation threshold and in combination with the spot shape features; The target recognition module is used to determine the target recognition result corresponding to the target remote sensing image according to the launch site image and the vehicle status image.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the target recognition method based on remote sensing images as described in any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the target recognition method based on remote sensing images as claimed in any one of claims 1 to 6 is implemented.

Images