Unmanned platform small target detection method, device and equipment based on rank sparse matrix

Abstract

The invention relates to an unmanned platform small target detection method, device and equipment based on a rank sparse matrix. The method comprises the steps of obtaining a target detection image of an environment domain where the unmanned platform is located; and constructing an image detection model based on the rank sparse matrix. The image detection model adopts robust principal component analysis to separate a foreground and a background of a target detection image to obtain a foreground target image and a background image. And screening a candidate region from the target detection image, and marking the candidate region in the target detection image to obtain a region detection feature map. And inputting the foreground target image as prior information and the region detection feature map into an image detection model for feature extraction to obtain a small target feature map. And training an image detection model according to the small target feature map to obtain a trained image detection model, and performing small target detection through the trained image detection model. By adopting the method, the small target detection precision can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of small target detection on unmanned platforms, and in particular to a method, device and equipment for detecting small targets on unmanned platforms based on a rank sparse matrix. Background Art

[0002] In the environment of unmanned platforms, small target detection has always been a challenge in the field of computer vision. Although existing technologies have made some progress in this regard, they have obvious limitations in solving background interference and extracting small target features, making it difficult to achieve the expected detection accuracy.

[0003] Traditional background subtraction technology mainly extracts foreground objects by modeling and subtracting static backgrounds. However, this method is prone to failure in dynamic environments, causing detection errors when the background changes or the illumination changes. In addition, background subtraction technology is less effective for detecting small objects, because small objects are easily mistaken for background noise and thus filtered out.

[0004] Many object detection algorithms use global feature extraction methods, such as Histogram of Oriented Gradients (HOG) and Scale-Invariant Feature Transform (SIFT). These methods perform well in large object detection, but for small objects, the features are often submerged in the global features and difficult to extract and distinguish effectively. Therefore, the detection accuracy of small objects is low.

[0005] Modern object detection methods make extensive use of convolutional neural networks (CNNs) such as Faster R-CNN, YOLO, and SSD. These methods perform well in large object detection, but have certain limitations in small object detection. The reason is that CNNs usually perform detection at a higher feature level, which results in low resolution feature maps of small objects and makes it difficult to capture enough details. In addition, CNN models are easily disturbed by background noise during training, resulting in small object features being ignored.

[0006] Principal component analysis (PCA) is often used for dimensionality reduction and feature extraction, but it is sensitive to noise and outliers. In complex unmanned platform environments, the changes in background and foreground are complex, and traditional PCA cannot effectively separate foreground objects and backgrounds, resulting in poor feature extraction results, which in turn affects the detection accuracy of small targets.

[0007] The above-mentioned existing technologies are difficult to effectively remove background interference in dynamic and complex backgrounds, making it difficult for foreground targets to stand out. Especially in unmanned platform environments, the background changes frequently and traditional background subtraction methods are difficult to adapt. Global feature extraction methods cannot effectively capture the details of small targets, resulting in the lack of obvious features of small targets in feature maps, making it difficult to accurately detect them. Summary of the invention

[0008] Based on this, it is necessary to provide an unmanned platform small target detection method, device and equipment based on rank sparse matrix, which can improve the unmanned platform's small target detection accuracy and efficiency.

[0009] A method for detecting small targets on an unmanned platform based on a rank sparse matrix, the method comprising:

[0010] Obtain target detection images of the environment where the unmanned platform is located.

[0011] Construct an image detection model based on rank sparse matrix.

[0012] The image detection model uses robust principal component analysis to separate the foreground and background of the target detection image into a foreground target image and a background image.

[0013] The candidate regions are screened from the target detection image, and the candidate regions are marked in the target detection image to obtain a region detection feature map.

[0014] The foreground target image is used as prior information and region detection features Figure 1 And input it into the image detection model for feature extraction to obtain the small target feature map.

[0015] The image detection model is trained according to the small target feature map to obtain a trained image detection model, and small target detection is performed using the trained image detection model.

[0016] In one embodiment, the method further includes: obtaining a target detection image in the current environment domain through a small optical module mounted on the unmanned platform:

[0017] f D (x,y)=f T (x,y)+f B (x,y)+f N (x,y)

[0018] Among them, f D (x,y) is the target detection image, f T (x,y) is a small target, f B (x, y) is the background image, f N(x, y) is noise, and (x, y) is the position of the pixel. The small optical module includes: a front downward infrared camera, a downward infrared camera, a front downward visible light camera, a downward visible light camera, and a three-axis servo linkage motor. The three-axis servo linkage motor includes: a pitch servo motor, an azimuth motor, and a roll motor.

[0019] In one of the embodiments, it also includes: constructing a foreground target matrix of small target pixels in the target detection image based on a rank sparse matrix, and constructing an image detection model according to the foreground target matrix and the low-rank background matrix.

[0020] In one embodiment, the image detection model further includes: using robust principal component analysis to separate the foreground and background of the target detection image, and if the binarization of the target detection image is greater than a preset threshold, the foreground target image is separated. If the binarization of the target detection image is less than the preset threshold, the background image is separated.

[0021] In one of the embodiments, it also includes: using a region candidate network to screen candidate regions from the target detection image to obtain a candidate feature map, labeling the candidate regions according to the positional relationship between the candidate regions and the real frame of the object in the candidate feature map, and outputting the region detection feature map.

[0022] In one of the embodiments, it also includes: marking the relative position and size of the candidate area and the real frame of the object in the candidate feature map to obtain the candidate label value, inputting the candidate label value and the candidate area into the deep learning network for feature prediction, and outputting the area detection feature map corresponding to the predicted value.

[0023] In one embodiment, the method further includes: using a machine learning algorithm and standard sample data to pre-train an image detection model, using the foreground target image as prior information and regional detection features Figure 1 The image is then input into a pre-trained image detection model to extract the target position and pixel size of small target pixels in the foreground target image. The region detection feature map is extracted in the deep learning network based on the target position, pixel size, and preset feature extraction rules to obtain a small target feature map.

[0024] In one of the embodiments, it also includes: constructing a loss function based on the comparison result of the predicted value and the candidate label value, performing interactive training of the target unmanned platform and the unmanned platform on the image detection model based on the loss function and the small target feature map to obtain a trained image detection model, and performing small target detection through the trained image detection model.

[0025] A small target detection device for an unmanned platform based on a rank sparse matrix, the device comprising:

[0026] The target detection image acquisition module is used to obtain the target detection image of the environment domain where the unmanned platform is located.

[0027] Image detection model building module, used to build image detection model based on rank sparse matrix.

[0028] The image separation module is used for the image detection model to separate the foreground and background of the target detection image into a foreground target image and a background image by using robust principal component analysis.

[0029] The region detection feature map extraction module is used to screen candidate regions from the target detection image and mark the candidate regions in the target detection image to obtain a region detection feature map.

[0030] Small target feature map acquisition module is used to use the foreground target image as prior information and regional detection features Figure 1 And input it into the image detection model for feature extraction to obtain the small target feature map.

[0031] The detection module is used to train the image detection model according to the small target feature map to obtain a trained image detection model, and perform small target detection through the trained image detection model.

[0032] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0033] Obtain target detection images of the environment where the unmanned platform is located.

[0034] Construct an image detection model based on rank sparse matrix.

[0035] The image detection model uses robust principal component analysis to separate the foreground and background of the target detection image into a foreground target image and a background image.

[0036] The candidate regions are screened from the target detection image, and the candidate regions are marked in the target detection image to obtain a region detection feature map.

[0037] The foreground target image is used as prior information and region detection features Figure 1 And input it into the image detection model for feature extraction to obtain the small target feature map.

[0038] The image detection model is trained according to the small target feature map to obtain a trained image detection model, and small target detection is performed using the trained image detection model.

[0039] The above-mentioned unmanned platform small target detection method, device and equipment based on rank sparse matrix separate the foreground and background of the target detection image through robust principal component analysis (RPCA) and rank sparse matrix technology to reduce background interference and enhance the effect of foreground target feature extraction. First, the unmanned platform obtains the target detection image in the environment domain. Then, an image detection model is constructed based on the rank sparse matrix, and the redundant information and sparse features in the image are decomposed by the rank sparse matrix. The image is decomposed into low-rank background and sparse foreground through RPCA, and the foreground and background are effectively separated. Next, the foreground target image is used as prior information and combined with the regional detection feature map as a training sample data set to input into the image detection model for feature extraction, and the small target feature map is output, so that the model can focus on learning the features of the regional detection map and ignore the background noise. According to the positional relationship between the candidate area and the real frame of the object in the regional detection map, the candidate area is labeled to ensure that the model obtains accurate labeling information during the training process. Accurate labeling helps to improve the model training effect and detection accuracy. Then, the small target feature map is output to indicate the position and features of the small target in the image. The image detection model is trained using the feature map to optimize the model parameters. Finally, small target detection is performed through the trained image detection model, thereby ensuring that the model can accurately identify and detect small targets in the image, thereby improving detection accuracy and speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 The figure is a flow chart of a method for detecting small targets on an unmanned platform based on a rank sparse matrix in one embodiment;

[0041] Figure 2 It is a flowchart of a small target recognition step of an unmanned platform based on a rank sparse matrix in one embodiment;

[0042] Figure 3 It is a schematic block diagram of the internal structure of the target autonomous recognition algorithm in one embodiment;

[0043] Figure 4 A schematic diagram of a training process of an image detection model in one embodiment;

[0044] Figure 5 is a target image detection result in one embodiment;

[0045] Figure 6 It is a structural block diagram of a small target detection device for an unmanned platform based on a rank sparse matrix in one embodiment;

[0046] Figure 7 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0048] The unmanned platform small target detection method based on rank sparse matrix provided by the present invention can be applied to an unmanned platform equipped with a small optical module, wherein the small optical module includes: a front downward infrared camera, a downward infrared camera, a front downward visible light camera, a downward visible light camera and a three-axis servo linkage motor. The three-axis servo linkage motor includes: a pitch servo motor, an azimuth motor and a roll motor.

[0049] In one embodiment, Figure 1 As shown, a method for detecting small targets on an unmanned platform based on a rank sparse matrix is ​​provided. The method is applied to the above unmanned platform as an example for explanation, and includes the following steps:

[0050] Step 102: Acquire a target detection image of the environment domain where the unmanned platform is located.

[0051] Step 104, constructing an image detection model based on the rank sparse matrix.

[0052] Step 106: The image detection model uses robust principal component analysis to separate the foreground and background of the target detection image into a foreground target image and a background image.

[0053] Step 108, screening candidate regions from the target detection image, and marking the candidate regions in the target detection image to obtain a region detection feature map.

[0054] Step 110: Use the foreground target image as prior information and region detection features Figure 1 And input it into the image detection model for feature extraction to obtain the small target feature map.

[0055] Step 112, training the image detection model according to the small target feature map to obtain a trained image detection model, and performing small target detection using the trained image detection model.

[0056] In the above-mentioned unmanned platform small target detection method based on rank sparse matrix, the foreground and background of the target detection image are separated by robust principal component analysis (RPCA) and rank sparse matrix technology to reduce background interference and enhance the effect of foreground target feature extraction. First, the unmanned platform obtains the target detection image in the environment domain. Then, the image detection model is constructed based on the rank sparse matrix, and the redundant information and sparse features in the image are decomposed by the rank sparse matrix. The image is decomposed into low-rank background and sparse foreground by RPCA, and the foreground and background are effectively separated. Next, the foreground target image is used as prior information and combined with the region detection feature map as a training sample data set to input into the image detection model for feature extraction, and the small target feature map is output, so that the model can focus on learning the features of the region detection map and ignore the background noise. According to the positional relationship between the candidate region and the real frame of the object in the region detection map, the candidate region is labeled to ensure that the model obtains accurate labeling information during the training process. Accurate labeling helps to improve the model training effect and detection accuracy. Then, the small target feature map is output to indicate the position and features of the small target in the image. The image detection model is trained using the feature map to optimize the model parameters. Finally, small target detection is performed through the trained image detection model, thereby ensuring that the model can accurately identify and detect small targets in the image, thereby improving detection accuracy and speed.

[0057] In one embodiment, a target detection image in the current environment is obtained by a small optical module mounted on an unmanned platform:

[0058] f D (x,y)=f T (x,y)+f B (x,y)+f N (x,y)

[0059] Among them, f D (x,y) is the target detection image, f T (x,y) is a small target, f B (x, y) is the background image, f N (x, y) is noise, and (x, y) is the position of the pixel. The small optical module includes: a front downward infrared camera, a downward infrared camera, a front downward visible light camera, a downward visible light camera, and a three-axis servo linkage motor. The three-axis servo linkage motor includes: a pitch servo motor, an azimuth motor, and a roll motor.

[0060] In one of the embodiments, a foreground target matrix of small target pixels in a target detection image is constructed based on a rank sparse matrix, and an image detection model is constructed based on the foreground target matrix and a low-rank background matrix.

[0061] In one embodiment, the image detection model uses robust principal component analysis to separate the foreground and background of the target detection image, and if the binarization of the target detection image is greater than a preset threshold, the foreground target image is separated. If the binarization of the target detection image is less than the preset threshold, the background image is separated.

[0062] It is worth noting that the target detection image obtained by using a small optical module is an image composed of a mixture of infrared images and visible light of different wavelengths. The image detection model uses Robust Principal Component Analysis (RPCA) to separate the features of image pixels and convert the target separation process into matrix mathematical operations, so that the model parameters do not need to be estimated according to different images and do not affect the calculation of the decomposition results.

[0063] Specifically, the component analysis of infrared images can be expressed as:

[0064] Y=B+T+N

[0065] Among them, B is the low-rank background matrix, T is the sparse target matrix, and N is the noise matrix. Based on this, it is a class of Robust Principal Component Analysis (RPCA) problem:

[0066]

[0067] stY=B+T+N

[0068] Among them, rank(B) is, μ is the penalty coefficient, and λ is the weighting constant. The common solution is to transform it into a convex optimization problem, but this will lead to a suboptimal solution. This application designs a new non-convex function to approximate the matrix rank norm constraint and the L0 norm, and uses the non-convex RPCA method to obtain more accurate optimization results. Specifically, the non-convex function expression is as follows:

[0069]

[0070] Among them, f(x) is the norm value, x is the independent variable, The value determines the degree of approximation to the norm.

[0071] Then we have:

[0072]

[0073] Among them, σ i (B) is the i-th singular value of B, and Respectively represent the newly defined Rank norm and Norm. This non-convex function has the following properties: When σ i When (B) = 0 or ||T||1 = 0, f(x) = 0. Then, the original low-rank sparse model can be simplified as follows (noise is not considered):

[0074]

[0075] stY=B+T

[0076] Furthermore, the image detection model is a low-rank sparse model, and the optimization methods include gradient descent method, dual method, Lagrange multiplier method and alternating direction multiplier method. This embodiment uses the alternating direction multiplier method (ADMM) for model optimization, which decomposes the global problem into multiple simple sub-problems, and then iterates the solutions of the sub-problems for multiple rounds until the global optimal solution of the problem is obtained. Compared with other methods, ADMM has better convergence and accuracy.

[0077] Furthermore, the augmented Lagrangian function can be expressed as:

[0078]

[0079] Where M is the Lagrange multiplier matrix, μ is the penalty coefficient, and λ is the weighting constant. <·,·> represents the inner product operation, ||·|| F is the Frobenius norm. The specific solution steps are as follows:

[0080]

[0081] (3) Update M k , μ k , fixed B k , T k :

[0082] M k =M k-1 +μ k-1 (YB k -T k )

[0083] μ k =ρμ k-1

[0084] The value of each variable to be determined is continuously updated until the predetermined convergence condition is met, and then an optimized and trained stable image detection model is obtained, which is used to separate the foreground and background of the target detection image. In addition, the robustness of small target detection based on the RPCA method is not limited by the size and brightness of small targets, and can solve the problem of detecting small targets in cluttered backgrounds.

[0085] In one of the embodiments, a region candidate network is used to screen candidate regions from a target detection image to obtain a candidate feature map, the candidate regions are labeled according to the positional relationship between the candidate regions and the real frame of the object in the candidate feature map, and the region detection feature map is output.

[0086] In one of the embodiments, the relative position and size of the candidate region and the real frame of the object in the candidate feature map are marked to obtain the candidate label value, the candidate label value and the candidate region are input into the deep learning network for feature prediction, and the region detection feature map corresponding to the predicted value is output.

[0087] In one embodiment, a machine learning algorithm and standard sample data are used to pre-train the image detection model, and the foreground target image is used as prior information and regional detection features. Figure 1 The image is then input into a pre-trained image detection model to extract the target position and pixel size of small target pixels in the foreground target image. The region detection feature map is extracted in the deep learning network based on the target position, pixel size, and preset feature extraction rules to obtain a small target feature map.

[0088] In one of the embodiments, a loss function is constructed based on the comparison result of the predicted value and the candidate label value, and the image detection model is interactively trained between the target unmanned platform and the unmanned platform based on the loss function and the small target feature map to obtain a trained image detection model, and small target detection is performed using the trained image detection model.

[0089] In one embodiment, if Figure 2 As shown, a small target recognition step for an unmanned platform based on a rank sparse matrix is ​​provided:

[0090] S1. Input image: the target image of the environment where the unmanned platform is located is captured by the sensing device or recognition device;

[0091] S2, using the RPCA method to separate the foreground from the background, the map composed of the position and category of the characteristic pixel points of the detected target cluster is used as the foreground image, and the map composed of the remaining pixel points is used as the background image;

[0092] S3, judging whether the binarization value is greater than a threshold value, if so, the image composed of the current pixel points is a foreground image, otherwise, it is a background image;

[0093] S4. Extract the position and size feature information of the target pixel points of the foreground target.

[0094] In one embodiment, if Figure 3 As shown in the figure, target classification and recognition based on deep learning is adopted, DarkNet53 is used as the backbone network, and multi-scale feature fusion is used to improve the efficiency of feature learning and target recognition. Combined with the target autonomous recognition algorithm, a large number of samples are used for learning and training. The recognition rate has a great relationship with the sample size, and the closer the target is to the sample in the sample library, the more accurate the recognition. It is divided into two steps: offline learning and online recognition. The offline learning time is longer and needs to be pre-trained. The online recognition can run in real time on HI3559A.

[0095] In one embodiment, if Figure 4 As shown, a training process of an image detection model is provided. A series of candidate regions are generated on the image according to certain rules, and then the candidate regions are labeled according to the positional relationship between these candidate regions and the real frames of the objects on the image. Those candidate regions that are close enough to the real frame will be labeled as positive samples, and the position of the real frame will be used as the position target of the positive sample. Those candidate regions that deviate greatly from the real frame will be labeled as negative samples, and the negative samples do not need to predict the position or category. The network is used to extract image features and predict the position and category of the candidate region. In this way, each predicted box can be regarded as a sample, and the label value is obtained by labeling it according to its position and category relative to the real frame. Its position and category are predicted by the network model. The network prediction value and the label value are compared, and a loss function can be established to perform end-to-end training to obtain a detection model. As shown Figure 5 As shown, small object images are recognized by the detection model.

[0096] It should be understood that although Figure 1 , 2 The steps in the flowcharts of , 4 are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 , 2 At least part of the steps in 4 may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0097] In one embodiment, Figure 6As shown, a small target detection device for an unmanned platform based on a rank sparse matrix is ​​provided, comprising: a target detection image acquisition module 602, an image detection model construction module 604, an image separation module 606, a small target feature extraction module 608, a small target feature map acquisition module 610 and a detection module 612, wherein:

[0098] The target detection image acquisition module 602 is used to obtain the target detection image of the environment domain where the unmanned platform is located.

[0099] The image detection model construction module 604 is used to construct an image detection model based on a rank sparse matrix.

[0100] The image separation module 606 is used for the image detection model to separate the foreground and background of the target detection image into a foreground target image and a background image by using robust principal component analysis.

[0101] The region detection feature map extraction module 608 is used to screen candidate regions from the target detection image and mark the candidate regions in the target detection image to obtain a region detection feature map.

[0102] Small target feature map acquisition module 610 is used to use the foreground target image as prior information and the regional detection feature map Figure 1 And input it into the image detection model for feature extraction to obtain the small target feature map.

[0103] The detection module 612 is used to train the image detection model according to the small target feature map to obtain a trained image detection model, and perform small target detection using the trained image detection model.

[0104] For the specific definition of the unmanned platform small target detection device based on a rank sparse matrix, please refer to the definition of the unmanned platform small target detection method based on a rank sparse matrix above, which will not be repeated here. Each module in the above-mentioned unmanned platform small target detection device based on a rank sparse matrix can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0105] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 7As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a method for detecting small targets of an unmanned platform based on a rank sparse matrix is ​​implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a key, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.

[0106] Those skilled in the art will understand that Figure 4 , 6 The structure shown in 7 is only a block diagram of a part of the structure related to the scheme of the present invention, and does not constitute a limitation on the computer device to which the scheme of the present invention is applied. The specific computer device may include more or less components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0107] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0108] Obtain target detection images of the environment where the unmanned platform is located.

[0109] Construct an image detection model based on rank sparse matrix.

[0110] The image detection model uses robust principal component analysis to separate the foreground and background of the target detection image into a foreground target image and a background image.

[0111] The candidate regions are screened from the target detection image, and the candidate regions are marked in the target detection image to obtain a region detection feature map.

[0112] The foreground target image is used as prior information and region detection features Figure 1 And input it into the image detection model for feature extraction to obtain the small target feature map.

[0113] The image detection model is trained according to the small target feature map to obtain a trained image detection model, and small target detection is performed using the trained image detection model.

[0114] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0115] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0116] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A small target detection method for unmanned platform based on rank sparse matrix, characterized in that: Applied to an unmanned platform equipped with a small optical module, the method comprises: Obtain target detection images of the environment where the unmanned platform is located; Build an image detection model based on rank sparse matrix; The image detection model uses robust principal component analysis to separate the foreground and background of the target detection image into a foreground target image and a background image; Screening candidate regions from the target detection image, and marking the candidate regions in the target detection image to obtain a region detection feature map; Input the foreground target image as prior information together with the region detection feature map into the image detection model for feature extraction to obtain a small target feature map; The image detection model is trained according to the small target feature map to obtain a trained image detection model, and small target detection is performed using the trained image detection model.

2. The method according to claim 1, characterized in that Obtain target detection images of the environment where the unmanned platform is located, including: Obtain the target detection image in the current environment domain through the small optical module loaded on the unmanned platform: f D (x,y)=f T (x,y)+f B (x,y)+f N (x,y) Among them, f D (x,y) is the target detection image, f T (x,y) is a small target, f B (x,y) is the background image, f N (x,y) is the noise, (x,y) is the position of the pixel; The small optical module includes: a front downward-looking infrared camera, a downward-looking infrared camera, a front downward-looking visible light camera, a downward-looking visible light camera, and a three-axis servo linkage motor; The three-axis servo linkage motor includes a pitch servo motor, an azimuth motor and a roll motor.

3. The method according to claim 1, characterized in that The image detection model is constructed based on the rank sparse matrix, including: A foreground target matrix of small target pixels in the target detection image is constructed based on a rank sparse matrix, and an image detection model is constructed according to the foreground target matrix and a low-rank background matrix.

4. The method according to claim 3, characterized in that: The image detection model uses robust principal component analysis to separate the foreground and background of the target detection image into a foreground target image and a background image, including: The image detection model uses robust principal component analysis to separate the foreground and background of the target detection image. If the binarization of the target detection image is greater than a preset threshold, the foreground target image is separated; if the binarization of the target detection image is less than the preset threshold, the background image is separated.

5. The method according to claim 4, characterized in that Screening candidate regions from the target detection image and marking the candidate regions in the target detection image to obtain a region detection feature map includes: A region candidate network is used to screen candidate regions from the target detection image to obtain a candidate feature map, the candidate regions are labeled according to the positional relationship between the candidate regions and the real frame of the object in the candidate feature map, and a region detection feature map is output.

6. The method according to claim 5, characterized in that The candidate region is labeled according to the positional relationship between the candidate region and the real frame of the object in the candidate feature map, and the region detection feature map is output, including: The relative position and size of the candidate region and the real frame of the object in the candidate feature map are marked to obtain the candidate label value, the candidate label value and the candidate region are input into the deep learning network for feature prediction, and the region detection feature map corresponding to the predicted value is output.

7. The method according to claim 6, characterized in that The foreground target image is used as prior information and input into the image detection model together with the region detection feature map to extract features, thereby obtaining a small target feature map, including: The image detection model is pre-trained using a machine learning algorithm and standard sample data, and the foreground target image is used as prior information and input into the pre-trained image detection model together with the region detection feature map to extract the target position and pixel size of small target pixels in the foreground target image; The area detection feature map is subjected to feature extraction in a deep learning network according to the target position, the pixel size and a preset feature extraction rule to obtain a small target feature map.

8. The method according to claim 7, characterized in that The image detection model is trained according to the small target feature map to obtain a trained image detection model, and small target detection is performed using the trained image detection model, including: A loss function is constructed according to the comparison result between the predicted value and the candidate label value, and the image detection model is interactively trained between the target unmanned platform and the unmanned platform according to the loss function and the small target feature map to obtain a trained image detection model, and small target detection is performed using the trained image detection model.

9. A small target detection device for an unmanned platform based on a rank sparse matrix, characterized in that: The device comprises: The target detection image acquisition module is used to obtain the target detection image of the environment domain where the unmanned platform is located; An image detection model building module is used to build an image detection model based on a rank sparse matrix; An image separation module, used for the image detection model to separate the foreground and background of the target detection image into a foreground target image and a background image by using robust principal component analysis; A region detection feature map extraction module is used to screen candidate regions from the target detection image and mark the candidate regions in the target detection image to obtain a region detection feature map; A small target feature map acquisition module, used to input the foreground target image as prior information together with the region detection feature map into the image detection model for feature extraction to obtain a small target feature map; The detection module is used to train the image detection model according to the small target feature map to obtain a trained image detection model, and perform small target detection through the trained image detection model.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

Images