Target detection network detection result optimization method, device and equipment

Abstract

The application relates to a detection result optimization method, device and equipment for a target detection network. A geospatial category probability graph with different target category geospatial category relationships is first constructed, a target detection neural network is first used to detect a target in a SAR image, a target category probability and a target detection frame are obtained, a classifier weight in the target detection neural network is weighted and modulated according to the geospatial category probability graph to obtain a priori enhanced classifier weight, the target category probability is optimized according to the priori enhanced classifier weight, and the optimized target category probability and the target detection frame are used as the final target detection result. The method can improve the accuracy of the target detection result.

Description

Technical Field

[0001] This application relates to the field of SAR image target detection technology, and in particular to a method, apparatus and equipment for optimizing the detection results of target detection networks. Background Technology

[0002] With the rapid development of neural network technology in computer vision, target detection technology has achieved remarkable results, especially in the identification and localization of specific targets such as aircraft. Deep learning-based target detection algorithms not only perform well on traditional natural images but also demonstrate powerful detection capabilities in remote sensing image processing. Inspired by this, deep learning technology has been widely applied to target detection and recognition in SAR images. Researchers have designed various innovative and effective methods tailored to the characteristics of targets in SAR images, achieving good detection results.

[0003] Despite progress in intelligent aircraft target detection and recognition in SAR images, several challenges remain. First, the increasing resolution of SAR imaging places higher demands on detection efficiency and accuracy. Second, the scale and morphological diversity of different target types in SAR images increases the difficulty of extracting effective features. Finally, the significant differences in target features under different polarization modes, resolutions, and imaging geometries in SAR images further complicate target detection. Furthermore, inherent noise and interference in SAR images, such as speckle noise, severely impact image quality and subsequent processing effectiveness, necessitating more refined preprocessing and enhancement techniques.

[0004] To address these challenges, researchers have designed complex deep learning networks to learn data features, improving algorithm performance. However, this also increases the demand for hardware resources, hindering its use under resource-constrained conditions. Contextual information within a scene provides crucial information for better understanding target features in SAR images, enabling faster and more accurate object detection and identification. While the human visual system can easily utilize contextual information, computer vision systems still face challenges in contextual reasoning. Therefore, effectively extracting key information from large amounts of SAR image data to assist deep learning algorithms in achieving accurate and efficient intelligent SAR image target detection is a key research focus. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, apparatus, and equipment for optimizing the detection results of target detection networks that can effectively improve the detection accuracy, in order to address the aforementioned technical problems.

[0006] A method for optimizing the detection results of an object detection network, the method comprising:

[0007] Obtain a priori geographic dataset, which contains multiple geographic images labeled with target type and location information;

[0008] Based on the prior geographic dataset, construct a geospatial category probability map with geospatial relationships of different target categories;

[0009] Acquire a SAR image to be detected, and use a target detection neural network to detect targets in the SAR image to obtain target category probabilities and target detection boxes;

[0010] The prior-enhanced classifier weights are obtained by weighting and modulating the classifier weights in the target detection neural network based on the geospatial category probability map.

[0011] The target category probability is optimized based on the prior-enhanced classifier weights, and the optimized target category probability and the target detection box are used as the final target detection result.

[0012] In one embodiment, the prior geographic dataset includes geographic images of various different regions, and a corresponding geospatial category probability map is constructed for each region;

[0013] The probability of the target category is optimized using a geospatial category probability map corresponding to the region in the SAR image to be detected.

[0014] In one embodiment, constructing a corresponding geospatial category probability map for each type of region includes:

[0015] Construct a corresponding latitude and longitude coordinate system based on the latitude and longitude information of each region of interest in the prior geographic dataset, and then grid it according to a preset resolution;

[0016] Based on latitude and longitude coordinate axes, the center coordinates of the target location label boxes on multiple geographic images related to the corresponding region are converted into latitude and longitude coordinates, and then the probability distribution of each type of target at different locations is calculated using a two-dimensional Gaussian distribution.

[0017] For each different region, all relevant geographic images are traversed, and the two-dimensional Gaussian distributions obtained from each geographic image are superimposed to obtain the geospatial category probability map of the corresponding region.

[0018] In one embodiment, constructing a corresponding latitude and longitude coordinate system based on the latitude and longitude information of each region of interest in the prior geographic dataset, and then gridding it according to a preset resolution includes:

[0019] The latitude and longitude directions of the latitude and longitude range of the region of interest are respectively used as the horizontal and vertical axes of the latitude and longitude coordinate system;

[0020] The longitude and latitude resolutions are calculated based on the preset surface resolution, and the longitude and latitude coordinate system is gridded based on the longitude and latitude resolutions.

[0021] The longitude resolution and latitude resolution are calculated using the following formulas:

[0022]

[0023] In the above formula, R lon R lat These refer to the longitude resolution and latitude resolution, respectively; lon1 and lon2 are the longitude ranges of the region; lat1 and lat2 are the latitude ranges of the region; S lon S lat λ represents the surface distance along the longitude and latitude directions of the region, respectively, and λ is the preset surface resolution.

[0024] In one embodiment, a two-dimensional Gaussian distribution is used to calculate the probability distribution of each type of target at different locations, using the following formula:

[0025]

[0026] In the above formula, P(C i |X,G) represents category C i The probability density function at position G, (G x G y (Box) represents the latitude and longitude coordinates corresponding to each location (x, y) on the geographic image. ctrlon Box ctrlon ) represents the center longitude and latitude coordinates of the location label box, δ represents the standard deviation, and R lon R lat These refer to the longitude resolution and latitude resolution, respectively.

[0027] In one embodiment, the prior-enhanced classifier weights are obtained by weighting and modulating the classifier weights in the target detection neural network according to the geospatial category probability map, including:

[0028] The coordinates of the center point of the target detection box predicted by the target detection neural network are converted into latitude and longitude coordinates.

[0029] Based on the latitude and longitude coordinates, obtain the prior probability of the geographic category at the corresponding location in the geospatial category probability map;

[0030] The classifier weights are weighted and modulated based on the prior probability of the geographical category to obtain prior-enhanced classifier weights.

[0031] In one embodiment, the weighting of the classifier weights based on the prior probability of the geographical category is weighted using the following formula:

[0032]

[0033] In the above formula, X represents the SAR image to be detected, and C i Let β represent the i-th target category, and let β represent the weighting factor. This represents the category C of geographic location G on the geospatial category probability map. i The sum of prior probabilities, where j represents the number of historical images of the region of interest, and P posterior (C i (X, G) represents the target detection neural network's detection of target category C at geographical location G. i The prediction confidence level.

[0034] In one embodiment, the target to be detected in the SAR image is an aircraft target, and different areas are different airport areas.

[0035] This application also provides a device for optimizing the detection results of a target detection network, the device comprising:

[0036] The prior geographic dataset acquisition module is used to acquire a prior geographic dataset, which contains multiple geographic images labeled with target type and location information;

[0037] The geospatial category probability map construction module is used to construct geospatial category probability maps with geospatial relationships of different target categories based on the prior geographic dataset.

[0038] The target detection result acquisition module is used to acquire the SAR image to be detected, use the target detection neural network to detect the targets in the SAR image, and obtain the target category probability and the target detection box;

[0039] The classifier weight enhancement module is used to perform weighted modulation on the classifier weights in the target detection neural network according to the geospatial category probability map to obtain the prior enhanced classifier weights;

[0040] The target detection result optimization module is used to optimize the target category probability based on the prior enhanced classifier weights, and use the optimized target category probability and the target detection box as the final target detection result.

[0041] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps:

[0042] Obtain a priori geographic dataset, which contains multiple geographic images labeled with target type and location information;

[0043] Based on the prior geographic dataset, construct a geospatial category probability map with geospatial relationships of different target categories;

[0044] Acquire a SAR image to be detected, and use a target detection neural network to detect targets in the SAR image to obtain target category probabilities and target detection boxes;

[0045] The prior-enhanced classifier weights are obtained by weighting and modulating the classifier weights in the target detection neural network based on the geospatial category probability map.

[0046] The target category probability is optimized based on the prior-enhanced classifier weights, and the optimized target category probability and the target detection box are used as the final target detection result.

[0047] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0048] Obtain a priori geographic dataset, which contains multiple geographic images labeled with target type and location information;

[0049] Based on the prior geographic dataset, construct a geospatial category probability map with geospatial relationships of different target categories;

[0050] Acquire a SAR image to be detected, and use a target detection neural network to detect targets in the SAR image to obtain target category probabilities and target detection boxes;

[0051] The prior-enhanced classifier weights are obtained by weighting and modulating the classifier weights in the target detection neural network based on the geospatial category probability map.

[0052] The target category probability is optimized based on the prior-enhanced classifier weights, and the optimized target category probability and the target detection box are used as the final target detection result.

[0053] The aforementioned method, apparatus, and equipment for optimizing the detection results of target detection networks first construct a geospatial category probability map with geospatial category relationships for different target categories. Then, a target detection neural network is used to detect targets in SAR images, obtaining target category probabilities and target detection boxes. Based on the geospatial category probability map, the classifier weights in the target detection neural network are weighted and modulated to obtain prior-enhanced classifier weights. The target category probabilities are then optimized based on these prior-enhanced classifier weights, and the optimized target category probabilities and target detection boxes are used as the final target detection results. This method can improve the accuracy of target detection results. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating a method for optimizing the detection results of a target detection network in one embodiment;

[0055] Figure 2 This is a schematic diagram illustrating the aircraft category probabilities modeled based on a Gaussian distribution in one embodiment, wherein... Figure 2 (a) represents a slice image containing three aircraft. Figure 2 (b) A three-dimensional plot showing the probability distribution of aircraft target position and category based on Gaussian distribution modeling;

[0056] Figure 3 This is a visualization of a geospatial category probability map at an airport in one embodiment, wherein... Figure 3 (a) is a SAR image of airport (I). Figure 3 (b), (c), and (d) represent the geospatial category probability maps corresponding to the three different types of aircraft, respectively.

[0057] Figure 4 This is a schematic diagram of the prior probability distribution of categories on a geospatial category probability map in one embodiment;

[0058] Figure 5 This is a schematic diagram illustrating the implementation steps of a method for optimizing the detection results of a target detection network in one embodiment;

[0059] Figure 6 This is a structural block diagram of a device for optimizing the detection results of a target detection network in one embodiment;

[0060] Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0062] like Figure 1 As shown, this application provides a method for optimizing the detection results of an object detection network, the specific steps of which include:

[0063] Step S100: Obtain a priori geographic dataset, which contains multiple geographic images labeled with target type and location information.

[0064] Step S110: Based on the prior geographic dataset, construct a geospatial category probability map with geospatial relationships of different target categories.

[0065] Step S120: Obtain the SAR image to be detected, and use the target detection neural network to detect the target in the SAR image to obtain the target category probability and the target detection box.

[0066] Step S130: The classifier weights in the target detection neural network are weighted and modulated according to the geospatial category probability map to obtain the prior-enhanced classifier weights.

[0067] Step S140: Optimize the target category probability based on the prior enhanced classifier weights, and use the optimized target category probability and the target detection box as the final target detection result.

[0068] In this embodiment, considering the special scenarios of certain target detection, different target categories typically have corresponding parking locations within certain designated parking areas to optimize management. Based on this prior knowledge, the typical parking situations of different types of targets within a given area can be effectively simulated by analyzing the target distribution on historical images of a target parking area, making it easier to generalize to dynamic and complex environments. Furthermore, this prior knowledge of geospatial target category distribution is used to post-process the target results predicted by the target detection neural network to further improve the accuracy of the detection results.

[0069] In this embodiment, target recognition can be performed on different types of vehicles in a ground parking lot or on different types of aircraft parked in an airport area. This paper will use an aircraft as an example to illustrate the concept.

[0070] In steps S100 and S110, before using the target detection neural network to detect the image to be detected, a geospatial category probability map is first constructed using multiple historical geographic images with ground truth labels.

[0071] In this embodiment, all geographic images in the prior geographic dataset are SAR images of the same type as the image to be detected. Considering that target recognition in SAR images is more difficult than in optical images, and that neural network-based SAR image target recognition will also have errors in its target detection results due to the special properties of SAR images, this is because the discrete appearance and attitude sensitivity of targets (especially aircraft targets) in SAR images mean that potential aircraft targets may exhibit a low confidence level and may be very similar to background clutter or other similar aircraft types, leading to ambiguity in classification. Therefore, this application mainly focuses on target recognition for SAR images as the image to be detected. By directly optimizing the target detection results after the neural network outputs the results, this method improves accuracy while being more efficient. It does not require optimization of the neural network structure itself or extensive training to give it more accurate predictive capabilities. Moreover, the training of neural networks for SAR image target recognition has always suffered from the problem of insufficient training samples.

[0072] Understandably, if this method can effectively improve the detection accuracy on SAR images, then it can also obtain accurate detection results when the image to be detected is an optical image.

[0073] In step S110, considering that different airports have different specifications and different parking regulations for each aircraft, separate geospatial category probability maps are constructed for multiple different airports when building the geospatial category probability maps. Therefore, the prior geographic dataset includes multiple geographic images corresponding to different regions, i.e., airports, and corresponding geospatial category probability maps are constructed based on these geographic images. In step S130, the target category probability is optimized based on the geospatial category probability map corresponding to the region, i.e., the airport, in the SAR image to be detected.

[0074] In this embodiment, constructing a corresponding geospatial category probability map for each type of region includes: constructing a corresponding latitude and longitude coordinate system based on the latitude and longitude location information of each type of region in the prior geographic dataset, and gridding it according to a preset resolution; based on the latitude and longitude coordinate axes, converting the center coordinates of the target location label boxes on multiple geographic images related to the corresponding region into latitude and longitude coordinates, and using a two-dimensional Gaussian distribution to calculate the probability distribution of each type of target at different locations; for each different region, traversing all related geographic images, and superimposing the two-dimensional Gaussian distributions obtained from each geographic image to obtain the geospatial category probability map of the corresponding region.

[0075] Specifically, determining the coordinate system and map resolution is crucial in creating geospatial category probabilistic maps. This method employs a latitude and longitude coordinate system, which better aligns with the expressions of geographic information. Then, a detailed representation is achieved by setting the latitude and longitude of surface distances as the units for the horizontal and vertical axes. Simultaneously, a coordinate system is constructed based on the geographical location information of each airport.

[0076] Furthermore, based on the latitude and longitude location information of each type of region in the prior geographic dataset, a corresponding latitude and longitude coordinate system is constructed, and it is gridded according to a preset resolution. This includes: first, taking the latitude and longitude directions of the region's latitude and longitude range as the horizontal and vertical axes of the latitude and longitude coordinate system, respectively; then, calculating the longitude and latitude resolutions according to the preset surface resolution; and finally, gridding the latitude and longitude coordinate system according to the longitude and latitude resolutions.

[0077] The following formulas are used to calculate the longitude resolution and latitude resolution, respectively:

[0078]

[0079] In the above formula, R lon R lat These refer to the longitude resolution and latitude resolution, respectively; lon1 and lon2 are the longitude ranges of the region; lat1 and lat2 are the latitude ranges of the region; S lon S lat λ represents the surface distance along the longitude and latitude directions of the region, respectively, and λ is the preset surface resolution.

[0080] Preferably, the preset resolution is 5 meters.

[0081] Furthermore, the surface distances along the longitude and latitude directions of the region are calculated using the following formulas:

[0082]

[0083] In the above formula, R is the radius of the Earth.

[0084] After defining the geospatial category probability map, a two-dimensional Gaussian distribution (GD) is used to model the category probability distribution of targets. While there are other alternative methods for modeling geospatial probability maps, the Gaussian distribution has unique advantages in resolving complex spatial relationships. On the one hand, the well-understood properties of GD make it very suitable for probabilistic modeling. On the other hand, despite the inherent positioning bias in SAR imagery, the continuous probability distribution of GD helps to effectively characterize aircraft distribution information.

[0085] In this embodiment, after converting the center coordinates of the target location label box in each geographic image into latitude and longitude coordinates in the corresponding coordinate system, a two-dimensional Gaussian distribution is used to calculate the probability distribution of each type of target at different locations, using the following formula:

[0086]

[0087] In the above formula, P(C i |X,G) represents category C i The probability density function at position G, (G x G y (Box) represents the latitude and longitude coordinates corresponding to each location (x, y) on the geographic image. ctrlon Box ctrlon ) represents the center longitude and latitude coordinates of the location label box, δ represents the standard deviation, and R lon R lat These are the longitude resolution and latitude resolution, respectively.

[0088] like Figure 2 As shown, this is a probability diagram of aircraft categories based on Gaussian distribution modeling, where... Figure 2 (a) is a slice image containing three aircraft targets. Figure 2 (b) is based on Figure 2 (a) A three-dimensional plot of the probability distribution of aircraft target position and category modeled using Gaussian distribution.

[0089] Specifically, by calculating the probability density function of each target on each geographic image and superimposing the Gaussian distributions of targets belonging to the same airport, a normalized geospatial category probability map with multiple Gaussian distributions can be obtained, in which the parking location of the aircraft has a relatively high probability. Thus, the geospatial category relationships between different types of aircraft in historical images can be comprehensively summarized.

[0090] like Figure 3 The image shown is a geospatial category probability map for Airport I. Figure 3 In the geospatial category probability map, each grid represents a composite of multiple Gaussian distributions associated with neighboring targets from multiple historical images. Figure 3 (a) is a SAR image of Airport I, which covers three types of aircraft. Figure 3 (b) Figure 3 (c) and Figure 3 (d) represents the geospatial category probability map corresponding to the three different types of aircraft.

[0091] Next, the prior probability of the aircraft category at the location of the target detection box is obtained from the geospatial category probability map. The classifier weights in the target detection network are then weighted and modulated based on the prior probability of the aircraft target category to obtain the prior-enhanced classifier weights.

[0092] Based on the detection head in the object detection network, predicted bounding box locations and class probability predictions are generated. Non-maximum suppression (NMS) is used to remove overlapping predictions, retaining the predicted bounding boxes with the highest confidence and the highest probability of containing the target. Simultaneously, in retaining the highest-confidence predicted bounding boxes, the class weights of these boxes are compared with those of other overlapping predicted bounding boxes, and the box with the highest weight is retained by class as the classifier weight vector output.

[0093] Based on this, using the geospatial category probability map, the probability of different target categories existing at the predicted bounding box location is determined, i.e., the prior probability of the geospatial category. Then, the prior probability of the geospatial category is weighted with the classifier weights output by the detection head (i.e., the posterior probability) to update and adjust the target category probability at each predicted bounding box, resulting in prior-enhanced classifier weights. The prior-enhanced classifier weights are then processed using the softmax function to obtain the predicted target category at each predicted bounding box location, thereby achieving accurate target detection.

[0094] In step S130, the prior-enhanced classifier weights are obtained by weighting and modulating the classifier weights in the target detection neural network according to the geospatial category probability map. This includes: taking the center point coordinates of the target detection box predicted by the target detection neural network and converting the center point coordinates into latitude and longitude coordinates; obtaining the geospatial category prior probability at the corresponding location in the geospatial category probability map based on the latitude and longitude coordinates; and weighting and modulating the classifier weights according to the geospatial category prior probability to obtain the prior-enhanced classifier weights, thereby improving the accuracy and stability of classification prediction.

[0095] In this embodiment, the classifier weights are weighted and modulated based on the prior probability of geographical categories, using the following formula:

[0096]

[0097] In the above formula, X represents the SAR image to be detected, and C i Let β represent the i-th target category, and let β represent the weighting factor. This represents the sum of prior probabilities of category G on the geospatial category probability map for geographical location G. The calculation principle is as follows: Figure 4 As shown. Target category C at geographic location G. i The prior probability is a composite result of multiple Gaussian distributions associated with neighboring targets in multiple historical images. j represents the number of historical images of the region of interest, and P...posterior (C i (X, G) represents the target detection neural network's detection of target category C at geographical location G. i The prediction confidence level.

[0098] like Figure 5 The diagram shown is a schematic of the processing flow of this method.

[0099] In this embodiment, the neural network used for target detection in SAR images is not limited to one type; existing mainstream detection neural networks can be used, including Oriented R-CNN, RetinaNet-OBB, Oriented RepPoints, etc.

[0100] This paper also verifies the effectiveness of the proposed method through experiments. In the experiments, based on mainstream object detection algorithms, the performance of the proposed optimization method is compared. Accuracy is quantified using mAP, as shown in Table 1.

[0101] Table 1. Performance Comparison of Detection Result Optimization Methods Based on Mainstream Object Detection Networks

[0102]

[0103] According to Table 1, compared to the target detection results directly generated by Oriented R-CNN, RetinaNet-OBB, and Oriented RepPoints networks, the proposed optimization method demonstrates superior performance in fine-grained target detection of SAR images, achieving mAPs of 73.6%, 71.3%, and 63.8%, respectively. This result validates that integrating prior knowledge helps improve target detection performance.

[0104] The aforementioned method for optimizing the detection results of target detection networks involves first constructing a geospatial category probability map with geospatial category relationships for different target categories. Then, a target detection neural network is used to detect targets in SAR images, obtaining target category probabilities and target detection boxes. Based on the geospatial category probability map, the classifier weights in the target detection neural network are weighted and modulated to obtain prior-enhanced classifier weights. These prior-enhanced classifier weights are then used to optimize the target category probabilities. The optimized target category probabilities and target detection boxes are then used as the final target detection results. This method can improve the accuracy of target detection results.

[0105] It should be understood that, although Figure 1The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed 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 some of the sub-steps or stages of other steps.

[0106] In one embodiment, such as Figure 6 As shown, a device for optimizing the detection results of an object detection network is provided, comprising: a prior geographic dataset acquisition module 200, a geospatial category probability map construction module 210, an object detection result acquisition module 220, a classifier weight enhancement module 230, and an object detection result optimization module 240, wherein:

[0107] The prior geographic dataset acquisition module 200 is used to acquire a prior geographic dataset, which contains multiple geographic images labeled with target type and location information.

[0108] The geospatial category probability map construction module 210 is used to construct a geospatial category probability map with geospatial relationships of different target categories based on the prior geographic dataset.

[0109] The target detection result acquisition module 220 is used to acquire the SAR image to be detected, and to detect the target in the SAR image using a target detection neural network to obtain the target category probability and the target detection box;

[0110] The classifier weight enhancement module 230 is used to perform weighted modulation on the classifier weights in the target detection neural network according to the geospatial category probability map to obtain the prior enhanced classifier weights;

[0111] The target detection result optimization module 240 is used to optimize the target category probability based on the prior enhanced classifier weights, and use the optimized target category probability and the target detection box as the final target detection result.

[0112] Specific limitations regarding the device for optimizing detection results for target detection networks can be found in the limitations of the method for optimizing detection results for target detection networks described above, and will not be repeated here. Each module in the aforementioned device for optimizing detection results for target detection networks can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0113] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for optimizing detection results for a target detection network. The display screen can be an LCD screen or an e-ink display screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0114] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0115] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0116] Obtain a priori geographic dataset, which contains multiple geographic images labeled with target type and location information;

[0117] Based on the prior geographic dataset, construct a geospatial category probability map with geospatial relationships of different target categories;

[0118] Acquire a SAR image to be detected, and use a target detection neural network to detect targets in the SAR image to obtain target category probabilities and target detection boxes;

[0119] The prior-enhanced classifier weights are obtained by weighting and modulating the classifier weights in the target detection neural network based on the geospatial category probability map.

[0120] The target category probability is optimized based on the prior-enhanced classifier weights, and the optimized target category probability and the target detection box are used as the final target detection result.

[0121] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0122] Obtain a priori geographic dataset, which contains multiple geographic images labeled with target type and location information;

[0123] Based on the prior geographic dataset, construct a geospatial category probability map with geospatial relationships of different target categories;

[0124] Acquire a SAR image to be detected, and use a target detection neural network to detect targets in the SAR image to obtain target category probabilities and target detection boxes;

[0125] The prior-enhanced classifier weights are obtained by weighting and modulating the classifier weights in the target detection neural network based on the geospatial category probability map.

[0126] The target category probability is optimized based on the prior-enhanced classifier weights, and the optimized target category probability and the target detection box are used as the final target detection result.

[0127] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application 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. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0128] The technical features of the above embodiments can be combined in any way. For the sake of brevity, 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.

[0129] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for optimizing detection results of a target detection network, characterized in that, The method includes: Obtain a priori geographic dataset, which contains multiple geographic images labeled with target type and location information; Based on the prior geographic dataset, a geospatial category probability map with geospatial relationships of different target categories is constructed. The prior geographic dataset includes geographic images of various regions. A corresponding geospatial category probability map is constructed for each region. The geospatial category probability map is generated by overlaying latitude and longitude grids and two-dimensional Gaussian distributions of the locations of each target category. The probability distribution of each target category at different locations is calculated using the following formula: In the above formula, Indicate category In position The probability density function on, ( , (x, y) are the latitude and longitude coordinates corresponding to each location (x, y) on the geographic image. , The coordinates () represent the center longitude and latitude coordinates of the location label box. Indicates standard deviation, , These are the longitude resolution and the latitude resolution, respectively. Acquire a SAR image to be detected, and use a target detection neural network to detect targets in the SAR image to obtain target category probabilities and target detection boxes; The prior-enhanced classifier weights are obtained by weighting and modulating the classifier weights in the target detection neural network according to the geospatial category probability map. Specifically, the center point coordinates of the target detection box predicted by the target detection neural network are converted into latitude and longitude coordinates. The geospatial category prior probability at the corresponding location is obtained according to the latitude and longitude coordinates in the geospatial category probability map. The classifier weights are then weighted and modulated according to the geospatial category prior probability to obtain the prior-enhanced classifier weights. The target category probability is optimized based on the prior-enhanced classifier weights, and the optimized target category probability and the target detection box are used as the final target detection result.

2. The method for optimizing detection results according to claim 1, characterized in that, For each type of region, a corresponding geospatial category probability map is constructed, which also includes: The probability of the target category is optimized using a geospatial category probability map corresponding to the region in the SAR image to be detected.

3. The method for optimizing detection results according to claim 2, characterized in that, The construction of a corresponding geospatial category probability map for each type of region includes: Construct a corresponding latitude and longitude coordinate system based on the latitude and longitude information of each region of interest in the prior geographic dataset, and then grid it according to a preset resolution; Based on latitude and longitude coordinate axes, the center coordinates of the target location label boxes on multiple geographic images related to the corresponding region are converted into latitude and longitude coordinates, and then the probability distribution of each type of target at different locations is calculated using a two-dimensional Gaussian distribution. For each different region, all relevant geographic images are traversed, and the two-dimensional Gaussian distributions obtained from each geographic image are superimposed to obtain the geospatial category probability map of the corresponding region.

4. The method for optimizing detection results according to claim 3, characterized in that, Constructing a corresponding latitude and longitude coordinate system based on the latitude and longitude information of each region of interest in the prior geographic dataset, and then gridding it according to a preset resolution, includes: The latitude and longitude directions of the latitude and longitude range of the region of interest are respectively used as the horizontal and vertical axes of the latitude and longitude coordinate system; The longitude and latitude resolutions are calculated based on the preset surface resolution, and the longitude and latitude coordinate system is gridded based on the longitude and latitude resolutions. The longitude resolution and latitude resolution are calculated using the following formulas: In the above formula, , These refer to the longitude resolution and latitude resolution, respectively. , The longitude range of the region. , The latitude range of the region. , These represent the surface distances along the longitude and latitude directions of the region, respectively. The preset surface resolution is used.

5. The method for optimizing detection results according to claim 4, characterized in that, The weighting of the classifier based on the prior probability of the geographical category is performed using the following formula: In the above formula, This represents the SAR image to be detected. Indicates the first target categories, Indicates the weighting factor. Represents geographical location on a geospatial category probability map. Category The sum of prior probabilities, This indicates the number of historical images representing the region of interest. This indicates that the object detection neural network is located in the geographical location. Top of target category The prediction confidence level.

6. The method for optimizing detection results according to claim 5, characterized in that, The target to be detected in the SAR image is an aircraft target, and different areas are different airport areas.

7. A device for optimizing detection results of a target detection network, characterized in that, The device implements the detection result optimization method for the target detection network according to any one of claims 1-6, including: The prior geographic dataset acquisition module is used to acquire a prior geographic dataset, which contains multiple geographic images labeled with target type and location information; The geospatial category probability map construction module is used to construct geospatial category probability maps with geospatial relationships of different target categories based on the prior geographic dataset. The target detection result acquisition module is used to acquire the SAR image to be detected, use the target detection neural network to detect the targets in the SAR image, and obtain the target category probability and the target detection box; The classifier weight enhancement module is used to perform weighted modulation on the classifier weights in the target detection neural network according to the geospatial category probability map to obtain the prior enhanced classifier weights; The target detection result optimization module is used to optimize the target category probability based on the prior enhanced classifier weights, and use the optimized target category probability and the target detection box as the final target detection result.

8. 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, it implements the steps of the method according to any one of claims 1 to 6.

Images