Key point rotating target detection method and system based on fusion Gaussian measurement

Abstract

The invention discloses a key point rotating target detection method and system based on fusion Gaussian measurement. The method comprises the following steps: firstly, outputting a multi-scale feature map through a feature extraction module and a feature fusion module; secondly, on the basis of the multi-scale feature map, a detection head module predicts the center point position of a target, and an initial key point set is regressed by using deformable convolution; gaussian modeling and quality evaluation are carried out on the initial key point set to obtain a final key point set, a loss function fused with Gaussian measurement is constructed in the training stage, key point distribution is fitted through maximum likelihood estimation, and KL divergence is used as distribution distance measurement to obtain a final key point set; evaluating the quality of the key points from five dimensions of classification, positioning, direction, correlation and Gaussian distribution; and finally, converting the final key point set into a rotation bounding box, carrying out post-processing by adopting a rotation box confidence coefficient propagation module, and outputting a final rotation target detection result. The method is superior to an existing optimal method in public data sets, and the detection precision is remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and remote sensing image processing technology, specifically to a keypoint rotation target detection method and system based on fused Gaussian metrics. Background Technology

[0002] For target detection tasks in remote sensing imagery, it is often necessary to accurately locate targets with arbitrary orientations and extreme aspect ratios (such as aircraft, ships, bridges, etc.). Since these targets appear in any 360-degree direction in the image, traditional horizontal bounding boxes are insufficient to accurately describe their true boundaries, leading to large localization errors and high false negative rates. Existing methods typically directly regress the parameters of the rotated bounding box, but this suffers from periodic discontinuities in the angles, affecting model training stability and detection accuracy.

[0003] In existing technologies, five-parameter or eight-parameter methods are generally used to represent the rotated bounding box, and its geometric parameters are predicted by deep learning models. However, when dealing with targets with large aspect ratios, this method is susceptible to sudden changes in angle critical points and vertex order swaps, leading to unstable regression results. Furthermore, while keypoint set methods can avoid explicit angle prediction, they are prone to generating isolated keypoints, disrupting the overall target structure. These shortcomings limit the effectiveness of existing technologies in complex remote sensing scenarios. Summary of the Invention

[0004] To address the problems of discontinuous angle regression boundaries and vertex order swapping caused by using five-parameter or eight-parameter methods to represent rotating frames when processing targets with arbitrary orientations and large aspect ratios in remote sensing images, as well as the problem of isolated points affecting detection accuracy in keypoint set representation methods, this application provides a keypoint rotating target detection method and system based on fused Gaussian metrics.

[0005] Firstly, this application provides a keypoint rotation target detection method based on fused Gaussian metric, comprising the following steps:

[0006] The system acquires and preprocesses the image to be detected, extracts features through the feature extraction module, and receives feature layers of different depths from the feature extraction module, outputting multi-scale feature maps with consistent semantic intensity but different spatial resolutions.

[0007] Based on the multi-scale feature map, the detection head module predicts the center point position of its target, and uses deformable convolution to regress the initial key point set based on the center point position.

[0008] The initial key point set is subjected to Gaussian modeling and quality evaluation, wherein the quality evaluation screens a high-quality key point set from five dimensions of classification quality, positioning quality, direction quality, key point set correlation quality and Gaussian metric quality, and the deformable convolution is used to finely predict the high-quality key point set to obtain a final key point set;

[0009] In the training stage of the Gaussian modeling and quality evaluation module, a loss function integrating Gaussian metrics is constructed, and the loss function includes classification loss, Gaussian distribution distance loss, positioning loss and shape / relative position loss, which are used for supervised model training.

[0010] In the training stage, the ConvexHull method is used to convert the key point set into a bounding box for loss calculation; in the inference stage, the MinAreaRect function is used to convert the final key point set into a rotating bounding box, which is post-processed through a rotating box confidence propagation module, and a graph structure between detection boxes is constructed based on a Markov random field, and a reliable detection box is optimized by confidence propagation, and a final rotating target detection result is output.

[0011] In a second aspect, the application provides a key point rotating target detection system based on integrated Gaussian metrics, comprising:

[0012] A feature extraction and fusion module is configured to acquire an image to be detected and perform preprocessing, extract features through a feature extraction module, and output multi-scale feature maps with consistent semantic intensity but different spatial resolution through a feature fusion module receiving feature layers at different depths in the feature extraction module;

[0013] A detection head module is configured to predict the center point position of a target based on the multi-scale feature maps, and regress an initial key point set based on the center point position using a deformable convolution.

[0014] A key point Gaussian quality evaluation module is configured to perform Gaussian modeling and quality evaluation on the initial key point set, wherein the quality evaluation screens a high-quality key point set from five dimensions of classification quality, positioning quality, direction quality, key point set correlation quality and Gaussian metric quality, and the deformable convolution is used to finely predict the high-quality key point set to obtain a final key point set.

[0015] In the training stage of the Gaussian modeling and quality evaluation module, a loss function integrating Gaussian metrics is constructed, and the loss function includes classification loss, Gaussian distribution distance loss, positioning loss and shape / relative position loss, which are used for supervised model training.

[0016] A rotating box confidence propagation module is configured to, in a training stage, convert the key point set into a bounding box by using a differentiable ConvexHull method, so as to calculate a positioning loss and support gradient back propagation; in an inference stage, the MinAreaRect function is used to convert the final key point set into a rotating bounding box, the rotating box confidence propagation module is used for post-processing of the rotating bounding box, a graph structure between detection boxes is constructed based on a Markov random field, a reliable detection box is optimized by using confidence propagation, and finally a rotating target detection result is output.

[0017] In a third aspect, the application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the key point rotating target detection method based on fusion Gaussian metric when executing the computer program.

[0018] The application has the following beneficial effects:

[0019] Firstly, the application discards the traditional five-parameter or eight-parameter rotating box representation method, and instead uses a key point set to implicitly express the geometric structure of a target, and introduces a Gaussian distribution to model the key point set, thereby fundamentally avoiding the periodic boundary discontinuity problem in angle regression, making the model training more stable, and significantly improving the angle prediction accuracy for targets in any orientation.

[0020] Secondly, the key point Gaussian quality assessment (GKPA) mechanism proposed by the application can comprehensively evaluate the quality of the key point set from multiple dimensions, effectively identify and filter out isolated abnormal points caused by background interference or feature confusion, thereby greatly improving the detection robustness and positioning accuracy of long and narrow targets (such as bridges and ships) in remote sensing images.

[0021] Thirdly, the loss function designed by the application fuses Gaussian metrics, uses the distance between Gaussian distributions as a supervision signal, and directly constrains the overall shape and direction of the key point set, further optimizing the regression process.

[0022] Finally, the rotating box confidence propagation (RCP) module based on the Markov random field is used to replace the traditional non-maximum suppression (NMS), and through the confidence message passing between the detection boxes, the problem of false deletion or missed detection caused by unstable rotating box IoU calculation in a dense arrangement scene is effectively solved.

[0023] In summary, the application has achieved better detection accuracy (mAP) than the existing optimal method (SOTA) on multiple public remote sensing datasets such as DOTA, UCAS-AOD, and HRSC2016, and does not introduce additional computational overhead in the inference stage, balancing high precision and practicality. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings used in the embodiments of the present application will be briefly introduced below.

[0025] Figure 1 A general structure schematic diagram of a key point rotation target detection method based on fusion Gaussian metric provided for the embodiments of the present application is shown in the figure.

[0026] Figure 2 A key point set Gaussian modeling learning process schematic diagram provided for the embodiments of the present application is shown in the figure; (a) and (d) are real annotation boxes; (b) and (e) are optimization intermediate processes; (c) and (f) are modeling results.

[0027] Figure 3 A detection result visualization comparison diagram of the embodiments of the present application and a comparison method on a DOTA remote sensing data set is shown in the figure; (a), (c), (e) are detection results of the Oriented RepPoints method; (b), (d), (f) are detection results of the GRPDet method of the present application. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solutions and advantages of the present application more clear, the present application will be further described in detail below in combination with the drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and not to limit the present application.

[0029] The present application proposes a key point rotation target detection method (GRPDet) based on fusion Gaussian metric, which is suitable for processing a large number of targets with arbitrary direction and extreme aspect ratio existing in remote sensing images, such as airplanes, ships, bridges and vehicles, etc., including:

[0030] An image to be detected is obtained, and a multi-scale feature map is generated through a feature extraction module and a feature fusion module; based on the multi-scale feature map, an initial key point set is predicted by a deformable convolution with the target center point as the reference; a two-dimensional Gaussian distribution modeling is performed on the initial key point set, and a key point Gaussian quality assessment (GKPA) method is designed to comprehensively evaluate the quality of the initial key point set from five dimensions of classification, positioning, direction, key point set correlation and Gaussian metric. This step aims to avoid explicit angle regression through statistical modeling, thereby solving the problem of unstable angle prediction of rotation targets caused by the discontinuity of angle periodicity; a high-quality key point set is selected according to the comprehensive quality evaluation result, and a fine prediction of the high-quality key point set is performed by using a deformable convolution to obtain a final key point set. This step aims to filter out abnormal prediction caused by isolated points and improve the detection robustness of large aspect ratio targets.

[0031] Wherein in the training stage, a loss function integrating Gaussian metric is constructed, which includes classification loss, Gaussian distribution distance loss, and positioning loss for initialization stage and refinement stage respectively, for supervising model training; in the inference stage, the final key point set is converted into a rotating bounding box, and the detection result is post-processed through a rotating box confidence propagation (RCP) module, and the final rotating target detection result is output, replacing the traditional NMS, solving the problem of dense rotating target misdeletion and missed detection.

[0032] As shown in Figure 1 The overall network architecture of the embodiment of the present application mainly consists of a feature extraction module, a feature fusion module, a detection head module (including initial prediction and refinement prediction), a key point Gaussian quality evaluation module (GKPA), and a rotating box confidence propagation module (RCP). Figure 1 The complete process from the input image to the final output detection result is shown. The left side is the feature extraction and fusion module (ResNet50 + FPN), the middle is the detection head module (DCN), the upper side is the key point Gaussian quality evaluation (GKPA) module, and the right side is the rotating box confidence propagation (RCP) module. The dashed boxes mark the modules used in the training stage (only train) and the training / test stage (train & test) respectively.

[0033] As an optional implementation, the technical solution of the embodiment of the present application is as follows:

[0034] S1: First, remote sensing images usually have high resolution and large size (for example, the original image size in the DOTA dataset ranges from 800x800 to 4000x4000 pixels), and the target scale varies greatly (from small vehicles to large port facilities). In order to adapt to the fixed input requirements of deep neural networks and balance the calculation efficiency and detection performance, all input images are uniformly cropped and adjusted to 1024x1024 pixels in this embodiment. This standardization preprocessing helps to stabilize the model training and also optimizes the calculation efficiency. ResNet50 is used as the backbone network (Backbone) to extract basic features in this embodiment, and the Feature Pyramid Network (FPN) is used to fuse the features output by ResNet50 at different stages.

[0035] Specifically, the FPN receives feature layer outputs from different depths in the ResNet50 backbone network, usually labeled as C3, C4, C5, corresponding to the shallow, middle and deep layers of the network respectively. Among them, the C3 layer feature map has high resolution but weak semantic information, suitable for detecting small targets; the C5 layer feature map has low resolution but rich semantic information, suitable for detecting large targets. The FPN transmits strong semantic information from the deep layer to the high-resolution feature map of the shallow layer through the "top-down" and "lateral connection" methods. Finally, the FPN outputs a set of multi-scale feature maps with the same semantic strength but different spatial resolutions, usually labeled as P3, P4, P5. This multi-scale feature representation mechanism enables the embodiments of the present application to efficiently detect targets spanning multiple scales from tiny vehicles to huge bridges in remote sensing images in the same network, greatly improving the robustness and applicability of the detection system.

[0036] S2: In the detection head module, the model first predicts the center point position of the target on the multi-scale feature map. Then, based on each predicted center point, a set of initial keypoint sets is regressed through a deformable convolution (DCN) layer. The coordinate offset of the keypoint is normalized to the [0, 1] interval by the Sigmoid function and then mapped back to the original feature map coordinate system.

[0037] S3: Key point Gaussian modeling and quality assessment (GKPA) is performed on the initial keypoint set, aiming to solve the detection box distortion problem caused by isolated points in traditional keypoint methods.

[0038] Specifically, each keypoint is regarded as a two-dimensional sample , the coordinate is denoted as the coordinate vector of the ith keypoint , and it is assumed that the entire keypoint set obeys a two-dimensional Gaussian distribution in the two-dimensional plane. The probability density function of this Gaussian distribution is as follows:

[0039] ,

[0040] wherein, is the coordinate vector of the ith keypoint, is the two-dimensional mean vector of the keypoint set, is the variance matrix (covariance matrix).

[0041] The mean and variance matrix of the Gaussian distribution of the keypoint set are calculated by the maximum likelihood estimation (MLE) method:

[0042]

[0043]

[0044] where N is the number of key points.

[0045] On this basis, a key point Gaussian quality assessment (GKPA) method is designed to comprehensively evaluate the quality of the initial key point set from five dimensions of classification, positioning, direction, key point set correlation and Gaussian measurement. The calculation formula of the comprehensive quality score Q is as follows:

[0046]

[0047] where the weights of each quality are set as: , , , and According to the comprehensive quality evaluation result, a high-quality key point set is screened out, and the high-quality key point set is finely predicted by using a deformable convolution to obtain a final key point set.

[0048] The GKPA module provided in the embodiments of the present application includes quality evaluation of five dimensions, and each dimension is aimed at a specific technical pain point:

[0049] Classification quality ( ) is derived from the reverse value of the classification loss . The purpose is to ensure that only when the model has a high enough confidence in the target class, the key point set is considered reliable, which can effectively prevent false detection targets from being wrongly assigned a high weight, thereby improving the reliability of the overall detection.

[0050] Positioning quality ( ) is derived from the reverse value of the positioning loss . The purpose is to evaluate the accuracy of the overall position of the key point set. If the key point set is far away from the real target area, its positioning quality must be very low, so it is filtered out, solving the problem of key point set drift.

[0051] Direction quality ( ) is obtained by calculating the reverse value of the chamfer distance between the predicted key point set and the real annotation box. The purpose is to measure the direction consistency of the key point set. For slender targets (such as airplanes and ships), even if the position of the key point has a slight deviation, as long as the direction is consistent, an accurate bounding box can be formed, ensuring that the model can learn the correct direction information.

[0052] Correlation quality ( ) is obtained by calculating the cosine similarity between the feature vectors of the key points. The purpose is to evaluate the consistency within the key point set. A high-quality key point set should have features of each key point correlated with each other, pointing to the same target, and effectively suppressing isolated points caused by background noise or irrelevant objects.

[0053] Gaussian metric quality originated from the inverse value of KL divergence loss . It is the most core innovation of the GKPA module. It directly measures the matching degree between the predicted key point set Gaussian distribution and the real labeled Gaussian distribution. Not only the position of the key point is considered, but also the overall shape, scale and direction, ensuring that the target geometric structure learned by the model is complete and accurate, which is the fundamental solution to the two core problems of "unstable angle prediction of rotating targets" and "poor robustness of large aspect ratio target detection".

[0054] During the training phase, a loss function that integrates the Gaussian metric is constructed to supervise the model training. The total loss function is:

[0055]

[0056]

[0057] Among them, the specific meaning of each loss is as follows: is the classification loss, which adopts Focal Loss to predict the category of the target; is the Gaussian distribution distance loss, which calculates the KL divergence between the predicted key point set Gaussian model and the real labeled Gaussian model; is the positioning loss, which adopts ConvexHull method to measure the geometric similarity (such as GIoU) between the bounding box converted by the key point set and the real labeled box; is the shape / relative position loss, which is used to constrain the internal structure of the key point set (such as through the chamfer distance, etc.). The weight coefficients of each loss are set as: = 0.375, = 1.0 and = 1.0.

[0058] During training, the SGD optimizer is used, the initial learning rate is set to 2.5 × 10 -3 , the momentum is 0.9, the weight decay is 1 × 10 -4 , the batch size is 2, and the preheating learning rate strategy is used.

[0059] S4: During the training phase, the differentiable ConvexHull method is used to convert the keypoint set into bounding boxes to generate ground truth bounding boxes for the supervised Rotated Box Confidence Propagation (RCP) module and to support gradient backpropagation. During the inference phase, the efficient MinAreaRect function is used to convert the final keypoint set into rotated bounding boxes, thereby providing the RCP module with an initial set of detection boxes to be optimized. The rotated bounding box set converted from the final keypoint set is then post-processed by the Rotated Box Confidence Propagation module.

[0060] The RCP module in this application constructs a graph structure between detection boxes based on Markov random fields. It enhances reliable detection boxes and suppresses redundant boxes through positive and negative confidence suppression mechanisms.

[0061] Specifically, the RCP module in this application first constructs a graph connection based on the GIoU value between predicted boxes and sets a GIoU threshold; if the GIoU between two predicted boxes is greater than the threshold, an undirected edge is created between them. Subsequently, five confidence propagation iterations are performed.

[0062] In each iteration, the positive propagation message for each detection box is computed. and negative suppression messages And update the detection box accordingly. Confidence level:

[0063]

[0064]

[0065]

[0066] in, This indicates pointing to the i-th detection box. The active dissemination of messages, its computation depends on and Related set of weak detection boxes In the formula This is an iterable variable within the set. This indicates selecting the smooth confidence level from all neighbor boxes. The highest value. Among them, and Representing the detection boxes respectively and The confidence level after smoothing is designed to avoid numerical instability caused by extreme values ​​in the original confidence level; This is a preset balancing factor used to adjust the confidence ratio and GIoU ratio in calculating the weights of directional associations. The relative contribution of time; ultimately, negative suppression messages By detection boxOriginal confidence of The GIoU of its most strongly associated neighbor (even if the largest one ) is jointly determined, thus effectively suppressing redundant predictions.

[0067] Specifically, the positive propagation message is used to characterize the collective support strength of multiple low-confidence bounding boxes associated with a detection box , whose value is positively correlated with the number of these neighboring weak bounding boxes and their respective confidence levels; when more low-confidence bounding boxes are spatially clustered, it indicates a higher likelihood of the presence of a real target in that region, thus providing additional confidence gain for the reliable detection results within that region. The negative suppression message

[0068] is used to quantify the suppression effect of a high-confidence bounding box on a detection box , whose strength depends on the confidence of the neighboring strong bounding box and the spatial correlation weight between them , which is determined by the generalized intersection over union (GIoU) between the detection box and the neighboring strong bounding box ; a high-confidence bounding box will exert stronger suppression on the highly overlapping redundant prediction boxes around it to reduce the final confidence of the latter.

[0069] In addition, after each iteration, the GIoU threshold θ used to construct the bounding box association graph is dynamically adjusted according to the pre-set threshold fluctuation coefficient λ to adapt to the density changes of the detection results in different scenarios. After N iterations, the final rotated target detection result is output after post-processing.

[0070] The design of the propagation coefficient aims to balance the number of weak bounding boxes and the confidence weight through this coefficient to avoid the influence of extreme values on the results. As shown in

[0071] , the embodiments of the present application show the learning process of key point set Gaussian modeling. Figure 2 (a) and (d) in Figure 2 are real labeled boxes; Figure 2 (b) and (e) in Figure 2 are intermediate optimization processes; (c) and (f) in

[0072] are the final modeling results. This figure intuitively shows how the Gaussian ellipse gradually fits the real geometric shape of the target, especially suitable for characterizing the direction and scale characteristics of elongated targets in remote sensing images. Figure 3The detection results on the DOTA remote sensing dataset are shown for comparison. Figure 3 (a)(c)(e) in FIG. 1 are results of the Oriented RepPoints method; Figure 3 (b)(d)(f) in FIG. 1 are results of the embodiments of the present application. As can be seen from the comparison, the embodiments of the present application can generate more accurate and stable rotated boxes when processing complex scenes such as densely arranged vehicles and arbitrarily oriented aircrafts, effectively overcoming the problems of isolated point interference and angle regression instability, and verifying the superiority of the embodiments of the present application in the task of remote sensing image rotated object detection.

[0073] The hardware environment required by the embodiments of the present application is recommended as follows: Intel i7-11700 CPU, NVIDIA 3080 GPU and 32 GB memory; the software environment is based on Ubuntu 16.04, PyTorch 1.10.0, Python 3.9.15 and CUDA 11.3.

[0074] In an exemplary embodiment, a key point rotated object detection system based on fused Gaussian metric is provided, comprising:

[0075] A feature extraction and fusion module is configured to acquire and pre-process a to-be-detected image, and extract features through a feature extraction module; a feature fusion module receives feature layers at different depths in the feature extraction module, and outputs multi-scale feature maps with consistent semantic intensity but different spatial resolution;

[0076] A detection head module is configured to predict the center point position of a target based on the multi-scale feature maps, and utilize a deformable convolution to regress an initial key point set based on the center point position;

[0077] A key point Gaussian quality evaluation module is configured to perform Gaussian modeling and quality evaluation on the initial key point set, wherein the quality evaluation filters out a high-quality key point set from five dimensions of classification quality, positioning quality, direction quality, key point set correlation quality and Gaussian metric quality, and utilizes the deformable convolution to perform fine prediction on the high-quality key point set to obtain a final key point set;

[0078] In the training stage of the Gaussian modeling and quality evaluation module, a loss function fused with Gaussian metric is constructed, and the loss function includes classification loss, Gaussian distribution distance loss, positioning loss and shape / relative position loss, which are used to supervise the model training;

[0079] The rotating box confidence propagation module is configured to, in a training stage, convert the key point set into a bounding box by using a differentiable ConvexHull method, so as to calculate a positioning loss and support gradient back propagation; in an inference stage, the final key point set is converted into a rotating bounding box by using a MinAreaRect function, the rotating box confidence propagation module is used for post-processing of the rotating bounding box, a graph structure between detection boxes is constructed based on a Markov random field, a reliable detection box is optimized by using confidence propagation, and finally, a rotating target detection result is output.

[0080] In an exemplary embodiment, a computer device is provided, comprising a memory and a processor, the memory stores a computer program, and the processor implements the steps of the key point rotating target detection method based on fusion Gaussian metric when executing the computer program.

[0081] The above description of the embodiments is for the purpose of enabling a person of ordinary skill in the art to understand and apply the present application. Those skilled in the art can easily make various modifications to the above embodiments, and apply the general principles described herein to other embodiments without inventive labor. Therefore, the present application is not limited to the above embodiments, and any improvements and modifications made to the present application by those skilled in the art based on the disclosure of the present application shall be within the scope of protection of the present application.

Claims

1. A keypoint rotation target detection method based on fused Gaussian metric, characterized in that, Includes the following steps: The image to be detected is acquired and preprocessed, and features are extracted using the feature extraction module. feature The fusion module receives feature layers of different depths from the feature extraction module and outputs multi-scale feature maps with consistent semantic strength but different spatial resolutions. Based on the multi-scale feature map, the detection head module predicts the center point position of its target, and uses deformable convolution to regress the initial key point set based on the center point position. Gaussian modeling and quality assessment are performed on the initial keypoint set. The quality assessment selects high-quality keypoint sets from five dimensions: classification quality, localization quality, orientation quality, keypoint set relevance quality, and Gaussian metric quality. The deformable convolution is then used to perform refined prediction on the high-quality keypoint sets to obtain the final keypoint set. In the training phase of the Gaussian modeling and quality assessment module, a loss function that integrates Gaussian metrics is constructed. The loss function includes classification loss, Gaussian distribution distance loss, localization loss, and shape / relative position loss, which is used to supervise model training. During the training phase, the differentiable ConvexHull method is used to convert the keypoint set into bounding boxes to calculate the localization loss and support gradient backpropagation. During the inference phase, the MinAreaRect function is used to convert the final keypoint set into rotated bounding boxes. The rotated bounding box confidence propagation module is used for post-processing. At the same time, a graph structure between detection boxes is constructed based on Markov random fields, and the confidence propagation is used to optimize reliable detection boxes, outputting the final rotated target detection results.

2. The method according to claim 1, characterized in that, The preprocessing involves uniformly cropping and adjusting the remote sensing image to 1024×1024 pixels.

3. The method according to claim 1 or 2, characterized in that, The coordinate offsets of the initial keypoints are normalized to the [0,1] interval using the Sigmoid function, and then mapped back to the original feature map coordinate system.

4. The method according to claim 1, characterized in that, The Gaussian modeling of the initial keypoint set includes: Each keypoint is treated as a two-dimensional sample, and it is assumed that the initial set of keypoints follows a two-dimensional Gaussian distribution; The mean and variance matrix of the Gaussian distribution of the initial keypoint set are obtained by using the maximum likelihood estimation method.

5. The method according to claim 1 or 4, characterized in that, The classification quality is determined by the inverse value of the classification loss, ensuring that only the keypoint set of high-confidence targets is reliable. The positioning quality is determined by the inverse value of the positioning loss, which is used to evaluate the accuracy of the overall position of the key point set; If the keypoint set deviates from the actual target area, the positioning quality is low and it is filtered out, thus solving the keypoint drift problem; The directional quality is calculated based on the chamfer distance reverse value and is used to measure the directional consistency of the key point set; For slender targets, accurate bounding boxes can still be generated when minor offsets in the key point positions do not affect the consistency of orientation. The relevance quality of the keypoint set is obtained by calculating the cosine similarity between feature vectors, which is used to evaluate the consistency within the set; high-quality keypoint sets have features that are interconnected and point to the same target, thereby suppressing isolated points caused by background noise or irrelevant objects. The Gaussian metric quality is determined by the inverse KL divergence value, which measures the degree of matching between the predicted keypoint set and the true labeled Gaussian distribution.

6. The method according to claim 1, characterized in that, The classification loss uses Focal Loss to predict the target category; The Gaussian distribution distance loss is used to calculate the KL divergence between the predicted keypoint set Gaussian model and the true labeled Gaussian model. The localization loss is used to measure the geometric similarity between the bounding box transformed from the keypoint set and the true labeled box; The shape / relative position loss is used to constrain the internal structure of the keypoint set.

7. The method according to claim 1 or 6, characterized in that, The rotating frame confidence propagation module constructs a graph structure between detection frames based on a Markov random field, and optimizes reliable detection frames using confidence propagation, including the following steps: Graph connections are constructed based on the GIoU values ​​between predicted bounding boxes, and a GIoU threshold is set; if the GIoU between two predicted bounding boxes is greater than the threshold, an undirected edge is created. Perform confidence propagation iterations. In each iteration, update the confidence of the detection box by actively propagating messages and negatively suppressing messages, and dynamically adjust the GIoU threshold.

8. The method according to claim 7, characterized in that, The positive propagation message quantifies the collective support strength of neighboring low-confidence detection boxes for the detection box, and its value is positively correlated with the number and confidence of the neighboring low-confidence detection boxes. When low-confidence detection boxes cluster in space, the likelihood of a real target being present in that area is increased, providing additional confidence gain for reliable detection results; The negative suppression message quantifies the suppressive effect of high-confidence detection boxes on detection boxes. Its strength is determined by the confidence of neighboring strong detection boxes and the GIoU between them, thereby suppressing redundant detection boxes.

9. A keypoint rotation target detection system based on fused Gaussian metric, characterized in that, include: The feature extraction and fusion module is used to acquire the image to be detected and perform preprocessing, and to extract features through the feature extraction module; The feature fusion module receives feature layers of different depths from the feature extraction module and outputs multi-scale feature maps with consistent semantic strength but different spatial resolutions. The detection head module is used to predict the center point position of the target based on the multi-scale feature map, and to regress the initial key point set based on the center point position using deformable convolution. The key point Gaussian quality assessment module is used to perform Gaussian modeling and quality assessment on the initial key point set. The quality assessment selects high-quality key point sets from five dimensions: classification quality, localization quality, orientation quality, key point set relevance quality, and Gaussian metric quality. The deformable convolution is then used to perform refined prediction on the high-quality key point sets to obtain the final key point set. In the training phase of the Gaussian modeling and quality assessment module, a loss function that integrates Gaussian metrics is constructed. The loss function includes classification loss, Gaussian distribution distance loss, localization loss, and shape / relative position loss, which is used to supervise model training. The Rotation Box Confidence Propagation module, during the training phase, uses the differentiable ConvexHull method to convert the keypoint set into bounding boxes to calculate the localization loss and support gradient backpropagation. During the inference phase, it uses the MinAreaRect function to convert the final keypoint set into rotated bounding boxes, which are then post-processed by the Rotation Box Confidence Propagation module. Simultaneously, it constructs a graph structure between detection boxes based on Markov random fields, optimizes reliable detection boxes using confidence propagation, and outputs the final rotated target detection result.

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

Images