Small target detection method and system based on adaptive Gaussian learning label allocation

Abstract

The present invention relates to a small target detection method based on adaptive Gaussian learning label allocation, comprising: 1) obtaining a training data set and a training set annotation file; 2) constructing a small target detection network model, including a reference detector, an adaptive weighted Gaussian measurement module and a dynamic cyclic label allocation module; extracting a classification feature map and a positioning feature map through the reference detector for classification and positioning tasks respectively; for the positioning feature map, using an adaptive weighted Gaussian measurement module to obtain the Gaussian distribution corresponding to the real frame and the feature point and measure the similarity; in the label allocation process, using the dynamic cyclic label allocation module to dynamically set the threshold for positive and negative sample screening; 3) using the small target detection network model to detect targets. The method and system can provide a more accurate similarity score with the real frame for the candidate prior, and dynamically allocate appropriate positive samples to the real frame during the label allocation process, thereby improving the accuracy of small target detection.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision, and in particular relates to a small target detection method and system based on adaptive Gaussian learning label allocation. Background Art

[0002] Tiny objects can be found everywhere in object detection in various scenarios, including aerial image detection, medical image detection, industrial detection and other fields. In recent years, the rise of deep learning has set off a wave of using neural networks for recognition and positioning detection tasks. Although researchers have proposed a variety of baseline detectors and continuously updated and improved them, most of the proposed methods are for objects of normal size. For tiny objects, this is a challenging detection area, which usually occupies very few pixels, and it is difficult for the detector to extract discriminative features, resulting in many false detections and missed detections.

[0003] General object detectors can be roughly divided into anchor-based detectors and anchor-free detectors. Anchor-based detectors preset discrete prior boxes of different positions, scales, and aspect ratios, and then use IoU as a measure of the matching degree between the anchor and the ground truth (GT), thereby providing suitable samples in the subsequent positive and negative sample selection. Anchor-free detectors transfer the prior information from the box to the point, and distinguish positive and negative samples by whether the prior point is within the coverage of GT, thereby eliminating the workload of fine-tuning the anchor box.

[0004] Although the above two types of methods perform well in general object detection tasks, they perform poorly in small target detection tasks. The main reason is that the positive and negative sample determination mechanism cannot accurately measure the positional relationship between the anchor box (or point) and the gt. The same pixel deviation has different effects on objects of different sizes. Even a small pixel deviation in a tiny object can cause a significant drop in the IoU score, making it an extremely sensitive indicator. But for medium objects with larger pixels, the same deviation has less impact on the IoU score. Similarly, for anchor-free detectors, the reduction in pixels results in few anchors within the coverage of the gt, which cannot provide enough positive samples, directly affecting the performance of the detector.

[0005] Recently, Xu et al. tried to use the two-dimensional Gaussian distribution distance between gt and prior to measure their similarity. Specifically, two-dimensional Gaussian modeling was performed on gt boxes and candidate boxes respectively. The distance between the two distributions was calculated by introducing Wasserstein distance (WD), and then the obtained value was normalized to the final score. Xu et al. calculated the similarity by directly using the effective receptive field (ERF) to match gt, and introduced Kullback-Leibler divergence (KLD) as another measurement strategy. Although the above method solves the problems of extreme sensitivity of IoU and continuous score of IoU of 0, it still has shortcomings when measuring the similarity between two distributions. When different similarity metrics are used, there is an obvious inconsistency in similarity. This phenomenon of difference in similarity metrics leads to the fact that when only KLD / WD is used as a metric, some high-quality samples tend to obtain lower metric scores because the distance between the center points of the two distributions is large. Therefore, they are marked as negative samples in the subsequent positive and negative sample selection. Summary of the invention

[0006] The purpose of the present invention is to provide a small target detection method and system based on adaptive Gaussian learning label allocation, which can provide a more accurate similarity score with the true frame for the candidate prior and dynamically allocate appropriate positive samples to the true frame during the label allocation process, thereby improving the accuracy of small target detection.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is: a small target detection method based on adaptive Gaussian learning label allocation, comprising the following steps:

[0008] 1) Obtain training data set and training set annotation files;

[0009] 2) Construct a small target detection network model based on adaptive Gaussian learning label allocation, the small target detection network model includes a reference detector, an adaptive weighted Gaussian measurement module and a dynamic cyclic label allocation module; extract classification feature maps and positioning feature maps through the reference detector for classification tasks and positioning tasks respectively; for the positioning feature map, use the adaptive weighted Gaussian measurement module to perform two-dimensional Gaussian modeling on the real box and feature points to obtain the Gaussian distribution corresponding to the real box and feature points, and then use the adaptive weighted Gaussian measurement to measure the similarity of the two Gaussian distributions to more accurately measure the similarity between the Gaussian candidate box and the real box; in the label allocation process, use the dynamic cyclic label allocation module to dynamically set the threshold for positive and negative sample screening, so that the positive samples in each round of training have sufficient quantity and high quality to obtain more accurate label allocation results;

[0010] The small target detection network model is trained using the training data set and training set annotation files to obtain generalizable model parameters.

[0011] 3) Use the trained small target detection network model to detect targets and realize small target detection.

[0012] Furthermore, in step 1), a training data set and a corresponding training set annotation file are obtained from an official open source data set, wherein the training data set is an AI-TOD-v2 small target detection data set.

[0013] Furthermore, the specific implementation method of step 2) includes:

[0014] A) Obtain a classification feature map and a positioning feature map through a reference detector, and then input the obtained positioning feature map into an adaptive weighted Gaussian measurement module. The adaptive weighted Gaussian measurement module uses the annotation information of the real frame and the receptive field of the feature points to perform two-dimensional Gaussian modeling on all feature points on the positioning feature map and the real frame, and calculates the similarity between the Gaussian distribution of all the obtained feature points and the Gaussian distribution of the real frame through the Wasserstein distance. At the same time, the size of the overlapping area between the Gaussian distribution of each feature point and the Gaussian distribution of the real frame is calculated as an adaptive weight, which is used to adjust the priority of each feature point and provide a more accurate similarity score between each feature point and the real frame;

[0015] B) All similarity scores obtained from the adaptive weighted Gaussian metric module are input into the dynamic cyclic label assignment module, which assigns positive sample labels or negative sample labels to all candidate feature points by setting dynamic thresholds during the network training and inference stages to avoid some low-score samples from being assigned as positive samples.

[0016] Furthermore, the implementation method of the adaptive weighted Gaussian metric module is:

[0017] In order to measure the similarity between different candidate priors and the true box gt, the adaptive weighted Gaussian measurement module calculates the adaptive Gaussian weighted similarity score between the effective receptive field of the feature point on the positioning feature map and gt, and uses this score to replace the traditional intersection-over-union ratio IoU; first, two-dimensional Gaussian modeling is performed on the effective receptive field and gt, and the distance between the two Gaussian distributions is calculated using the Wasserstein distance; then, under the premise of two-dimensional Gaussian modeling, the priority of each feature point is adaptively adjusted by calculating the overlapping area score between each feature point and gt; the specific method is as follows:

[0018] First, calculate the n Layer Theory Receptive Field tr n :

[0019]

[0020] in, tr n It is n The theoretical receptive field of each feature point in the convolutional layer; k n and s n are the convolution kernel size and n The stride of the layer convolution operation;

[0021] For the effective receptive field, firstly, each feature point ( x n , y n ) is set to the mean vector of its standard two-dimensional Gaussian distribution ; Since the effective receptive field of each feature point is a subset of the theoretical receptive field, half of the theoretical receptive field radius is expressed as er n , set as the radius of the effective receptive field, and use er n The square of the two-dimensional Gaussian distribution constitutes the covariance For the real frame gt ( x g , y g , w g , h g ), the center point of each gt ( x g , y g ) is specified as the mean vector , and use ( w g / 2, h g / 2) squared constitutes the covariance in a two-dimensional Gaussian distribution ; The effective receptive field and gt are modeled as Gaussian distributions respectively and ; , , and They are respectively expressed as follows:

[0022]

[0023]

[0024] For the Gaussian distribution of effective receptive field and gt and , the second-order Wasserstein distance between them is expressed as follows:

[0025]

[0026] Through D wd Normalize it and control its value range within (0,1), thus giving a value range similar to IoU; after normalization, a new normalized metric score is obtained S d , which is expressed as:

[0027]

[0028] Among them, β represents the adjustment factor in the normalization process;

[0029] Furthermore, the overlapping area of ​​Gaussian distribution is used to design adaptive Gaussian weights for Wasserstein distance w Gauss ; Specifically:

[0030] First, the covariance matrix of the overlapping area of ​​the two Gaussian distributions is calculated by the Kalman filter , as shown below:

[0031]

[0032] Based on the covariance matrix , the overlapping area between two Gaussian distributions A pt It is expressed as:

[0033]

[0034] Overlap area fraction S o It is expressed as:

[0035]

[0036] in, A e represents the area of ​​the Gaussian prior, A g represents the area of ​​gt, S o Further expressed as:

[0037]

[0038]

[0039]

[0040] In the above formula S o The value range of is (0, 1 / 3); therefore, the condition factor α is introduced to normalize the value range to (0, 1), which is consistent with S d Stay consistent; final S o It is expressed as follows:

[0041]

[0042] In order to dynamically adjust the similarity score based on the overlap area S o , adaptive Gaussian weight w Gauss It is expressed as:

[0043]

[0044] Among them, n is a hyperparameter used to control different w Gauss relative gap;

[0045] Then, the adaptive weighted Gaussian metric is defined as ,Right now

[0046]

[0047] In order to distinguish between different Gaussian priors S AWG , further adding an exponential modulation factor γ to expand S AWG Gap; Final Adaptive Gaussian Weighted Similarity Score S' AWG It is expressed as:

[0048] .

[0049] Furthermore, the implementation method of the dynamic cyclic label allocation module is:

[0050] Using the similarity score obtained from the adaptive weighted Gaussian metric module S' AWG Obtain the score matrix between Gaussian prior and gt; for the score matrix, first select the first k priors for gt to prevent the introduction of too many low-score samples from affecting the setting of subsequent thresholds; the positive sample candidate set T(i) of the i-th gt is expressed as:

[0051]

[0052] in, Indicates the number of candidate samples for the i-th gt;

[0053] For a set T(i), by calculating the mean of all scores in the set m i and standard deviation sd i To set the dynamic threshold t i , the two are respectively expressed as:

[0054]

[0055]

[0056] For the i-th gt, only the similarity scores of the candidate priors S' AWG Greater than t i Only then can it be identified as a positive sample to avoid low-scoring samples being considered as positive samples; dynamic threshold t i Defined as:

[0057]

[0058] In the first stage, given a dynamic threshold t i , the similarity score is greater than t i The feature points are assigned as positive sample labels, and the remaining feature points are assigned as negative sample labels, thereby obtaining the assignment result P1; then, in the second stage, in order to avoid that some gts are not assigned enough positive samples, a cyclic label assignment process is introduced to further expand the number of positive samples; specifically, before entering the cyclic process, all Gaussian priors are transformed, that is, the effective radius er n Multiply by a reduction factor σ to provide a new Gaussian prior for the subsequent allocation process; then, repeat the sample selection process of the first stage to add an appropriate number of positive samples; after the second label allocation, the second allocation result P2 is obtained, and then the final allocation result is obtained .

[0059] The present invention also provides a small target detection system based on adaptive Gaussian learning label allocation, including a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the above method can be implemented.

[0060] Compared with the prior art, the present invention has the following beneficial effects: the method and system can accurately measure the difference between different candidate priors and the true frame, and effectively avoid low-score samples in the positive and negative sample screening process, further improving the detection performance of small targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is an architecture diagram of a small target detection network model based on adaptive Gaussian learning label allocation in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0063] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0064] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0065] This embodiment provides a small target detection method based on adaptive Gaussian learning label allocation, including the following steps:

[0066] 1) Obtain the training data set and training set annotation files.

[0067] 2) A small target detection network model based on adaptive Gaussian learning label allocation is constructed, and the small target detection network model includes a benchmark detector, an adaptive weighted Gaussian metric module and a dynamic cyclic label allocation module; the classification feature map and the positioning feature map are extracted by the benchmark detector for classification tasks and positioning tasks respectively; for the positioning feature map, the adaptive weighted Gaussian metric module is used to perform two-dimensional Gaussian modeling on the true box and feature points to obtain the Gaussian distribution corresponding to the true box and feature points, and then the adaptive weighted Gaussian metric is used to measure the similarity of the two Gaussian distributions to more accurately measure the similarity between the Gaussian candidate box and the true box; in the label allocation process, the dynamic cyclic label allocation module is used to dynamically set the threshold for positive and negative sample screening, so that the positive samples in each round of training have sufficient quantity and high quality to obtain more accurate label allocation results; the adaptive weighted Gaussian metric module and the dynamic cyclic label allocation module can effectively improve the small target detection performance and can simultaneously adapt to two mainstream benchmark detectors: anchor-free (without anchor box) and anchor-based (with anchor box).

[0068] 3) Use the trained small target detection network model to detect targets and realize small target detection.

[0069] In this embodiment, a training data set and a corresponding training set annotation file are obtained from an official open source data set, and the training data set is an AI-TOD-v2 small target detection data set.

[0070] Figure 1 is an architecture diagram of a small target detection network model based on adaptive Gaussian learning label allocation in this embodiment. Figure 1 As shown in Figure 1, the processing process of the small target detection network model on the input data is as follows.

[0071] A) A classification feature map and a positioning feature map are obtained through a reference detector, and then the positioning feature map is input into an adaptive weighted Gaussian metric module. The adaptive weighted Gaussian metric module uses the annotation information of the real frame and the feature point receptive field to perform two-dimensional Gaussian modeling on the receptive fields of all feature points on the positioning feature map and the real frame. The Gaussian distribution of the receptive fields of all the obtained feature points is respectively calculated with the Gaussian distribution of the real frame through the Wasserstein distance (WD) to calculate the similarity. At the same time, the size of the overlapping area between the Gaussian distribution of each feature point and the Gaussian distribution of the real frame is calculated as an adaptive weight, which is used to adjust the priority of each feature point and provide each feature point with a more accurate similarity score with the real frame.

[0072] In this embodiment, the specific implementation method of the adaptive weighted Gaussian metric module is as follows.

[0073] For standard convolutional neural networks, the theoretical receptive field refers to the size of the area mapped on the original image by the feature points on the feature map output by each layer, that is, the range that the feature points can "see". The higher (deeper) the layer of feature points, the more comprehensive the original image information described, and the more semantic information can be expressed. For each feature point, on the basis of the theoretical receptive field, the radius of the receptive field is reduced by half, and the new receptive field obtained is called the effective receptive field. In order to better measure the similarity between different candidate priors and the true box (gt), the present invention calculates the adaptive Gaussian weighted similarity score between the effective receptive field of the feature point on the positioning feature map and gt, and uses this score to replace the traditional intersection-over-union (IoU); first, the effective receptive field and gt are modeled in two-dimensional Gaussian, and the distance between the two Gaussian distributions is calculated using WD, thereby avoiding the extreme sensitivity of the traditional IoU due to the small number of pixels occupied by small targets, and the inability of some candidate priors and gt to accurately measure the similarity between each prior and gt when there is no intersection. Then, under the premise of two-dimensional Gaussian modeling, the priority of each feature point is adaptively adjusted by calculating the overlapping area score between each feature point and gt; the specific method is as follows:

[0074] First, calculate the n Layer Theory Receptive Field tr n :

[0075]

[0076] in, tr n It is n The theoretical receptive field of each feature point in the convolutional layer; k n and s n are the convolution kernel size and n The stride of the convolution operation at the layer.

[0077] For the effective receptive field, firstly, each feature point ( x n , y n ) is set to the mean vector of its standard two-dimensional Gaussian distribution ; Since the effective receptive field of each feature point is a subset of the theoretical receptive field, half of the theoretical receptive field radius is expressed as er n , set as the radius of the effective receptive field, and use er n The square of the two-dimensional Gaussian distribution constitutes the covariance For the real frame gt ( xg , y g , w g , h g ), the center point of each gt ( x g , y g ) is specified as the mean vector , and use ( w g / 2, h g / 2) squared constitutes the covariance in a two-dimensional Gaussian distribution ; The effective receptive field and gt are modeled as Gaussian distributions respectively and ; , , and They are respectively expressed as follows:

[0078]

[0079]

[0080] For the Gaussian distribution of effective receptive field and gt and , the second-order Wasserstein distance between them is expressed as follows:

[0081]

[0082] Through D wd By normalizing, its value range can be controlled within (0,1), thus giving a value range similar to that of IoU; after normalization, a new normalized metric score is obtained S d , which is expressed as:

[0083]

[0084] Among them, β represents the adjustment factor in the normalization process.

[0085] Furthermore, an adaptive Gaussian weight is designed for WD using the overlapping area of ​​Gaussian distribution ( w Gauss The specific weight design is as follows:

[0086] First, the covariance matrix of the overlapping area of ​​the two Gaussian distributions is calculated by the Kalman filter , as shown below:

[0087]

[0088] Based on the covariance matrix , the overlapping area between two Gaussian distributions A pt It can be expressed as:

[0089]

[0090] Furthermore, the overlap area fraction S o It can be expressed as:

[0091]

[0092] in, A e represents the area of ​​the Gaussian prior, A g represents the area of ​​gt, S o It can be further expressed as:

[0093]

[0094]

[0095]

[0096] In the above formula S o The value range of is (0, 1 / 3); therefore, a condition factor α is introduced to normalize the value range to (0, 1), which is consistent with S d Stay consistent; final S o It can be expressed as follows:

[0097]

[0098] Therefore, in order to dynamically adjust the similarity score, based on the overlap area S o , adaptive Gaussian weight w Gauss It can be expressed as:

[0099]

[0100] Among them, n is a hyperparameter used to control different w Gauss relative gap.

[0101] Then, the adaptive weighted Gaussian metric can be defined as ,Right now

[0102]

[0103] In order to distinguish between different Gaussian priors S AWG , further adding an exponential modulation factor γ to expand S AWG Gap; Final Adaptive Gaussian Weighted Similarity Score S' AWG It can be expressed as:

[0104] .

[0105] B) All similarity scores obtained from the adaptive weighted Gaussian metric module are input into a dynamic cyclic label assignment module, which assigns positive sample labels or negative sample labels to all candidate feature points by setting a dynamic threshold during the network training phase to avoid some low-score samples from being assigned as positive samples.

[0106] In this embodiment, the specific implementation method of the dynamic cyclic label allocation module is as follows.

[0107] Using the similarity score obtained from the adaptive weighted Gaussian metric module S' AWG Obtain the score matrix between Gaussian prior and gt; for the score matrix, first select the first k priors for gt to prevent the introduction of too many low-score samples from affecting the setting of subsequent thresholds; the positive sample candidate set T(i) of the i-th gt can be expressed as:

[0108]

[0109] in, Indicates the number of candidate samples for the i-th gt.

[0110] For a set T(i), by calculating the mean of all scores in the set m i and standard deviation sd i To set the dynamic threshold t i , the two can be expressed as:

[0111]

[0112]

[0113] For the i-th gt, only the similarity scores of the candidate priors S' AWG Greater than ti Only then can it be identified as a positive sample to avoid low-score samples being considered as positive samples; dynamic threshold t i Defined as:

[0114]

[0115] In the first stage, given a dynamic threshold t i , the similarity score is greater than t i The feature points are assigned as positive sample labels, and the remaining feature points are assigned as negative sample labels, thereby obtaining the assignment result P1; then, in the second stage, in order to avoid that some gts are not assigned enough positive samples, a cyclic label assignment process is introduced to further expand the number of positive samples; specifically, before entering the cyclic process, all Gaussian priors are transformed, that is, the effective radius er n Multiply by a reduction factor σ to provide a new Gaussian prior for the subsequent allocation process, where σ is set to 0.9 by default; then, repeat the sample selection process of the first stage to add an appropriate number of positive samples; after the second label allocation, the second allocation result P2 is obtained, and then the final allocation result is obtained .

[0116] This embodiment also provides a small target detection system based on adaptive Gaussian learning label allocation, including a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the above method can be implemented.

[0117] In this embodiment, the AI-TOD-v2 dataset is used to use DetectoRS as the benchmark detector for comparative verification. Table 1 shows the comparison results of the method proposed in the present invention and other small target detection methods on the AI-TOD-v2 dataset. It can be seen from Table 1 that the method proposed in the present invention has higher accuracy and robustness than other small target detection methods, which is specifically reflected in AP, AP 50 、AP vt 、AP t and AP s Best (AP stands for Average Precision in the field of target detection and is used as the main evaluation indicator. AP 50 It represents the average precision under the premise of IoU threshold of 0.5, AP vt 、AP t 、AP s and AP m They represent the average accuracy of detecting objects of different areas, specifically, , , and .

[0118] Table 1

[0119]

[0120] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0121] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0122] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0123] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0124] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. A small target detection method based on adaptive Gaussian learning label allocation, characterized in that: The following steps are involved: 1) Obtain training data set and training set annotation files; 2) constructing a small target detection network model based on adaptive Gaussian learning label allocation, wherein the small target detection network model includes a reference detector, an adaptive weighted Gaussian metric module and a dynamic cyclic label allocation module; The classification feature map and the positioning feature map are extracted by the benchmark detector for classification and positioning tasks respectively; for the positioning feature map, the adaptive weighted Gaussian measurement module is used to perform two-dimensional Gaussian modeling on the real box and feature points to obtain the Gaussian distribution corresponding to the real box and feature points, and then the adaptive weighted Gaussian measurement is used to measure the similarity of the two Gaussian distributions, and the adaptive Gaussian weight is designed using the overlapping area of ​​the two Gaussian distributions to dynamically adjust the similarity distribution to more accurately measure the similarity between the Gaussian candidate box and the real box; in the label assignment process, all similarity scores obtained from the adaptive weighted Gaussian measurement module are input into the dynamic cyclic label assignment module, and the dynamic cyclic label assignment module is used to dynamically set the threshold for positive and negative sample screening, so that the positive samples in each round of training have sufficient quantity and high quality to obtain more accurate label assignment results; The small target detection network model is trained using the training data set and training set annotation files to obtain generalizable model parameters. 3) Use the trained small target detection network model to detect targets and realize small target detection.

2. The method for detecting small targets based on adaptive Gaussian learning label allocation according to claim 1, characterized in that: In step 1), a training data set and a corresponding training set annotation file are obtained from the official open source data set, wherein the training data set is the AI-TOD-v2 small target detection data set.

3. The method for detecting small targets based on adaptive Gaussian learning label allocation according to claim 1, characterized in that: The specific implementation method of step 2) includes: A) obtaining a classification feature map and a positioning feature map through a reference detector, and then inputting the obtained positioning feature map into an adaptive weighted Gaussian measurement module, wherein the adaptive weighted Gaussian measurement module performs two-dimensional Gaussian modeling on all feature points on the positioning feature map and the real frame using the annotation information of the real frame and the feature point receptive field, and calculates the similarity between the Gaussian distribution of all the obtained feature points and the Gaussian distribution of the real frame through the Wasserstein distance, and at the same time calculates the size of the overlapping area between the Gaussian distribution of each feature point and the Gaussian distribution of the real frame as an adaptive weight, which is used to adjust the priority of each feature point, and provide each feature point with a more accurate similarity score with the real frame; B) All similarity scores obtained from the adaptive weighted Gaussian metric module are input into a dynamic cyclic label assignment module, which assigns positive sample labels or negative sample labels to all candidate feature points by setting dynamic thresholds during the network training and inference stages to avoid some low-scoring samples from being assigned as positive samples.

4. The method for detecting small targets based on adaptive Gaussian learning label allocation according to claim 3, characterized in that: The implementation method of the adaptive weighted Gaussian metric module is: In order to measure the similarity between different candidate priors and the true box gt, the adaptive weighted Gaussian measurement module calculates the adaptive Gaussian weighted similarity score between the effective receptive field of the feature point on the positioning feature map and gt, and uses this score to replace the traditional intersection-over-union ratio IoU; first, two-dimensional Gaussian modeling is performed on the effective receptive field and gt, and the distance between the two Gaussian distributions is calculated using the Wasserstein distance; then, under the premise of two-dimensional Gaussian modeling, the priority of each feature point is adaptively adjusted by calculating the overlapping area score between each feature point and gt; the specific method is as follows: First, the theoretical receptive field tr of the nth layer is calculated as follows n : Among them, tr n is the theoretical receptive field of each feature point in the nth convolutional layer; k n is the convolution kernel size of the nth convolutional layer, s i is the stride of the i-th convolution operation; For the effective receptive field, firstly, each feature point (x n ,y n ) is set to the mean vector μ of its standard two-dimensional Gaussian distribution. e ; Since the effective receptive field of each feature point is a subset of the theoretical receptive field, half of the theoretical receptive field radius is denoted as er n , set to the radius of the effective receptive field, and use er n The square of the two-dimensional Gaussian distribution constitutes the covariance Σ e ; For the real box gt(x g ,y g ,w g ,h g ), the center point (x g ,y g ) is specified as the mean vector μ g , and use (w g / 2,h g / 2) constitutes the covariance ∑ in the two-dimensional Gaussian distribution g ; The effective receptive field and gt are modeled as Gaussian distributions respectively and μ e ,Σ e , μ g and∑ g They are respectively expressed as follows: For the effective receptive field and Gaussian distribution of gt and The second-order Wasserstein distance between them is expressed as follows: Through D wd Normalize it and control its value range within (0,1), thus giving a value range similar to IoU; after normalization, a new normalized metric score S is obtained d , which is expressed as: Among them, β represents the adjustment factor in the normalization process; Furthermore, the Gaussian distribution overlap area is used to design an adaptive Gaussian weight w for the Wasserstein distance. Gauss ; Specifically: First, the Kalman filter is used to calculate the covariance matrix ∑ of the overlapping area of ​​the two Gaussian distributions pt , as shown below: ∑ pt =S e -S e (∑ e +∑ g ) -1 ∑ e Based on the covariance matrix ∑ pt , the overlapping area A between the two Gaussian distributions pt It is expressed as: Overlap area fraction S o It is expressed as: Among them, A e represents the area of ​​the Gaussian prior, A g represents the area of ​​gt, S o Further expressed as: In the above formula, S o The value range of is (0,1 / 3); therefore, the condition factor α is introduced to normalize the value range to (0,1), which is consistent with S d Stay consistent; the final S o ' means as follows: In order to dynamically adjust the similarity score, based on S o ', adaptive Gaussian weight w Gauss It is expressed as: w Gauss =(1+S o ′) q Among them, q is a hyperparameter used to control different w Gauss relative gap; Then, the adaptive weighted Gaussian metric is defined as S AWG =w Gauss ×S d ,Right now In order to distinguish S between different Gaussian priors AWG , further adding an exponential modulation factor γ to expand S AWG Gap; Final adaptive Gaussian weighted similarity score S' AWG It is expressed as:

5. The method for detecting small targets based on adaptive Gaussian learning label allocation according to claim 3, characterized in that: The implementation method of the dynamic cyclic label allocation module is: Using the similarity score S' obtained from the adaptive weighted Gaussian metric module AWG Obtain the score matrix between Gaussian prior and gt; for the score matrix, first select the first k priors for gt to prevent the introduction of too many low-score samples from affecting the setting of subsequent thresholds; the positive sample candidate set T(i) of the i-th gt is expressed as: Among them, N i Indicates the number of candidate samples for the i-th gt; For a set T(i), by calculating the mean m of all scores in the set i and standard deviation sd i To set the dynamic threshold t i , the two are respectively expressed as: For the i-th gt, only the similarity score S' of the candidate prior AWG Greater than t i can be identified as positive samples to avoid low-scoring samples being considered positive samples; dynamic threshold t i Defined as: t i =m i +sd i In the first stage, given a dynamic threshold t i , the similarity score is greater than t i The feature points of are assigned as positive sample labels, and the remaining feature points are assigned as negative sample labels, thereby obtaining the assignment result P1; then, in the second stage, in order to avoid that some gts are not assigned enough positive samples, a cyclic label assignment process is introduced to further expand the number of positive samples; specifically, before entering the cyclic process, all Gaussian priors are transformed, that is, the effective radius er n Multiply by a reduction factor σ to provide a new Gaussian prior for the subsequent allocation process; then, repeat the sample selection process of the first stage and add an appropriate number of positive samples; after the second label allocation, obtain the second allocation result P2, and then obtain the final allocation result P = P1 ∪ P2.

6. A small target detection system based on adaptive Gaussian learning label allocation, characterized in that: The method comprises a memory, a processor, and computer program instructions stored in the memory and executable by the processor. When the processor executes the computer program instructions, the method according to any one of claims 1 to 5 can be implemented.

Images