Aviation infrared small target detection optimization method based on neural network

Abstract

The invention discloses an aviation infrared small target detection optimization method based on a neural network, and relates to the technical field of target detection, and the method comprises the following steps: constructing a re-parameterized Gaussian-Wasserstein network which comprises a backbone network, a neck network and a head network; the heavy parameterized Gaussian-Wasserstein network is trained; the trained re-parameterized Gaussian-Wasserstein network is adopted for detection, and a small target detection result is output; the backbone network adopts a backbone structure of a re-parameterization style, the neck network is designed with efficient re-parameterization convolution, and the head network is designed with an improved head loss function based on a Wasserstein distance and a minimum point distance. According to the method, the re-parameterized Gaussian-Wasserstein network is constructed, and the backbone network, the neck network and the head network are respectively improved, so that the network parameter quantity and the calculation complexity are reduced, and the infrared small target detection precision and robustness are improved.

Description

Technical Field

[0001] This application relates to the field of target detection technology, and in particular to an optimized method for detecting small airborne infrared targets based on neural networks. Background Technology

[0002] Airborne infrared image processing has been widely explored in fields such as intelligent transportation, public safety, environmental monitoring, and military defense. However, compared with visible light images, it lacks color and relatively fine texture information. Therefore, existing airborne infrared small target detection methods generally suffer from problems such as small target size, low contrast, sparse target features, and severe occlusion when processing small targets on roads, leading to low detection accuracy and efficiency. Especially in complex environments, detection systems are prone to missed detections, false detections, and slow processing speeds. How to effectively reduce model computational complexity while ensuring high-accuracy detection and improve robustness in environments with small targets, low contrast, and multi-target interference remains an urgent technical problem to be solved. Therefore, research is needed on optimized methods for airborne infrared small target detection.

[0003] In the prior art, Chinese patent CN118918308A discloses an infrared small target detection method based on multi-scale feature fusion. The method involves inputting an infrared image into a backbone network for deep feature extraction; using a path aggregation network framework to collect feature information from different levels into a collection module; and using average pooling and bilinear interpolation to unify the feature map scale. A fusion module integrates features from various levels to generate global information, and an SE module adaptively adjusts feature weights to enhance important features and suppress secondary features. An injection module combines global information with local features to enhance the detection capability for small targets. Finally, an anchor-free method based on center points is used to regress the original image and output the target location information.

[0004] However, although the aforementioned existing technologies enhance features through multi-scale fusion and SE modules, the superposition of multiple modules leads to structural redundancy, the SE modules are not targeted enough in enhancing the features of small targets, the number of network parameters and computational complexity need to be reduced, and the accuracy and robustness of infrared small target detection need to be improved. Summary of the Invention

[0005] This application provides an optimized method for detecting small targets in airborne infrared systems based on neural networks, which addresses the problems of reducing the number of network parameters and computational complexity, and improving the accuracy and robustness of infrared small target detection in existing airborne infrared small target detection technologies.

[0006] On the one hand, this application provides an optimized method for detecting small airborne infrared targets based on neural networks, including the following steps: Step 1: Based on the YOLOv8 detection framework, construct a reparameterized Gaussian-Wotherstein network, which includes a backbone network, a neck network, and a head network.

[0007] Step 2: Obtain the aerial infrared image dataset and train the reparameterized Gaussian-Wotherstein network.

[0008] Step 3: Obtain the actual aerial infrared image to be detected, and use the trained reparameterized Gaussian-Wotherstein network to detect the aerial infrared image, outputting the small target detection result.

[0009] The backbone network adopts a reparameterized backbone structure, the neck network is designed with efficient reparameterized convolutions, and the head network is designed with an improved head loss function based on Wasserstein distance and minimum point distance.

[0010] In one possible implementation, the backbone structure is based on the modified CSPDarknet53 structure in the YOLOv8 detection framework, and a multi-branch convolutional module is introduced in the latter part of feature extraction to form a reparameterized backbone structure.

[0011] In one possible implementation, the multi-branch convolutional module consists of several improved RepConv units, each of which has a normalized branch added.

[0012] In one possible implementation, the backbone structure retains the CBS module and spatial pyramid pooling module of the modified CSPDarknet53 structure.

[0013] In one possible implementation, the efficient reparameterized convolution is set on the feature output side of the neck network, and the efficient reparameterized convolution consists of a plurality of ERC residual units.

[0014] In one possible implementation, the ERC residual unit includes: performing an addition operation on two two-dimensional convolutional branches and a direct pass branch, followed by sequentially passing them through a normalization layer and an activation layer, and multiplying the output of the normalization layer with the output of the activation layer.

[0015] In one possible implementation, the improved head loss function is calculated by subtracting the fusion-type cross-union ratio.

[0016] The fusion-type cross-union ratio is obtained by weighted summation of the Gauss-Wotherstein distance cross-union ratio and the minimum point distance cross-union ratio.

[0017] In one possible implementation, an adaptive contrast enhancement method is introduced for preprocessing in both steps two and three. Step two involves preprocessing the images in the aviation infrared image dataset, while step three involves preprocessing the aviation infrared image to be detected.

[0018] In one possible implementation, in step two, the trained reparameterized Gaussian-Wotherstein network is evaluated at multiple levels using accuracy metrics, average precision metrics, mean average precision metrics, and overlap metrics.

[0019] The neural network-based optimization method for detecting small airborne infrared targets in this application has the following advantages: By constructing a reparameterized Gaussian-Wotherstein network and improving the backbone, neck, and head networks respectively, the detection accuracy and robustness of infrared small targets are improved while reducing the number of network parameters and computational complexity.

[0020] The proposed multi-branch convolutional module consists of several improved RepConv units, each of which has a normalized branch added, reducing the parameter dimension and improving feature extraction efficiency.

[0021] The proposed efficient reparameterized convolution setting is located on the feature output side of the neck network. The efficient reparameterized convolution consists of several ERC residual units, which improves the multi-scale feature fusion efficiency of the network and achieves a balance between network parameters and detection accuracy.

[0022] The proposed improved head loss function is calculated by subtracting the fusion cross-union ratio from one. The fusion cross-union ratio is obtained by weighted summation of the Gauss-Wotherstein distance cross-union ratio and the minimum point distance cross-union ratio. It covers both single small target detection and densely occluded small target detection, improves scene adaptability and detection robustness, and enhances network training optimization efficiency.

[0023] By introducing an adaptive contrast enhancement method for preprocessing, we can not only highlight the brightness and edge information of small targets and enhance the network's recognition ability in low-resolution areas, but also automatically adjust the image contrast under complex environments such as flight altitude, shooting angle changes, nighttime, and foggy weather, thereby improving the robustness of detection. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 A flowchart illustrating an optimized method for detecting small aerial infrared targets based on a neural network, provided in an embodiment of this application; Figure 2 A schematic diagram of the overall structure of the reparameterized Gaussian-Wotherstein network provided in the embodiments of this application; Figure 3 A schematic diagram of the structure of the improved RepConv unit and the multi-branch convolution module provided in the embodiments of this application; Figure 4 A schematic diagram of the reparameterized style backbone structure provided in the embodiments of this application; Figure 5 This is a schematic diagram of the structure of the ERC residual unit provided in the embodiments of this application; Figure 6 This is a schematic diagram of the structure of the efficient reparameterized convolution provided in the embodiments of this application; Figure 7 This is a schematic diagram of the neck network structure provided in an embodiment of this application. Detailed Implementation

[0026] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0027] like Figure 1 As shown in the figure, this application provides an optimization method for detecting small airborne infrared targets based on neural networks, including the following steps: Step 1: Based on the YOLOv8 detection framework, construct a reparameterized Gaussian-Wotherstein network, which includes a backbone network, a neck network, and a head network.

[0028] Step 2: Obtain the aerial infrared image dataset and train the reparameterized Gaussian-Wotherstein network.

[0029] Step 3: Obtain the actual aerial infrared image to be detected, and use the trained reparameterized Gaussian-Wotherstein network to detect the aerial infrared image, outputting the small target detection result.

[0030] The backbone network adopts a reparameterized backbone structure, the neck network is designed with efficient reparameterized convolutions, and the head network is designed with an improved head loss function based on Wasserstein distance and minimum point distance.

[0031] Specifically, such as Figure 2 As shown, the structure of the backbone, neck, and head of a reparameterized Gaussian-Wotherstein network is illustrated in the diagram. Figure 2 The right half of the diagram shows the basic structure of the main modules on the right, while the left half shows the basic structure of the main modules on the right, including the CBS module (CBS), the C2f module (C2f), upsampling, the efficient reparameterized convolution (ERC block), the multi-branch convolution module (iRepBlock), the bottleneck layer (Bottleneck), the head, and the fusion-type crossover ratio (GWPIoU function) (GWPIoU).

[0032] The backbone network is used to extract features from the image, the neck network is used for feature fusion, and the head network is used to output the detection results.

[0033] For example, the backbone structure is based on the modified CSPDarknet53 structure in the YOLOv8 detection framework, and a multi-branch convolution module is introduced in the latter half of the feature extraction to form a reparameterized backbone structure.

[0034] For example, the multi-branch convolution module consists of several improved RepConv units, each of which has a normalized branch added.

[0035] Specifically, the multi-branch convolutional module combines the advantages of the InceptionNet and ResNet network architectures, reduces the parameter dimensionality through an improved RepConv unit, and improves feature extraction efficiency by using multiple improved RepConv units.

[0036] like Figure 3 As shown, Figure 3 The left half is an improved RepConv unit (denoted as iRepConv), and the right half is a multi-branch convolutional module (denoted as iRepBlock). To address the gradient vanishing problem, this embodiment adds a normalized branch ( Figure 2The Batch Normalization (BN) layer directly transmits the input to the output layer to simulate the ResNet network architecture. Two-dimensional 1×1 convolutions are used for dimensionality reduction and rectified linear activation, significantly reducing computational costs. Several improved RepConv units form a multi-branch convolutional module; the number of improved RepConv units, n, varies depending on the task. In this embodiment, n equals 2. These preparations alter the original structure of the modified CSPDarknet53 architecture. By using a multi-branch convolutional module, the number of parameters is reduced without sacrificing performance, resulting in a new backbone structure called the reparameterized style backbone structure. Through these operations, even when the target is not readily apparent, the most useful features can still be obtained from aerial infrared images.

[0037] For example, the backbone structure retains the CBS module and spatial pyramid pooling module of the modified CSPDarknet53 structure.

[0038] Specifically, the main structure of the reparameterized style is as follows: Figure 4 As shown, the modified CSPDarknet53 structure retains the CBS module and the Spatial Pyramid Pooling (SPFF) module, thus preserving the original advantages. A multi-branch convolutional module is introduced in the latter part of feature extraction to reduce parameters and improve performance.

[0039] For example, the efficient reparameterized convolution is set on the feature output side of the neck network, and the efficient reparameterized convolution consists of a number of ERC residual units.

[0040] For example, the ERC residual unit includes: performing an addition operation on two two-dimensional convolutional branches and a direct pass branch, and then sequentially passing them through a normalization layer and an activation layer, with the output of the normalization layer being multiplied by the output of the activation layer.

[0041] Specifically, such as Figure 5 and Figure 6The diagrams show the structural schematics of ERC residual units and efficient reparameterized convolutions. ERC residual units (denoted as ERC units) avoid the vanishing gradient problem by directly passing the input to the output (corresponding to the direct pass branch). A 1×1 2D convolution (Conv2d 1×1) provides dimensionality reduction and rectified linear activation. Normalization layers (BN) and activation layers (Sigmoid activation) are then used to obtain ERC residual units. Each ERC residual unit forms the basis for constructing the efficient reparameterized convolution, and the number of ERC residual units depends on the training conditions. In this embodiment, four ERC residual units constitute an upsampling and C2f structure utilizing YOLOv8. Therefore, multi-scale feature fusion improves with increasing detection accuracy. Considering the mutual constraints between parameter reduction and detection accuracy, this embodiment places the efficient reparameterized convolution on the feature output side of the neck network, such as... Figure 7 The diagram shows the structure of the neck network.

[0042] Efficient reparameterized convolutions leverage the positional advantage within the neck network, not only outputting feature information to head1 but also influencing head2 and head3 by outputting feature information to the CBS structure. This structural optimization of the neck network plays a crucial role in processing infrared images with low contrast and few features.

[0043] For example, the improved head loss function is calculated by subtracting the fusion-type cross-union ratio.

[0044] The fusion-type cross-union ratio is obtained by weighted summation of the Gauss-Wotherstein distance cross-union ratio and the minimum point distance cross-union ratio.

[0045] Specifically, since most road objects detected in UAV aerial images are small in size, the accuracy of general deep learning detectors is low. The lack of color and texture features in infrared images exacerbates this problem. Therefore, this application redesigns an improved head loss function and constructs its Cross-Union Ratio (CUI) calculation method. Traditional CUI calculation methods are not very useful, especially in cases of overlap and occlusion. In traditional CUI calculations, when the detection area of ​​a small target changes slightly relative to the ground truth, the calculated CUI and loss function change significantly, reducing detection accuracy. This application proposes an improved head loss function based on Wasserstein distance and minimum point distance, calculated by subtracting a fused CUI. The fused CUI (GWPIoU) is obtained by weighted summation of the Gaussian-Wotherstein distance CUI (NWDIoU) and the minimum point distance CUI (MPDIoU). The Gaussian-Wotherstein distance CUI is used to address the problem of detecting small targets using statistical methods, while the minimum point distance CUI is used to address the problem of overlapping targets.

[0046] The intersection-union ratio of the Gauss-Wotherstein distance is calculated as follows: .

[0047] in, is a Gaussian distributed distance matrix, and c is a constant usually associated with the dataset.

[0048] Minimum point distance intersection-union ratio increases the detection rate of occluded or overlapping small objects because it is related to the coordinates of the predicted region, the ground reality region, and the length and width of the image, calculated as follows: .

[0049] in, This is the actual ground area. It is the prediction region. and They are respectively with the region and The distances between the top left and bottom right corners are calculated as follows: .

[0050] The formula for calculating the crossover ratio (CRO) of a fusion-type crossover is as follows: .

[0051] in, It is a constant, with a value range of (0-1). It is selected based on the dataset; the smaller the object, the better. The larger the value (0.9 in this example), the better. This is a constant (0.1 in this embodiment). Based on loss function theory, the head loss function is improved. The calculation is as follows: .

[0052] Specifically, in step two, when selecting the aerial infrared image dataset, since private datasets have the disadvantage of being limited to a single shooting scene and other researchers being unable to conduct subsequent comparative studies, this embodiment selects two public datasets and uses the HIT-AUV dataset and the DroneVehicle dataset to train the reparameterized Gauss-Wotherstein network.

[0053] The HIT-UAV dataset contains 2898 infrared images from roads, parking lots, and schools. This dataset covers a variety of complex scenes and multiple target categories, making it more effective at testing the robustness of detection algorithms compared to multi-target datasets for single scenes. The targets include people, cars, bicycles, and other vehicles. People and bicycles are significantly smaller than vehicles and occupy fewer pixels in the infrared images, exhibiting overall small target characteristics. Therefore, this dataset effectively reflects the challenge of small target detection in complex backgrounds and is suitable as a benchmark for validating the performance of the model proposed in this application.

[0054] The DroneVehicle dataset provides a large-scale dual-mode image dataset in visible light and infrared, containing a total of 28,439 paired images for daytime and nighttime conditions (with and without lighting). This dataset covers various typical scenes such as parking lots, urban roads, and overpasses, and annotates five types of road targets: cars, trucks, buses, vans, and freight vehicles. Cars have the most instances, while trucks have the fewest. All targets exhibit small-sized features across different scenes. Due to the diversity of target categories and the complexity of the scenes, this dataset effectively reflects the challenges of small target detection and is therefore suitable as a benchmark for evaluating the performance of the model proposed in this application.

[0055] For example, in both steps two and three, an adaptive contrast enhancement method is introduced for preprocessing. In step two, the images in the aviation infrared image dataset are preprocessed, and in step three, the aviation infrared image to be detected is preprocessed.

[0056] Specifically, in UAV infrared remote sensing images, infrared imaging differs from visible light imaging, often lacking clear texture and edge information. This results in a small grayscale difference between the target and the background, especially in the HIT-AUV and DroneVehicle datasets, where many instance targets have a resolution less than 1 / 10 of the original image, representing typical small targets that easily blend into the background and have blurred edges. To address this, this application introduces an adaptive contrast enhancement method in the data preprocessing stage to dynamically optimize the grayscale distribution of the original infrared image, thereby improving the overall contrast and making the difference between the target and the background more significant. This method not only highlights the brightness and edge information of small targets, enhancing the network's recognition ability in low-resolution areas, but also automatically adjusts image contrast under complex environments such as flight altitude, shooting angle changes, and nighttime or foggy conditions, improving detection robustness. The preprocessed image retains more detailed features and highlights potential small target regions, thus providing better input for the reparameterized Gaussian-Wotherstein network, making it more stable and efficient in subsequent feature extraction and target detection stages.

[0057] For example, in step two, the trained reparameterized Gaussian-Wotherstein network is evaluated at multiple levels using accuracy metrics, average accuracy metrics, mean average accuracy metrics, and overlap metrics.

[0058] Specifically, in this embodiment, to verify the effectiveness of the reparameterized Gaussian-Wotherstein network constructed in this application for infrared small target detection on UAVs, a multi-level performance evaluation of the trained reparameterized Gaussian-Wotherstein network is performed using accuracy metrics, average accuracy metrics, mean average accuracy metrics, and overlap metrics. These metrics not only reflect the accuracy of the detection results but also measure the applicability and robustness of the model under different target categories and complex scenarios.

[0059] Precision index ) is used to measure the true positive rate in the test results. ) and false positives ( The proportion of correctly detected results reflects the percentage of correct detections in the test results. The formula is: .

[0060] in, This indicates the number of detected targets that match the actual labels. This indicates the number of detected targets that do not match the actual labels. Higher accuracy means a lower false alarm rate during prediction and more reliable detection results.

[0061] Average Precision This is used to evaluate the overall detection performance of a specific category of targets. For targets belonging to the same category... Given a set of positive samples, calculate the detection probability of each positive sample and average them, as shown in the following formula: .

[0062] in, They belong to the same category. Positive samples from a road sample Indicates the first The detection accuracy probability of a positive sample. This metric reflects the overall performance of the model in detecting a specific category of objects. For example, in the HIT-UAV dataset, the accuracy probability of detecting "people," "cars," and "bicycles" can be calculated separately. value.

[0063] Mean Average Precision (MAP) () is a measure of a model's overall detection capability across multiple categories. It is defined as the ability to detect all categories. The average is calculated using the following formula: .

[0064] in, This indicates the number of categories contained in the dataset. For the first Average precision for each category. The overall performance of a model across different categories is comprehensively reflected, making it the most commonly used evaluation metric in object detection tasks. In the DroneVehicle dataset, the model needs to simultaneously detect five categories of targets: cars, trucks, buses, freight vehicles, and vans. Its overall performance is measured by this metric. measure.

[0065] Overlap index ( Mean accuracy when the threshold is 0.5 Bounding box regression and target location estimation are important components of small target detection. This embodiment uses the intersection-over-union (IoU) ratio to measure the overlap between the predicted and ground truth bounding boxes. .

[0066] in, For the prediction area, This is the real area. When... When the value is 0.5, the prediction is considered a correct detection, and the result is... The value is called . It is a key indicator for measuring the effectiveness of an algorithm, reflecting both the accuracy of bounding box localization and the model's performance in the overall object detection task.

[0067] By introducing accuracy ( ), average accuracy ( ), mean precision ( ), degree of overlap ( This application employs multi-level performance evaluation metrics such as [list of metrics] to comprehensively and quantitatively analyze the model's performance in single-class object detection, and also to comprehensively measure its overall effectiveness and operational stability in multi-class object detection and complex scenarios. Results show that the HIT-UAV dataset and the DroneVehicle dataset perform well under ERGW-net [performance evaluation criteria]. The scores were 81.5% and 82.5% respectively, higher than other target algorithms such as YOLO8 and YOLO7. Therefore, this application demonstrates significant advantages in small target detection, occlusion scene processing, and detection accuracy in low-contrast infrared images.

[0068] This application embodiment constructs a reparameterized Gaussian-Wotherstein network and improves the backbone network, neck network, and head network respectively, thereby reducing the number of network parameters and computational complexity while improving the detection accuracy and robustness of infrared small targets.

[0069] The proposed multi-branch convolutional module consists of several improved RepConv units, each of which has a normalized branch added, reducing the parameter dimension and improving feature extraction efficiency.

[0070] The proposed efficient reparameterized convolution setting is located on the feature output side of the neck network. The efficient reparameterized convolution consists of several ERC residual units, which improves the multi-scale feature fusion efficiency of the network and achieves a balance between network parameters and detection accuracy.

[0071] The proposed improved head loss function is calculated by subtracting the fusion cross-union ratio from one. The fusion cross-union ratio is obtained by weighted summation of the Gauss-Wotherstein distance cross-union ratio and the minimum point distance cross-union ratio. It covers both single small target detection and densely occluded small target detection, improves scene adaptability and detection robustness, and enhances network training optimization efficiency.

[0072] By introducing an adaptive contrast enhancement method for preprocessing, we can not only highlight the brightness and edge information of small targets and enhance the network's recognition ability in low-resolution areas, but also automatically adjust the image contrast under complex environments such as flight altitude, shooting angle changes, nighttime, and foggy weather, thereby improving the robustness of detection.

[0073] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0074] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. An optimized method for detecting small airborne infrared targets based on neural networks, characterized in that, Includes the following steps: Step 1: Based on the YOLOv8 detection framework, construct a reparameterized Gaussian-Wotherstein network, which includes a backbone network, a neck network, and a head network. Step 2: Obtain the aerial infrared image dataset and train the reparameterized Gaussian-Wotherstein network; Step 3: Obtain the actual aerial infrared image to be detected, and use the trained reparameterized Gaussian-Wotherstein network to detect the aerial infrared image, and output the small target detection result; The backbone network adopts a reparameterized backbone structure, the neck network is designed with efficient reparameterized convolutions, and the head network is designed with an improved head loss function based on Wasserstein distance and minimum point distance.

2. The optimized method for detecting small airborne infrared targets based on neural networks according to claim 1, characterized in that, The backbone structure is based on the modified CSPDarknet53 structure in the YOLOv8 detection framework. A multi-branch convolution module is introduced in the latter part of feature extraction to form a reparameterized backbone structure.

3. The optimized method for detecting small airborne infrared targets based on neural networks according to claim 2, characterized in that, The multi-branch convolution module consists of several improved RepConv units, each of which has a normalized branch added.

4. The optimized method for detecting small airborne infrared targets based on neural networks according to claim 2, characterized in that, The backbone structure retains the CBS module and spatial pyramid pooling module of the modified CSPDarknet53 structure.

5. The optimized method for detecting small airborne infrared targets based on neural networks according to claim 1, characterized in that, The efficient reparameterized convolution is set on the feature output side of the neck network, and the efficient reparameterized convolution consists of several ERC residual units.

6. The optimized method for detecting small airborne infrared targets based on neural networks according to claim 5, characterized in that, The ERC residual unit includes: performing addition operations on two two-dimensional convolutional branches and a direct pass branch, followed by sequentially passing them through a normalization layer and an activation layer, and multiplying the output of the normalization layer with the output of the activation layer.

7. The optimized method for detecting small airborne infrared targets based on neural networks according to claim 1, characterized in that, The improved head loss function is calculated by subtracting the fusion-type cross-union ratio; The fusion-type cross-union ratio is obtained by weighted summation of the Gauss-Wotherstein distance cross-union ratio and the minimum point distance cross-union ratio.

8. The optimized method for detecting small airborne infrared targets based on neural networks according to claim 1, characterized in that, In both steps two and three, an adaptive contrast enhancement method is introduced for preprocessing. In step two, the images in the aviation infrared image dataset are preprocessed, and in step three, the aviation infrared image to be detected is preprocessed.

9. The optimized method for detecting small airborne infrared targets based on neural networks according to claim 1, characterized in that, In step two, the trained reparameterized Gaussian-Wotherstein network is evaluated at multiple levels using accuracy, average precision, mean average precision, and overlap metrics.

Images