Aerial photography infrared small target detection method

Abstract

The invention provides an aerial photography infrared small target detection method, and the method comprises the steps: firstly, designing an FPSC module at a backbone part of a network, and enabling the FPSC module to comprise three expansion convolutions with different expansion coefficients and a shared convolution layer, so as to enhance the capability of a model for capturing multi-scale context information; and secondly, feature extraction is performed on the information of the P2 feature layer through SPD convolution and a designed ECSD module, the SPD convolution is used for enhancing feature information, and the ECSD is used for refining image features, so that the recognition capability of small target details is remarkably improved. Besides, a C2f module is improved, a C2f-IF module is designed, and the efficiency of small target detection is improved by means of identical mapping and jump connection. And finally, solving the problems of expansion and low convergence speed of the anchor frame in a probability distribution modeling mode by using a PIoU loss function. The detection effect visualization graph shows the excellent detection performance of the detection method. According to the method, the infrared target can be quickly and accurately identified, and the conditions of missing detection and false detection are greatly improved.

Description

Technical Field

[0001] The invention relates to the field of artificial intelligence and a target detection method, and in particular to a method for detecting small infrared targets in aerial photography. Background Art

[0002] With the continuous advancement of visual system technology, the functions of drones are becoming increasingly enhanced, and they have been widely used in both civil and military fields. Especially in the military field, micro drones have shown unique advantages due to their small size, light weight, strong concealment and high flexibility, and are widely deployed in modern warfare, mainly for reconnaissance, surveillance and target strike missions. At present, most drones are equipped with visible light imaging systems, which can provide images with rich textures and high resolution, but their imaging effects are easily affected by environmental factors, especially in low light conditions at night, it is difficult to capture clear images. In contrast, infrared imaging systems do not rely on external light sources, can achieve all-weather monitoring, and have good resistance to environmental interference such as rain, snow and smoke, are not affected by local strong light or backlight, and have strong resistance to complex lighting. However, infrared imaging relies on temperature difference to capture images, and can usually only generate single-channel grayscale images, resulting in relatively low resolution and blurred visual effects. In addition, the targets in infrared images are often relatively small, making infrared target detection more difficult than visible light target detection.

[0003] In traditional infrared target detection algorithms, filtering, local contrast measurement and other methods are often used to target the characteristics of infrared targets. These methods can be single-frame processing or sequence-based processing, with the goal of distinguishing noise from targets in the image. These methods can achieve target detection to a certain extent, but they have some obvious shortcomings. First, in complex scenes, target objects may show diverse appearances, shapes and scales. Traditional methods are difficult to accurately identify and locate these target objects, and are prone to false detection and missed detection. In addition, when processing large-scale image data, traditional methods may consume a lot of computing resources and time, resulting in poor real-time performance of the algorithm, which is not applicable in scenarios with high real-time requirements. Therefore, in order to improve the accuracy and real-time performance of infrared target detection, more advanced algorithms need to be designed to meet these challenges. Summary of the invention

[0004] In view of this, the embodiment of the present invention provides an aerial infrared small target detection method to solve the problem that the current detection algorithm has low target detection accuracy for drone infrared aerial images and difficulty in detecting infrared small targets.

[0005] An embodiment of the present invention provides a method for detecting small infrared targets in aerial photography, the method comprising the following steps:

[0006] S1. Obtain the HIT-UAV dataset, which contains more than 2,800 infrared thermal images extracted from 43,470 frames captured by drones in different scenes (schools, parking lots, roads, playgrounds, etc.), including five categories: cars, pedestrians, other vehicles, bicycles, and no attention. Process the dataset to obtain training set, test set, and validation set;

[0007] Furthermore, the specific content of processing the HIT-UAV dataset in S1 is:

[0008] Convert the format in the dataset into YOLO format and adjust the image pixel size to 640×640;

[0009] The data set is divided into training set, test set and validation set according to a certain ratio.

[0010] S2. Build a UAV infrared small target detection model and train the training set on the built model;

[0011] Furthermore, the specific steps of building the UAV infrared small target detection model in S2 are as follows:

[0012] Build the YOLOv8 basic network model;

[0013] The FPSC module is designed in the backbone network part;

[0014] The information of the P2 feature layer is extracted through SPD convolution and the designed ECSD module;

[0015] Improve the C2f module and design the C2f-IF module;

[0016] PIoU is introduced instead of CIoU as a new boundary regression loss function.

[0017] Furthermore, the FPSC module is designed in the backbone network. The FPSC module contains convolutional layers with different expansion rates and shared convolutional layers. The convolutional layers with different expansion rates capture features in the image at different scales, thereby expanding the receptive field of the model. The shared convolutional layer uses a 1×1 convolutional layer to efficiently adjust the number of channels and perform feature fusion. In this way, the FPSC module enhances the model's ability to capture multi-scale contextual information. The specific content of the FPSC module is processed according to the following formula:

[0018] Y=Concat((x,W),Conv2d(y0,W shared ,d i ))

[0019] Y out =Conv(Y,W)

[0020] Where x represents the feature map of the last downsampling output of the backbone network, Concat(·) represents the feature connection operation, Conv(·) represents the convolution operation, and Conv2d represents the use of three 3×3 convolution kernels for dilated convolution with a dilation factor of d. i , d i ∈{1,3,5}, y0 is the first convolution feature map, W and W shared are the convolutional layer and shared convolutional layer parameters, respectively. Y represents the intermediate features of the input features after a series of convolution operations. Y out Represents the final output feature, that is, the result after all convolution and feature fusion operations.

[0021] Furthermore, the information of the received P2 feature layer is extracted through SPD convolution and the designed ECSD module. The ECSD module contains three branches. The detail branch contains 1×31 deep convolution to focus on detail features. The local branch contains 31×1, 31×31 and 1×1 deep convolutions to process medium-scale features. The global branch contains dual-domain (frequency domain and spatial domain) channel attention module (DCAM) and frequency-based spatial attention module (FSAM) to capture the overall pattern. First, the information of the P2 feature layer is processed by SPDConv to obtain enhanced small target information features. Subsequently, these features are passed to the P3 layer for deep fusion processing and then processed by the three branches of the ECSD module. In the global branch, DCAM first performs frequency channel processing (FCA) on the features fused by the P3 layer. The formula is expressed as follows:

[0022]

[0023] Where X Global is the feature after fusion at the P3 layer, F and I are fast Fourier transform and its inverse operation respectively; X FCA represents the output of FCA, W(·) represents the convolutional layer operation, and GAP(·) represents the global average pooling operation; Represents the multiplication operation of the elements. Through Fourier processing, the global features of the image are effectively refined. After global modulation in the spectral domain, the obtained features are further input into the spatial channel attention module (SCA), which is expressed as:

[0024]

[0025] Where X DCAM is the output of DCAM. DCAM only enhances the dual-domain features at the coarse granularity in the channel direction. After the feature information is processed by DCAM, it is processed by FASM. FSAM is a frequency-based attention module applied in the spatial dimension. Its function is to refine the spectrum at the fine-grained level. It can be expressed as:

[0026]

[0027] Where X FSAM is the output of FSAM. In this way, the model can focus on the frequency components with rich information, thereby achieving efficient feature processing. Finally, the information of the three branches is fused through the convolution layer, which significantly improves the model's ability to recognize small target details.

[0028] Furthermore, the C2f module is improved to design the C2f-IF module. The C2f-IF module improves the Bottleneck part of the C2f module with IdentityFormer. IdentityFormer directly transmits these features through identity mapping and skip connections without any additional token mixing or conversion operations, ensuring the direct and efficient flow of feature information. Finally, these unmodified feature sequences are output for further processing and analysis by subsequent layers of the network, thereby maintaining the sensitivity and accuracy of the network to the input data without adding additional computational burden. It is defined as:

[0029] Identity Mapping(X)=X

[0030] In the formula, X is the input feature sequence of the baseline network. This formula means that Identity Mapping does not perform any Token Mixing, that is, it does not perform any operations to enhance the model feature fusion capability, and directly maps the input X to the output X.

[0031] Furthermore, PIoU is introduced instead of CIoU as a new boundary regression loss function. Unlike CIoU, which simply considers the complete intersection between target boxes, PIoU is modeled through probability distribution. Even if there is no overlap between bounding boxes, it can still output an effective loss signal, avoiding the problem of gradient 0, making training more stable. In addition, PIoU uses smooth probability estimation to enable the loss function to still give stable gradient information when the bounding box position is slightly deviated, thereby accelerating the convergence speed. The PIoU loss function expression is:

[0032]

[0033] Where M represents the set of all positive samples; M' is the number of positive samples; b and b' are the real target and predicted target border respectively; PIoU(b, b') represents the ratio of the intersection area to the union area of ​​the real target b and the predicted border b' at the pixel level after binarization. PIoU It is the PIoU value output after calculation.

[0034] S3, use the test set of S1 to test the infrared small target detection model;

[0035] S4, evaluate using evaluation indicators such as precision, recall, average precision, FPS, etc.;

[0036] S5. The measured qualified model processes the UAV infrared image to obtain the target detection result.

[0037] In summary, the beneficial effects of the present invention are as follows:

[0038] 1. According to the special properties of the infrared images taken by drones, the target detection method of the present invention is designed, which can accurately detect the target in the case of complex background and when the target is blocked.

[0039] 2. Aiming at the shortcomings of YOLOv8 in detecting feature information, FPSC is designed to enhance the feature extraction capability of the backbone network.

[0040] 3. By adjusting the network structure and adding a small target detection layer designed for small targets, that is, extracting features from the P2 feature layer through SPD convolution and the designed ECSD module, the network's detection capability for objects of different scales can be improved.

[0041] 4. By using IdentityFormer to improve the C2f module, the C2f-IF module is designed to improve the detection performance of small targets while reducing the number of parameters.

[0042] 5. By using the PIoU loss function to accurately process the target's location information, the error in positioning small targets is reduced and the detection accuracy is improved.

[0043] 6. Based on the test results on the HIT-UAV dataset, the present invention has excellent target detection performance in UAV aerial infrared images. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solution of the embodiment of the present invention, the drawings required for use in the embodiment of the present invention will be briefly introduced below. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work, and these are all within the protection scope of the present invention.

[0045] Figure 1 This is a flow chart of a method for detecting small infrared targets in aerial photography provided by an embodiment of the present invention;

[0046] Figure 2 This is an overall structural diagram of an aerial infrared small target detection method provided by an embodiment of the present invention;

[0047] Figure 3It is a FPSC structure diagram of an aerial infrared small target detection method provided by an embodiment of the present invention;

[0048] Figure 4 It is an ECSD structure diagram of an aerial infrared small target detection method provided by an embodiment of the present invention;

[0049] Figure 5 It is a structural diagram of DCAM and FSAM modules in an ECSD module of an aerial infrared small target detection method provided by an embodiment of the present invention;

[0050] Figure 6 It is a C2f-IF structure diagram of an aerial infrared small target detection method provided by an embodiment of the present invention;

[0051] Figure 7 This is a visual comparison diagram of the detection effect of an aerial infrared small target detection method provided by an embodiment of the present invention and a baseline network. DETAILED DESCRIPTION

[0052] The features and exemplary embodiments of the present invention in various aspects will be described in detail below. In order to make the purpose, technical solutions and advantages of the present invention clearer and easier to understand, the following content will be described in conjunction with the drawings and examples. It should be understood that the specific embodiments described herein are only configured to explain the present invention and are not configured to limit the present invention. For those skilled in the art, the present invention can be implemented without the need for some of these specific details. The following description of the embodiments is only to provide a better understanding of the present invention by illustrating examples of the present invention.

[0053] See also Figure 1 The embodiment of the present invention provides a method for detecting small infrared targets in aerial photography, which is specifically implemented by the following steps:

[0054] S1, obtain the HIT-UAV dataset, which contains more than 2,800 infrared thermal images extracted from 43,470 frames captured by drones in different scenes (schools, parking lots, roads, playgrounds, etc.), including five categories: cars, pedestrians, other vehicles, bicycles, and no attention. The dataset is processed to obtain the training set, test set, and validation set, including the following specific steps:

[0055] (1) Convert the format in the dataset into YOLO format and adjust the image pixel size to 640×640;

[0056] (2) Divide the data set into a training set, a test set, and a validation set according to a certain ratio.

[0057] S2, build a UAV infrared small target detection model, the model is shown in the figure Figure 2As shown in the figure, the Backbone part includes modules such as conv, C2f-IF, FPSC, the Neck part includes C2f-IF, Concat, Upsample, Conv, and the Head part includes Detect. Conv represents the convolution operation, and Concat represents the splicing operation. The C2f module consists of Conv, Split, IdentityFormer, and Concat, where Split represents the segmentation operation, and IdentityFormer represents the bottleneck layer, which aims to reduce the amount of calculation by reducing the number of channels while maintaining or improving the expressiveness of the features. FPSC consists of Conv, Conv2d (two-dimensional convolution layer), and Concat. The Detect detection module consists of Conv, Conv2d (two-dimensional convolution layer), Bbox.loss (bounding box loss), and C1s.loss (category loss). Training the training set on the built model includes the following steps;

[0058] (1) First, the input image of size 640×640×3 is passed through multiple convolutional layers (Conv) and C2f-IF modules in sequence to extract features from different levels. The C2f-IF module ensures the direct and efficient flow of feature information through identity mapping during the feature extraction process, thereby improving the feature expression capability. In this process, the model downsamples the feature map through convolution with a stride of 2, continuously compresses the spatial size of the image, and extracts higher-level semantic features. Finally, the feature map size is further adjusted through the FPSC (Feature Pyramid Shared Conv) module to prepare for subsequent feature fusion.

[0059] Among them, the FPSC module structure is as follows Figure 3 As shown in the figure, convolutional layers with different expansion rates capture features in the image at different scales, thereby expanding the receptive field of the model, and the shared convolutional layer efficiently adjusts the number of channels through a 1×1 convolutional layer for feature fusion.

[0060] Among them, the C2f-IF module structure is as follows Figure 6 As shown in Figure 2, the C2f-IF module is obtained by improving the Bottleneck part of the C2f module using IdentityFormer. IdentityFormer uses a simple identity mapping and adds the input directly to the output through skip connections, thereby alleviating the problem of gradient disappearance or explosion and helping deep networks to train better.

[0061] (2) In the Neck part, SPConv is first used to further process the features extracted by Backbone to generate a high-dimensional feature representation. This part uses a multi-layer feature fusion strategy, such as the combination of Concat (feature concatenation) and C2f-IF modules, to share information between features of different layers, which helps to improve the detection effect of small targets. At the same time, the Upsample operation is used to restore the spatial size of the feature map so that features of different levels can be concatenated on the same spatial scale. ECSD further enhances the spatial selectivity of the features and improves the model's sensitivity to the target position.

[0062] Among them, the ECSD module structure is as follows Figure 4 and Figure 5 As shown in the figure, it enhances the multi-scale feature extraction of the model through three branches: detail branch, local branch and global branch. The detail branch is mainly responsible for processing small signals, while the local branch uses deep convolutions of different sizes (including 1×31, 31×1, 31×31 and 1×1) to process medium-scale features. The global branch introduces the dual-domain channel attention module (DCAM) and the frequency-based spatial attention module (FSAM), which work together to refine the global features of the image. DCAM first performs frequency channel processing (FCA) on the features fused from the P3 layer, and refines the global features of the image through fast Fourier transform (FFT) and its inverse operation (IFT). Then the processed feature information is processed by FASM. FSAM is a frequency-based attention module applied in the spatial dimension to refine the spectrum at a fine-grained level, so that the model can focus on the information-rich frequency components and achieve efficient feature extraction.

[0063] (3) In the Head part, the features are further fused and processed through a series of Concat, C2f-IF and convolution layers, and finally the features are sent to the Detect module for target detection. The Detect module here uses the PIoU loss function. Through smooth probability estimation, the loss function can still give stable gradient information when the position of the bounding box is slightly deviated, thereby accelerating the convergence speed.

[0064] Among them, PIoU is modeled through probability distribution. Even if there is no overlap between bounding boxes, it can still output an effective loss signal, avoiding the problem of gradient being 0, making the training more stable. And PIoU uses smooth probability estimation to make the loss function still give stable gradient information when the bounding box position is slightly deviated, speeding up the convergence speed. The PIoU loss function expression is:

[0065]

[0066] Where M represents the set of all positive samples; M' is the number of positive samples; b and b' are the true target and predicted target bounding box, respectively; PIoU(b, b') represents the ratio of the intersection area to the union area of ​​the true target b and the predicted bounding box b' at the pixel level after binarization.

[0067] S3, uses the test set of S1 to test the infrared small target detection model;

[0068] S4, evaluates using evaluation indicators such as precision, recall, average precision, and FPS;

[0069] S5, the measured qualified model processes the UAV infrared image to obtain the target detection result.

[0070] In order to solve the problems of missed detection and false detection caused by large changes in target scale, dense distribution of small-sized targets and complex background in aerial infrared images of drones, this application proposes an aerial infrared small target detection method. First, the FPSC module is designed to enhance the feature extraction capability of the backbone network. Secondly, a small target detection layer is introduced, that is, the information of the received P2 feature layer is extracted through SPD convolution and the designed ECSD module, which increases the size of the feature map to be detected, improves the network's perception of small targets, introduces the PIoU loss function, reduces the large gradients of extreme samples, and improves the overall performance of the network model.

[0071] The recall rate (Recal1), mAP@0.5 and mAP@0.5:0.95 of the method mentioned in the present invention on the HIT-UAV dataset are 78.7%, 83.6% and 54.7%, respectively, which are 5.0%, 6.3% and 3.6% higher than the original YOLOv8n network, and the number of parameters is reduced by 20%. Experiments show that the improved model can better detect small targets in images and can be effectively applied to various application scenarios of target detection in drone aerial infrared images.

[0072] The experimental environment of the present invention is as follows: This experiment is built on a Windows system, and the deep learning framework used is Pytorch-GPU version 2.4.0. The GPU model used in the experiment is NVIDIA GeForce RTX 4090, with 24GB video memory. The programming language selected is Python 3.8, and the CUDA version is 11.8. All experiments are performed under the same hyperparameters.

[0073] In the YOLO series of models, the main indicators for evaluating its network performance are as follows: Precision (P), Recall (R), and Mean Average Precision (mAP). This experiment uses mAP@0.5 and mAP@0.5:0.95 as performance reference indicators. mAP@0.5% and mAP@0.5:0.95% represent the mAP value when the IoU threshold is 0.5 and the average mAP value when the IoU starts from 50% and increases to 95% with a step size of 0.05, respectively. The larger the mAP, the higher the overall accuracy of the model. The calculation formulas for each indicator are as follows:

[0074]

[0075] Among them, TP represents the number of positive samples correctly detected, that is, true positives; FP represents the number of false samples correctly detected; FN represents the number of true samples incorrectly detected; N represents the number of categories.

[0076] In order to verify that the method proposed in this application has significantly improved the performance of drone aerial image detection, an ablation experiment was conducted on the HIT-UAV dataset. Due to the overall difference between the present invention and YOLOv8n, YOL0v8n was numbered A for experiment. On the basis of A, FPSC, improved small target layer, C2f-IF were added in sequence, and the network structure obtained by PIoU was recorded as B, C, D, and E in sequence. The experimental results are shown in Table 1. After the introduction of the FPSC module, the accuracy has been improved to a certain extent, indicating that the model can better capture the details and complex patterns in the image. After adding the improved small target layer, the accuracy of the model in small target detection has been significantly improved, mAP@0.5 has increased by 3.3%, R% has increased by 1.6%, and the amount of calculation has only increased to 12GFLOPs, which verifies that adding the target detection layer is conducive to the effectiveness of capturing small targets and achieving a balance between computational efficiency and performance; when C2f-IF is added, it is found that under the premise that the performance indicators are improved, the number of parameters and the amount of calculation of the model have been significantly reduced, and the lightweight model has been achieved. After introducing the PIoU loss function, the model can better handle the problems of target score imbalance and bounding box overlap, and mAP@0.5 is improved by 1.5% and R is improved by 2.4%. Experiments show that the improved model can better achieve the target detection task of aerial infrared images.

[0077] Table 1 Comparison of ablation experiment results

[0078] method mAP@0.5 R Gflops Params(M) A 77.3 73.7 8.7 3.2 B 78.0 74.3 8.8 3.4 C 81.3 75.9 12.0 3.2 D 82.1 76.3 10.8 2.5 E 83.6 78.7 10.8 2.5

[0079] In order to verify the superiority of the proposed method in the detection performance of UAV aerial images, the proposed method including the traditional two-stage detection algorithm Faster R-CNN and the YOLO series algorithm were compared on the HIT-UAV dataset. The method comparison experiment is shown in Table 2. Compared with other excellent two-stage target detection methods, the mAP value of the proposed method reached 83.6%, achieving the best detection performance while successfully realizing parameter lightweight, which shows that the model simplification and performance improvement of the proposed method are very significant.

[0080] Table 2 Comparison of experimental results

[0081] Model mAP@0.5 mAP@0.5:0.95 Gflops Params(M) Faster R-CNN 72.7 44.6 206.7 41.2 SSD 70.4 41.5 52.7 24.5 Yolov3-Tiny 75.7 51.4 18.5 8.7 Yolv5s 77.6 51.7 24.1 9.1 Yolov6 77.1 50.5 14.7 16.3 Yolov8n (baseline) 77.3 51.1 8.7 3.2 Yolov8s 81.7 53.6 22.2 5.9 Yolov9s 81.8 53.4 27.4 7.2 Ours 83.6 54.7 10.8 2.5

[0082] By selecting drone aerial images from the HIT-UAV test set for detection, the detection results are as follows: Figure 7 As shown, the left side is the detection effect of the original model, and the right side is the detection effect of the model of the present invention. The present invention can accurately detect and locate targets such as cars and pedestrians under different height conditions and scene backgrounds. For example, the present invention is aimed at the original model's missed detection of pedestrians in pedestrian-sparse scenes, missed detection of edge bicycles in bicycle-dense scenes, and the problem of low accuracy of low-altitude and high-altitude detection of drones. Specifically, under complex backgrounds, the method successfully suppresses the interference of background information such as buildings and trees, thereby achieving accurate positioning and classification of targets. This test result shows that the method of the present invention performs well in actual scenes and has strong adaptability and robustness. It can not only cope with complex backgrounds and different height conditions, but also accurately detect targets of various sizes, making it highly practical and reliable in the task of detecting small infrared targets in drone aerial photography.

[0083] In summary, an embodiment of the present invention provides an aerial infrared small target detection method.

[0084] It should be clear that the present invention is not limited to the specific configuration and processing described above and shown in the figures. For the sake of simplicity, a detailed description of the known method is omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown, and those skilled in the art can make various changes, modifications and additions, or change the order between the steps after understanding the spirit of the present invention.

[0085] It should also be noted that the exemplary embodiments mentioned in the present invention describe some methods or systems based on a series of steps or devices. However, the present invention is not limited to the order of the above steps, that is, the steps can be performed in the order mentioned in the embodiments, or in a different order from the embodiments, or several steps can be performed simultaneously.

[0086] The above is only a specific implementation of the present invention. Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working process of the system, module and unit described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. It should be understood that the protection scope of the present invention is not limited to this. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be covered within the protection scope of the present invention.

Claims

1. A method for detecting small infrared targets in aerial photography, characterized in that: The method comprises the following steps: S1. Obtain the HIT-UAV dataset, which contains more than 2,800 infrared thermal images extracted from 43,470 frames captured by drones in different scenes (schools, parking lots, roads, playgrounds, etc.), including five categories: cars, pedestrians, other vehicles, bicycles, and no attention. Process the dataset to obtain training set, test set, and validation set; S2. Build a UAV infrared small target detection model and train the training set on the built model; S3, use the test set of S1 to test the infrared small target detection model; S4, evaluate using evaluation indicators such as precision, recall, average precision, FPS, etc.; S5. The measured optimal model processes the UAV infrared image and obtains the target detection result.

2. The method for detecting small infrared targets in aerial photography according to claim 1, characterized in that: The target image is processed as described in S1, specifically as follows: Convert the format in the dataset into YOLO format and adjust the image pixel size to 640×640; The data set is divided into training set, test set and validation set according to a certain ratio.

3. The method for detecting small infrared targets by aerial photography according to claim 1, characterized in that: The infrared small target detection model described in S2 is as follows: Build the YOLOv8 basic network model; The FPSC module is designed in the backbone network part; The information of the P2 feature layer is extracted through SPD convolution and the designed ECSD module; Improve the C2f module and design the C2f-IF module; PIoU is introduced instead of CIoU as a new boundary regression loss function.

4. The method for detecting small infrared targets by aerial photography according to claim 3, characterized in that: The FPSC module is designed in the backbone network. The FPSC module contains convolutional layers with different expansion rates and shared convolutional layers. The convolutional layers with different expansion rates capture features in the image at different scales, thereby expanding the receptive field of the model. The shared convolutional layer uses a 1×1 convolutional layer to efficiently adjust the number of channels and perform feature fusion. In this way, the FPSC module enhances the model's ability to capture multi-scale contextual information. The specific content of the FPSC module is processed according to the following formula: Y=Concat((x,W),Conv2d(y0,W shared ,d i )) Y out =Conv(Y,W) Where x represents the feature map of the last downsampling output of the backbone network, Concat(·) represents the feature connection operation, Conv(·) represents the convolution operation, and Conv2d represents the use of three 3×3 convolution kernels for dilated convolution with a dilation factor of d. i , d i ∈{1,3,5}, y0 is the first convolution feature map, W and W shared are the convolutional layer and shared convolutional layer parameters, respectively. Y represents the intermediate features of the input features after a series of convolution operations. Y out Represents the final output feature, that is, the result after all convolution and feature fusion operations.

5. The method for detecting small infrared targets by aerial photography according to claim 3, characterized in that: The information of the received P2 feature layer is extracted through SPD convolution and the designed ECSD module. The ECSD module contains three branches. The detail branch contains 1×31 deep convolution to focus on detail features. The local branch contains 31×1, 31×31 and 1×1 deep convolutions to process medium-scale features. The global branch contains dual-domain (frequency domain and spatial domain) channel attention module (DCAM) and frequency-based spatial attention module (FSAM) to capture the overall pattern. First, the information of the P2 feature layer is processed by SPDConv to obtain enhanced small target information features. Subsequently, these features are passed to the P3 layer for deep fusion processing and then processed by the three branches of the ECSD module. In the global branch, DCAM first performs frequency channel processing (FCA) on the features fused by the P3 layer. The formula is expressed as follows: Where X Global is the feature after P3 layer fusion, F and I are fast Fourier transform and its inverse operation respectively, X FCA represents the output of FCA, W(·) represents the convolutional layer operation, and GAP(·) represents the global average pooling operation; represents the element-wise multiplication operation, which effectively refines the global features of the image through Fourier processing; after global modulation in the spectral domain, the obtained features are further input into the spatial channel attention module (SCA), which is expressed as: Where X DCAM is the output of DCAM. DCAM only enhances the dual-domain features at the coarse granularity in the channel direction. After the feature information is processed by DCAM, it is processed by FASM. FSAM is a frequency-based attention module applied in the spatial dimension. Its function is to refine the spectrum at the fine-grained level. It can be expressed as: Where X FSAM is the output result of FSAM; in this way, the model can focus on the frequency components rich in information, thereby achieving efficient feature processing. The information of the last three branches is fused through the convolutional layer, which significantly improves the model's ability to recognize small target details.

6. The method for detecting small infrared targets by aerial photography according to claim 3, characterized in that: The C2f module is improved and the C2f-IF module is designed. The C2f-IF module improves the Bottleneck part of the C2f module with IdentityFormer. IdentityFormer directly transmits these features through identity mapping and skip connections without any additional token mixing or conversion operations, ensuring the direct and efficient flow of feature information. Finally, these unmodified feature sequences are output for further processing and analysis by subsequent layers of the network, thereby maintaining the network's sensitivity and accuracy to the input data without adding additional computational burden. It is defined as: Identity Mapping(X)=X In the formula, X is the input feature sequence of the baseline network. This formula means that Identity Mapping does not perform any Token Mixing, that is, it does not perform any operations to enhance the model feature fusion capability, and directly maps the input X to the output X.

7. The method for detecting small infrared targets by aerial photography according to claim 3, characterized in that: PIoU is introduced to replace CIoU as a new boundary regression loss function. Unlike CIoU, which simply considers the complete intersection between target boxes, PIoU is modeled through probability distribution. Even if there is no overlap between bounding boxes, it can still output an effective loss signal to avoid the problem of gradient 0, making training more stable. In addition, PIoU uses smooth probability estimation to enable the loss function to still give stable gradient information when the bounding box position is slightly deviated, thereby accelerating the convergence speed. The PIoU loss function expression is: Where M represents the set of all positive samples; M' is the number of positive samples; b and b' are the true target and predicted target bounding box, respectively; PIoU(b, b') represents the ratio of the intersection area to the union area of ​​the true target b and the predicted bounding box b' at the pixel level after binarization.

Images