Infrared small target real-time detection system and method
By utilizing a real-time infrared small target detection system based on the YOLOv8n architecture, and combining a lightweight adaptive feature enhancement unit and an efficient dynamic convolution module with normalized second-order Wasserstein distance and a custom dataset, the system solves the problems of missed detection and false alarms in infrared small target detection, and achieves efficient real-time detection and edge computing deployment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ORIGINAL JIWEI TECHNOLOGY (SHANGHAI) CO LTD
- Filing Date
- 2026-05-07
- Publication Date
- 2026-06-05
AI Technical Summary
Existing technologies struggle to significantly improve the detection accuracy of small infrared targets while maintaining real-time detection speed, addressing issues of missed detections and false alarms, and simultaneously reducing model complexity to adapt to edge computing deployments.
A real-time infrared small target detection system based on the YOLOv8n architecture is adopted, including a lightweight adaptive feature enhancement unit, PANet structure and efficient dynamic convolution module. Normalized second-order Wasserstein distance is used to replace the IoU metric. The system is trained with a custom infrared small target dataset, and the super-resolution preprocessing step is omitted.
While ensuring real-time detection speed, it significantly improves the detection accuracy of small infrared targets, effectively solves the problems of missed detection and false alarm, reduces model complexity, adapts to edge computing deployment, and enhances the model's generalization ability and detection performance.
Smart Images
Figure CN122156916A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computers, and in particular to a real-time infrared small target detection system and method. Background Technology
[0002] Small infrared targets refer to targets in infrared images that occupy very few pixels, have weak signals, and low signal-to-noise ratios. They are usually only represented as dots or subdots, lacking structural features such as shape, texture, and outline, and are easily overwhelmed by background noise or clutter.
[0003] Currently, infrared small target detection algorithms mainly include traditional algorithms, target detection-based methods, and semantic segmentation-based methods.
[0004] 1. Traditional small target detection algorithms based on filtering mainly include Top-Hat filtering, spatial high-pass filtering, bilateral filtering, two-dimensional least mean square filtering, and wavelet transform. These algorithms rely on manually designed features, have poor robustness, and are not suitable for complex detection scenarios.
[0005] 2. Semantic segmentation-based methods mainly include network models such as MDvsFA, UIU-Net, DNANet, ISNet, IAANet, and AGPCNet. These models typically have high accuracy, but they also have high computational complexity, a large number of parameters and computational load, and are time-consuming, making them unsuitable for real-time detection and edge computing deployments.
[0006] 3. Object detection-based methods mainly include R-CNN, FAST R-CNN, Faster R-CNN, RetinaNet, and the YOLO series of algorithms. These algorithms are generally fast, but their drawback is lower accuracy in detecting small objects, making them less effective for small object detection tasks.
[0007] In recent years, researchers have proposed several infrared small target detection methods based on improvements to YOLO. In 2023, Mou et al. designed the YOLO-FR infrared small target detection system based on the YOLOv5 model. In the same year, Ronghao Li et al. proposed the YOLOSR-IST method, introducing a Swin Transformer module to replace the bottleneck layer in the network's C3 module. In 2024, Hao et al. proposed the YOLO-SR model, introducing a BTB and a neck C3 module. These methods more effectively address the problems of missed detections and false alarms.
[0008] However, before using the YOLO-improved model for infrared image small target detection, a super-resolution model is usually used to perform super-resolution upscaling of the infrared small target image. This preprocessing step usually takes hundreds of milliseconds, making it difficult to apply these models to real-time small target detection scenarios.
[0009] Therefore, the existing technology has the following technical problems to be solved: how to significantly improve the detection accuracy of small infrared targets while ensuring real-time detection speed, solve the problems of missed detection and false alarm, and at the same time reduce the model complexity to adapt to edge computing deployment. Summary of the Invention
[0010] The summary of this invention introduces a series of simplified concepts, all of which are simplifications of existing technologies in the field, and will be further explained in detail in the detailed description section. This summary is not intended to limit the key features and essential technical features of the claimed technical solution, nor is it intended to determine the scope of protection of the claimed technical solution.
[0011] The technical problem to be solved by the present invention is to provide a novel lightweight real-time infrared small target detection system that significantly improves the detection accuracy of infrared small targets while ensuring real-time detection speed, and effectively solves the problems of missed detection and false alarm.
[0012] To solve the above-mentioned technical problems, the present invention provides an infrared small target real-time detection system, comprising: The input layer is used to receive the input infrared image; The backbone network layer, based on the YOLOv8n architecture, includes multiple lightweight adaptive feature enhancement units connected in sequence. Each lightweight adaptive feature enhancement unit includes an efficient dynamic convolution module, a spatial depth transformation module, and a gradient flow fusion module connected in sequence. It extracts feature maps from the input infrared image to provide feature representations for subsequent target detection. The neck network layer adopts the PANet structure, and the convolutional layer uses an efficient dynamic convolution module, which fuses the feature maps output by the backbone network layer to generate a unified feature representation rich in multi-scale semantic information. The output layer is used to output the detection results; Among them, the weights of the efficient dynamic convolution module are dynamically adjusted according to the input features, and the spatial depth transformation module transfers the spatial information of the input feature map into the channel dimension.
[0013] Preferably, in a further improvement to the infrared small target real-time detection system, the loss function of the backbone network layer uses normalized second-order Wasserstein distance instead of the IoU metric in the original loss function of the YOLOv8 architecture. The loss function first models the predicted bounding boxes and the ground truth bounding boxes using a two-dimensional Gaussian distribution, and then calculates the normalized second-order Wasserstein distance between the predicted and ground truth bounding boxes as the bounding box similarity metric.
[0014] Preferably, in a further improved version of the infrared small target real-time detection system, the weights of the efficient dynamic convolution module are input adaptive weights, which are dynamically adjusted according to the input features, enabling the network to learn more feature information from a large-scale dataset.
[0015] Preferably, the infrared small target real-time detection system further includes: an image acquisition module, which includes a cooled mid-wave infrared detector for capturing infrared small target images and transmitting them to a host computer.
[0016] Preferably, in a further improvement of the infrared small target real-time detection system, the host computer further includes an image preprocessing module, which sequentially performs blind pixel replacement, non-uniformity correction, and tone mapping processing on the infrared small target image.
[0017] Preferably, in a further improvement to the infrared small target real-time detection system, the infrared detector in the image acquisition module is an uncooled infrared detector or a long-wave infrared detector.
[0018] Preferably, the infrared small target real-time detection system is further improved by using a custom infrared small target dataset for training. The infrared small target dataset is formed by manually annotating high-quality bounding box labels in the format required for YOLO model training by filtering and summarizing publicly available infrared small target datasets.
[0019] Preferably, the infrared small target real-time detection system is further improved by being deployed on an edge computing device.
[0020] This invention provides a real-time infrared small target detection method, which is based on the YOLOv8n architecture and implemented through the aforementioned real-time infrared small target detection system, and includes the following steps: Step 1: Infrared image acquisition and preprocessing; the processed infrared image is displayed in real time and output to the detection network. Step 2: Construct the IRST-YOLO detection network based on the YOLOv8n network architecture. The IRST-YOLO detection network uses normalized second-order Wasserstein distance to replace the IoU metric in the original YOLOv8 loss function, transferring the spatial information of the feature map to the channel dimension. It uses an efficient dynamic convolution module to dynamically adjust the convolution kernel weights according to the input features to achieve adaptive feature extraction. Step 3: Train the IRST-YOLO model using a custom infrared small target dataset; Step 4: Input the preprocessed infrared image from Step 1 into the trained IRST-YOLO model. The model outputs the detection results of small infrared targets in the image. Exemplary detection results include category, location, and confidence level.
[0021] Preferably, the infrared small target real-time detection method is further improved by the preprocessing including blind pixel replacement, non-uniformity correction and tone mapping processing sequentially through ISPPipeline.
[0022] Preferably, the infrared small target real-time detection method is further improved by customizing the infrared small target data by filtering and summarizing the publicly available infrared small target dataset and manually annotating it with high quality according to the bounding box label format required for YOLO model training.
[0023] The working principle of this invention is as follows: 1. Traditional YOLO networks use IoU (Intersection over Union) as a metric for bounding box similarity. For small objects, even slight offsets between the predicted and ground truth bounding boxes can cause a sharp drop in IoU, making the loss function overly sensitive to changes in the location of small objects.
[0024] This invention remodels the bounding box using a two-dimensional Gaussian distribution, representing the bounding box as a two-dimensional Gaussian distribution. Then, it calculates the second-order Wasserstein distance between the two two-dimensional Gaussian distributions, and further normalizes the Wasserstein distance. This normalized second-order Wasserstein distance is used as a similarity metric to replace the IoU in the original loss function, making the network more robust to changes in the position of small targets and effectively improving the detection accuracy of small targets.
[0025] 2. Traditional convolution uses fixed weights, while the weights of efficient dynamic convolution modules are adaptive to the input. This adaptability allows the network to dynamically adjust the convolution kernel according to the input features, extracting infrared small target features in complex backgrounds more accurately, and performing better, especially with large-scale training data.
[0026] 3. The spatial depth transformation module of this invention effectively reduces spatial resolution while preserving information by transferring the spatial information of the feature map to the channel dimension. Taking scale=2 as an example, a feature map X of size S×S×C_1 is sliced into 4 sub-feature maps, each of size S / 2×S / 2×C_1. These 4 sub-feature maps are then connected along the channel dimension to obtain a new feature map X'. After processing by the spatial depth transformation module, the feature map X (S×S×C_1) is transformed into an intermediate feature map X' (S / 2×S / 2×(4×C_1)). This transformation ensures that the network can extract rich features, making it particularly suitable for handling difficult tasks such as low-resolution infrared images and small target detection.
[0027] Based on the above working principle, the present invention can achieve at least the following technical effects; 1. Traditional YOLO networks use IoU as a bounding box similarity metric. For small targets, even slight offsets between the predicted and ground truth boxes can cause a sharp drop in IoU, making the loss function overly sensitive to changes in the small target's position, resulting in low localization accuracy. Traditional IoU and its variants (such as GIoU, DIoU, and CIoU) are all based on the geometric overlap area of the bounding boxes. For pixel-level small targets, even slight positional shifts can cause drastic changes in the overlap area. This is an inherent flaw determined by its mathematical definition and cannot be solved by simple improvements.
[0028] This invention remodels the bounding boxes using a two-dimensional Gaussian distribution and introduces the normalized second-order Wasserstein distance as a similarity metric. The Wasserstein distance, based on optimal transport theory, measures the distance between two probability distributions. For small targets, even with slight offsets between the predicted and ground truth boxes, the NWD value changes smoothly, making the loss function more robust to changes in the small target's location. Furthermore, the Wasserstein distance considers the shape and size information of the bounding boxes, reflecting the actual similarity of small targets better than IoU.
[0029] 2. Traditional convolution uses fixed weights, which are difficult to adapt to the complex and varied background features of small infrared targets; direct downsampling leads to the loss of spatial information, which is not conducive to small target detection. The fixed weight design of traditional convolution cannot adaptively adjust according to the input content and lacks dynamic feature extraction capabilities; standard downsampling operations (such as pooling and stride convolution) directly discard spatial information and lack a spatial-channel information transfer mechanism for spatial depth transformation.
[0030] This invention replaces traditional convolution with an efficient dynamic convolution module and introduces a spatial depth transformation module inserted before the gradient flow fusion module. The efficient dynamic convolution dynamically adjusts the kernel weights based on input features, enabling it to learn more feature information from large-scale datasets and adaptively extract infrared small target features against complex backgrounds. The spatial depth transformation module transfers spatial information to the channel dimension, preserving more information during downsampling, avoiding spatial feature loss, and enhancing the feature representation capability for low-resolution small targets. This invention solves the problems of low resolution and difficult feature extraction in infrared images, improving feature representation capabilities.
[0031] 3. Existing semantic segmentation methods (such as MDvsFA, UIU-Net, etc.) have high computational complexity and a large number of parameters; super-resolution-based preprocessing methods take hundreds of milliseconds, making real-time detection difficult. Existing semantic segmentation methods require pixel-level classification, which naturally results in a large computational load; super-resolution preprocessing requires additional neural network forward propagation, which is unavoidably time-consuming.
[0032] This invention is based on the lightweight YOLOv8n architecture, introducing efficient dynamic convolution and spatial depth transformation modules to eliminate the super-resolution preprocessing step. YOLOv8n itself is a lightweight design, and the efficient dynamic convolution improves feature extraction capabilities while avoiding the addition of a large number of parameters through a dynamic weighting mechanism. The spatial depth transformation module operates without parameters, does not increase computation, and solves the problems of high model complexity hindering real-time detection and edge deployment.
[0033] 4. Existing publicly available infrared small target datasets are small in scale and have monotonous backgrounds. Direct use of these datasets leads to insufficient model training and poor generalization ability. The monotonous backgrounds of single datasets also result in a lack of diversity. Furthermore, the quality of automatic annotation is difficult to guarantee, which affects the training effect.
[0034] This invention filters and aggregates publicly available infrared small target datasets (e.g., four or more), manually annotates them with high quality according to the bounding box label format required for YOLO model training, and creates multiple (e.g., 2855) infrared small target datasets with bounding box label formats. Through data augmentation and high-quality annotation, the diversity and scale of the training data are increased, enabling the model to learn richer infrared small target features and improving the model's generalization ability and detection accuracy.
[0035] 5. This invention, while ensuring real-time detection speed, outperforms existing YOLO series models and other deep learning models in all aspects. Furthermore, the model successfully avoids missed detections and false positives, demonstrating excellent detection performance in complex scenarios. Attached Figure Description
[0036] The accompanying drawings are intended to illustrate the general characteristics of the methods, structures, and / or materials used in specific exemplary embodiments of the invention, supplementing the description in the specification. However, the drawings are schematic diagrams not drawn to scale and may not accurately reflect the precise structural or performance characteristics of any of the given embodiments. The drawings should not be construed as limiting or restricting the range of numerical values or properties covered by exemplary embodiments of the invention. The invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0037] Figure 1 This is a schematic diagram of the overall architecture of the infrared small target detection system provided in an embodiment of the present invention.
[0038] Figure 2 This is a schematic diagram of the ISP Pipeline processing flow provided in an embodiment of the present invention.
[0039] Figure 3 A schematic diagram of the IRST-YOLO network structure provided in an embodiment of the present invention.
[0040] Figure 4 This is a schematic diagram illustrating the working principle of the spatial depth conversion module provided in this embodiment of the invention.
[0041] Figure 5 This is a comparative diagram of the efficient dynamic convolution module provided in this embodiment of the invention and the traditional convolution module.
[0042] Figure 6 This is a comparison chart of the detection effects of different models provided in the embodiments of the present invention. Detailed Implementation
[0043] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can fully understand other advantages and technical effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through different specific embodiments, and various details in this specification can also be applied based on different viewpoints, with various modifications or changes made without departing from the overall design concept of the invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. The following exemplary embodiments of the present invention can be implemented in many different forms and should not be construed as being limited to the specific embodiments set forth herein. It should be understood that these embodiments are provided to make the disclosure of the present invention thorough and complete, and to fully convey the technical solutions of these exemplary embodiments to those skilled in the art. It should be understood that when an element is referred to as "connected" or "combined" to another element, the element can be directly connected or combined to the other element, or there may be intermediate elements. The difference is that when an element is referred to as "directly connected" or "directly combined" to another element, there are no intermediate elements. Throughout the drawings, the same reference numerals always denote the same elements.
[0044] Example, reference Figure 1 As shown, the present invention provides an infrared small target real-time detection system, which can be implemented based on existing hardware devices, including: The input layer is used to receive the input infrared image; The backbone network layer, based on the YOLOv8n architecture, includes multiple lightweight adaptive feature enhancement units connected in sequence. Each lightweight adaptive feature enhancement unit includes an efficient dynamic convolution module, a spatial depth transformation module, and a gradient flow fusion module (C2f) connected in sequence. It extracts feature maps from the input infrared image to provide feature representations for subsequent target detection. The neck network layer adopts the PANet structure, and the convolutional layer uses an efficient dynamic convolution module to fuse the feature maps output by the backbone network layer to generate a unified feature representation rich in multi-scale semantic information. The output layer is used to output the detection results; Among them, the weights of the efficient dynamic convolution module are dynamically adjusted according to the input features, and the spatial depth transformation module S2D transfers the spatial information of the input feature map into the channel dimension.
[0045] Based on the above overall design concept, each component of the present invention will be described in detail below; refer to Figure 2 As shown, a hardware system for acquiring infrared small target images is built using a cooled mid-wave infrared detector (such as an InSb detector). The infrared detector captures infrared radiation through an optical lens, converts it into an electrical signal, and then transmits it to a host computer via a high-speed data interface (such as CameraLink or GigE interface). The infrared detector can be replaced with an uncooled infrared detector or other band (long-wave) infrared detectors depending on the application scenario.
[0046] After receiving the raw infrared image data, the host computer processes it through the ISP (Image Signal Processing) pipeline, as shown in the reference. Figure 3 As shown, it includes: Blind pixel replacement: Detecting and replacing blind pixels (pixels that respond abnormally) in the detector, for example, by using a neighborhood interpolation method; Non-uniformity correction: For example, using two-point correction or scene adaptive correction algorithms to eliminate detector response non-uniformity; Tone mapping: For example, mapping 14-bit or 16-bit raw data to an 8-bit display range, using histogram equalization or adaptive gamma correction to enhance image contrast.
[0047] like Figure 4 As shown, this embodiment provides the specific structure of the IRST-YOLO network, and makes the following three improvements based on the YOLOv8n architecture: 1. Loss Function Optimization: To address the sensitivity of the IoU metric and its variants to small object bounding boxes, this paper proposes a novel similarity metric based on Wasserstein distance. First, the bounding boxes are remodeled using a two-dimensional Gaussian distribution. Then, the similarity between the predicted and ground truth bounding boxes is calculated using the normalized second-order Wasserstein distance (NWD). Finally, NWD is used as the similarity metric to replace the IoU metric in the YOLOv8 loss function, significantly enhancing the detection performance of YOLOv8 in applications targeting small object detection.
[0048] 2. Perform spatial depth transformation. By transferring the spatial information of the feature map to the channel dimension, the spatial resolution can be effectively reduced while preserving information, ensuring that the network can extract rich features. This is especially suitable for handling difficult tasks such as low-resolution infrared images and small target detection.
[0049] refer to Figure 3 As shown, the working principle of the S2D module is illustrated. Taking scale=2 as an example, the feature map X of size S×S×C_1 is sliced into 4 sub-feature maps, each of size S / 2×S / 2×C_1. Then, the 4 sub-feature maps are connected along the channel dimension to obtain a new feature map X'. After processing by the S2D module, the feature map X (S×S×C_1) is transformed into an intermediate feature map X' (S / 2×S / 2×(4×C_1)).
[0050] 3. The convolutional layers employ efficient dynamic convolution modules (efficient dynamic convolution). Unlike the fixed weights of traditional convolution, the weights W' of dynamic convolution are adaptive to the input. This adaptability allows IRST-YOLO to dynamically adjust the convolution kernel based on input features, thereby enabling more accurate extraction of features and small infrared targets in complex backgrounds, thus improving the model's detection accuracy, especially when handling large-scale training data. (Reference) Figure 5 As shown, this illustrates the difference in working principles between traditional convolutional modules and efficient dynamic convolutional modules.
[0051] Based on the embodiments of the present invention, a comparative experiment was conducted between IRST-YOLO and deep learning models of the YOLO series and non-YOLO series. The results of the comparative experiment are shown in Tables 1 and 2 below.
[0052] Table 1. Comparative experimental results of IRST-YOLO and other versions of the YOLO series; Model Name Precision Recall mAP@0.5 mAP@0.5:0.95 Params(M) FLOPs(G) YOLOv5s 0.934 <![CDATA[ 0.922 ]]> 0.967 0.521 9.11 23.8 YOLOv6s 0.937 0.87 0.927 0.463 16.30 44.0 YOLOv8n 0.938 0.912 0.952 0.498 3.01 8.1 YOLOv8s 0.942 <![CDATA[ 0.922 ]]> 0.954 0.515 11.13 28.4 YOLOv10s 0.9 0.912 0.948 <![CDATA[ 0.533 ]]> 8.04 24.4 YOLO11s 0.942 <![CDATA[ 0.951 ]]> <![CDATA[ 0.969 ]]> 0.532 9.41 21.3 SRResNet+YOLOSR-IST-s 0.937 0.91 0.946 0.463 6.4 12.7 ESRGAN+YOLO-MST <![CDATA[ 0.958 ]]> 0.918 0.964 0.475 12.7 20.9 IRST-YOLO <![CDATA[ 0.988 ]]> <![CDATA[ 0.951 ]]> <![CDATA[ 0.989 ]]> <![CDATA[ 0.534 ]]> 5.18 32.2 Table 2 shows the experimental results comparing IRST-YOLO with other deep learning models; Model Name Precision Recall F1 score Faster R-CNN 0.532 0.467 0.497 SSD 0.602 0.554 0.577 RetinaNet 0.615 0.599 0.606 RFBNet 0.586 0.575 0.580 RefineDet 0.692 0.627 0.657 MDvsFA 0.846 0.845 0.845 ACM 0.866 0.873 0.869 ALCNet <![CDATA[ 0.883 ]]> <![CDATA[ 0.899 ]]> <![CDATA[ 0.890 ]]> IRST-YOLO <![CDATA[ 0.988 ]]> <![CDATA[ 0.951 ]]> <![CDATA[ 0.969 ]]> IRST-YOLO significantly outperforms all other models in precision, demonstrating a substantial improvement in addressing false positives in small target detection. It also achieves first place in recall, tying YOLO11s, indicating a significant reduction in missed detections. Furthermore, it achieves the best results in the two most important accuracy metrics for YOLO models, mAP@0.5 and mAP@0.5:0.95, demonstrating excellent performance across various scenarios with both lenient and stringent accuracy requirements. Simultaneously, the model's params and FLOPs remain within acceptable limits, enabling successful deployment on edge computing devices for real-time infrared small target detection.
[0053] Figure 6 Several representative images from the SIRST dataset test set are shown as examples of visualizations of the detection results. Areas with missed detections and false positives are marked with boxes. Experimental results show that the YOLO series algorithms exhibit varying degrees of missed detections and false positives, with some models outperforming others. YOLOv5s, YOLOv6s, YOLOv8n, and YOLOv8s perform poorly in detecting complex images of small infrared targets, frequently resulting in missed detections. In contrast, YOLOv9s, YOLOv10s, and YOLO11s models demonstrate some reasonable inference in complex infrared small target scenes, but they also experience missed detections and false positives. The proposed IRST-YOLO model successfully avoids these problems, outperforming all other YOLO series models in detection inference.
[0054] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It will also be understood that, unless explicitly defined herein, terms such as those defined in a general dictionary shall be interpreted as having the meaning consistent with their meaning in the relevant field context, and not as having an idealized or overly formal meaning.
[0055] The present invention has been described in detail above through specific embodiments and examples, but these are not intended to limit the invention. Many modifications and improvements can be made by those skilled in the art without departing from the principles of the invention, and these should also be considered within the scope of protection of the present invention.
Claims
1. A real-time infrared small target detection system, characterized in that, include: The input layer is used to receive the input infrared image; The backbone network layer, based on the YOLOv8n architecture, includes multiple lightweight adaptive feature enhancement units connected in sequence. Each lightweight adaptive feature enhancement unit includes an efficient dynamic convolution module, a spatial depth transformation module, and a gradient flow fusion module connected in sequence. It extracts feature maps from the input infrared image to provide feature representations for subsequent target detection. The neck network layer adopts the PANet structure, and the convolutional layer uses an efficient dynamic convolution module, which fuses the feature maps output by the backbone network layer to generate a unified feature representation rich in multi-scale semantic information. The output layer is used to output the detection results; Among them, the weights of the efficient dynamic convolution module are dynamically adjusted according to the input features, and the spatial depth transformation module transfers the spatial information of the input feature map into the channel dimension.
2. The infrared small target real-time detection system as described in claim 1, characterized in that: The loss function of the backbone network layer uses normalized second-order Wasserstein distance instead of the IoU metric in the original loss function of the YOLOv8 architecture.
3. The infrared small target real-time detection system as described in claim 1, characterized in that: The weights of the efficient dynamic convolution module are input-adaptive weights that are dynamically adjusted based on the input features.
4. The infrared small target real-time detection system as described in claim 1, characterized in that, Also includes: The image acquisition module includes a cooled mid-wave infrared detector for capturing images of small infrared targets and transmitting them to a host computer.
5. The infrared small target real-time detection system as described in claim 4, characterized in that: The host computer also includes an image preprocessing module, which sequentially performs blind pixel replacement, non-uniformity correction, and tone mapping processing on the infrared small target image.
6. The infrared small target real-time detection system as described in claim 4, characterized in that: The infrared detector in the image acquisition module is an uncooled infrared detector or a long-wave infrared detector.
7. The infrared small target real-time detection system as described in claim 1, characterized in that: The real-time infrared small target detection system is trained using a custom infrared small target dataset. This dataset is created by manually annotating high-quality bounding box labels in the format required for YOLO model training, after filtering and summarizing publicly available infrared small target datasets.
8. The infrared small target real-time detection system as described in any one of claims 1-7, characterized in that: It is deployed on edge computing devices.
9. A real-time infrared small target detection method, based on the YOLOv8n architecture, implemented using the real-time infrared small target detection system described in claim 1, characterized in that... Includes the following steps: Step 1: Infrared image acquisition and preprocessing; the processed infrared image is displayed in real time and output to the detection network. Step 2: Construct the IRST-YOLO detection network based on the YOLOv8n network architecture. The IRST-YOLO detection network uses normalized second-order Wasserstein distance to replace the IoU metric in the original YOLOv8 loss function, transferring the spatial information of the feature map to the channel dimension. It uses an efficient dynamic convolution module to dynamically adjust the convolution kernel weights according to the input features to achieve adaptive feature extraction. Step 3: Train the IRST-YOLO model using a custom infrared small target dataset; Step 4: Input the preprocessed infrared image from Step 1 into the trained IRST-YOLO model, and the model outputs the detection results of small infrared targets in the image.
10. The real-time infrared small target detection method as described in claim 9, characterized in that: The preprocessing includes sequential blind replacement, non-uniformity correction, and tone mapping via the ISP Pipeline.
11. The real-time infrared small target detection method as described in claim 9, characterized in that: Custom infrared small target data is created by filtering and summarizing publicly available infrared small target datasets, and manually annotating them with high quality according to the bounding box label format required for YOLO model training.