Unmanned vehicle infrared target detection method

Abstract

The invention provides an infrared target detection method for an unmanned carrier. The method comprises the following steps: preprocessing an infrared image acquired by the unmanned carrier; inputting the preprocessed image into an improved IFD-YOLO detection model for feature extraction and target detection; and obtaining the position and category information of the target in the infrared image according to the output of the improved IFD-YOLO detection model. Feature extraction and target detection are carried out through an improved IFD-YOLO detection model, the improved IFD-YOLO detection model can reduce parameter redundancy and enhance the feature extraction capability, while the feature expression capability of the model is maintained, the calculation amount and the parameter amount are significantly reduced, the positioning precision of a high-quality target and the overall detection performance of the model can be improved, and the method is suitable for large-scale popularization and application. Therefore, the purposes of reducing the calculation amount and improving the detection precision are achieved.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and in particular relates to a method for infrared target detection of unmanned vehicles. Background Technology

[0002] With the rapid development of unmanned aerial vehicle (UAV) technology, it has been widely used in fields such as national defense reconnaissance, maritime surveillance, disaster relief, and border patrol. UAVs have advantages such as mobility, flexibility, convenient deployment, and low cost, enabling them to perform missions efficiently in complex or dangerous environments, greatly expanding aerial monitoring and information acquisition capabilities.

[0003] Among various sensing methods, infrared imaging has become an important technology for UAVs to perform monitoring and identification tasks due to its strong resistance to light interference, day and night usability, and high environmental adaptability. Infrared imaging achieves target perception in the environment by capturing the target's thermal radiation. It can still work stably even at night, in low visibility conditions such as fog, haze, or smoke, thus having irreplaceable advantages in scenarios such as military reconnaissance, search and rescue, and monitoring of complex environments.

[0004] Because drones fly at high altitudes and at long imaging distances, targets in infrared images often appear small on the imaging plane. Furthermore, infrared imaging itself is limited by low resolution, weak target thermal signals, and complex background clutter, resulting in low target contrast, blurred edges, and easy indistinguishability from the background. These factors combined make it difficult for traditional target detection algorithms to maintain high detection accuracy and stability in drone infrared scenarios.

[0005] Before the rise of deep learning methods, infrared target detection primarily relied on traditional image processing and pattern recognition techniques. These methods typically rely on differences in grayscale, texture, or spatial distribution features between the target and the background, employing manually designed feature extraction and discrimination rules for detection. Common approaches include using spatial or frequency domain filters to suppress the background and enhance the target response, such as high-pass filtering, Laplacian operators, Top-hat filtering, and Gaussian-Laplacian (LoG) filtering. While these methods are simple in structure and computationally efficient, they are prone to false detections in complex backgrounds or under strong noise conditions.

[0006] Another type of method is based on statistical modeling principles, achieving target separation by constructing a statistical distribution model of the background or detecting anomalies in pixel grayscale. Examples include singularity detection based on background prediction (RX algorithm), local mean and variance discrimination, or Bayesian modeling methods. These methods perform well in static or stable backgrounds but struggle with dynamic scenes or clutter interference. In addition, local contrast methods are widely used. Their core idea is to utilize the difference in brightness or energy between the target region and the surrounding background for detection. Typical examples include Local Contrast Metric (LCM) and Local Energy Contrast (LEC). However, traditional infrared detection methods generally rely on manually set features and thresholds, making them highly sensitive to imaging conditions, target scale, and background complexity, resulting in insufficient robustness in multi-scene and low signal-to-noise ratio environments. These limitations provide important directions and impetus for subsequent research on infrared target detection based on deep learning.

[0007] With the rapid development of deep learning, target detection technology has become increasingly mature. Many problems in infrared small target detection can be solved using deep learning methods. Currently, mainstream target detection algorithms are divided into two types: two-stage methods and one-stage methods. In the field of UAV aerial infrared small target detection, the two-stage method is typically represented by the R-CNN series of algorithms (such as R-CNN, Fast R-CNN, etc.). Its core process usually consists of two stages: first, generating candidate regions using heuristic algorithms such as Selective Search or convolutional neural networks (CNN); then, performing target classification and bounding box regression tasks on these candidate regions respectively. This method has a significant advantage in detection accuracy due to its detailed candidate region analysis, but because it relies on the two-step process of "candidate region generation - accurate detection," its inference speed is relatively slow, making it difficult to meet the requirements of real-time detection in infrared scenes.

[0008] In contrast, one-stage methods, centered around the YOLO (You Only Look Once) series, perform dense predictions across the entire image, extract features using CNNs, and directly complete target classification and bounding box regression, achieving true end-to-end detection. This method requires only one forward propagation to output detection results, significantly improving inference efficiency while greatly reducing model complexity. Although its accuracy is slightly lower than two-stage methods, its lightweight and efficient nature makes it more advantageous for detecting small infrared targets on UAVs.

[0009] Unmanned aerial vehicle (UAV) infrared (UAV) missions often face challenges such as small target size, complex backgrounds, and low signal-to-noise ratios. Furthermore, limitations in onboard computing resources and power consumption place higher demands on the real-time performance and lightweight design of detection algorithms. Against this backdrop, one-stage methods, with their compact structure and high computational efficiency, are better suited to the practical needs of infrared small target detection. Among these, the YOLO series of algorithms has become the mainstream choice due to its end-to-end prediction mechanism and superior speed performance. Compared to other one-stage detectors such as SSD (Single Shot MultiBox Detector) and RetinaNet, YOLO performs better in terms of inference speed, model lightweighting, and ease of engineering deployment. With the continuous evolution of the YOLO series (such as YOLOv5, YOLOv8, and YOLOv11), its feature extraction and small target perception capabilities in infrared scenes have been significantly improved. In summary, the YOLO series of algorithms provides a high-efficiency solution for UAV infrared small target detection that combines speed and accuracy.

[0010] However, in the task of detecting small targets in infrared by UAVs, due to the influence of thermal radiation characteristics on infrared imaging, the targets usually exhibit characteristics such as small size, low contrast, and blurred edges. At the same time, there is complex thermal noise and false target interference in the background. In addition, the imaging environment is significantly affected by factors such as climate, temperature and sensor performance, making it difficult for existing methods to guarantee high recognition accuracy and detection precision in this scenario.

[0011] To address these issues, numerous scholars have conducted extensive research in the field of infrared target detection in UAV aerial photography scenarios. Yuan et al. proposed the IRSDD-YOLOv5 model, which improves the accuracy of infrared small target detection for UAVs by adding an infrared small target detection module (IRSTDM), optimizing the loss function, and constructing the SIDD dataset. However, it is not fully adapted to multi-target dense scenes, and its real-time performance is slightly reduced. Wang et al. designed the PHSI-RTDETR algorithm, which achieves efficient detection with a lightweight module, HiLo attention, and Inner-GIoU loss. However, in multi-class scenes with severe target occlusion, the semantic feature mining of small targets is still insufficient. Aibibu et al. proposed the ERGW-net model, which improves detection performance by improving the backbone and neck structure and innovating L_GWPIoU loss. However, its adaptability to weakly featured small targets on small sample datasets is limited, and it does not fully consider the impact of extreme imaging conditions. Pan et al. proposed AIMED-Net, which optimizes YOLOv7 through shallow and deep feature enhancement networks and improves the accuracy of infrared small target detection by combining sa-CIoU loss. However, it is not adaptable enough to multi-target dense occlusion scenes. The DMF-YOLO proposed by Liu et al. optimizes multi-scale detection with DDSConv, MFAM modules and EW-BBRLF loss. Although it performs well on public datasets, the increase in parameters affects the real-time performance of micro UAVs. Liu et al. proposed Z-shaped cropping enhancement and improved YOLOv11 to improve the confidence of small occluded target detection. However, it has limited simulation of real irregular occlusion and its ability to distinguish categories in complex scenes needs to be strengthened. Summary of the Invention

[0012] In view of the problems existing in the prior art, the present invention provides an infrared target detection method for unmanned vehicles, which at least partially solves the problems of insufficient accuracy and large model size in the prior art.

[0013] This disclosure provides an infrared target detection method for unmanned vehicles, including: Preprocessing of infrared images collected by unmanned vehicles; The preprocessed image is input into the improved IFD-YOLO detection model for feature extraction and target detection; Based on the output of the improved IFD-YOLO detection model, the location and category information of the target in the infrared image are obtained; The improved IFD-YOLO detection model includes: RepViT is used as the backbone network to extract local detail features and global semantic features of the image; The processing procedure of the RepViT Block includes: The Token Mixer module performs spatial feature mixing and recombination on the input features. The Token Mixer module is implemented using a 3×3 depthwise separable convolution. The output of the Token Mixer module is as follows: , in, This represents a 3×3 depthwise convolution operation, and BN stands for batch normalization. It is the ReLU activation function. For input features.

[0014] Optionally, the improved IFD-YOLO detection model further includes: Feature extraction is performed using the C3k2_DyGhost module, which integrates dynamic convolution mechanism and Ghost feature generation mechanism. An adaptive fusion IoU loss function is used as the bounding box regression loss function.

[0015] Optionally, the RepViT backbone network uses a structural reparameterization mechanism. During the training phase, the RepViT backbone network adopts a multi-branch structure to enhance feature representation, and during the inference phase, the RepViT backbone network merges the branches into a single convolutional path. The RepViT backbone network includes an input stage, multiple feature extraction stages, and a downsampling module. The feature extraction stage contains multiple RepViT Blocks or RepViT SEBlocks.

[0016] Optionally, in the RepViT Block, the network uses a combination of channel-wise convolution and point-wise convolution for feature modeling.

[0017] Optionally, the processing of the RepViT Block further includes: The Channel Mixer module performs cross-channel feature modeling and nonlinear enhancement on the output of the Token Mixer. The Channel Mixer module consists of two 1×1 convolutional layers and the GELU activation function. When the step size is 1, a residual connection is introduced between the output and input features of the Channel Mixer.

[0018] Optionally, the C3k2_DyGhost module includes a DyGhost module, which generates features through the following steps: K sets of convolutional kernel parameters are preset, and corresponding weighting coefficients are generated based on the input features through a lightweight routing branch; The input features are convolved using the K sets of convolution kernels to obtain K sets of feature responses; The dynamic convolution output features are obtained by weighting and summing the K sets of feature responses based on weighted coefficients. The Ghost feature is obtained by concatenating the dynamic convolution output feature with the redundant feature generated by its linear transformation. The Ghost features are batch normalized and then fused with the input features via residual connections to obtain the output features.

[0019] Optionally, the preset K sets of convolutional kernel parameters are used to generate corresponding weighting coefficients based on the input features through a lightweight routing branch, including: The lightweight routing branch extracts global information of the input features through global average pooling (GAP), generates weighted coefficients through a linear mapping function, and then normalizes the weighted coefficients through the Softmax function.

[0020] Optionally, the adoption of an adaptive fusion IoU loss function as the bounding box regression loss function includes: Scale the predicted bounding box and the ground truth bounding box, and calculate the scaled IoU value. ; Calculate the loss value based on the scaled IoU value and the confidence score (conf) of the predicted bounding box. ,in An exponential factor for controlling the weights of difficult samples.

[0021] Optionally, the step of using the adaptive fusion IoU loss function as the bounding box regression loss function further includes: A dynamic annealing strategy is introduced, in which the scaling ratio is controlled by the hyperparameter ratio to change with the training rounds. In the early stage of training, the hyperparameter ratio is larger, so that the model focuses on low-quality prediction boxes to accelerate convergence. In the later stage of training, the hyperparameter ratio is reduced, so that the model focuses on high-quality prediction boxes to improve localization accuracy.

[0022] Optionally, the formula for calculating the hyperparameter ratio is: Where T is the total number of training rounds, The time unit refers to the process of the entire training dataset undergoing one forward and backward propagation through the neural network.

[0023] The infrared target detection method for unmanned vehicles provided by this invention uses an improved IFD-YOLO detection model for feature extraction and target detection. The improved IFD-YOLO detection model can reduce parameter redundancy and enhance feature extraction capabilities. While maintaining the model's feature representation capabilities, it significantly reduces the amount of computation and parameters, thereby improving the positioning accuracy of high-quality targets and the overall detection performance of the model. This achieves the goal of reducing computation while improving detection accuracy. Attached Figure Description

[0024] The above and other objects, features and advantages of this disclosure will become more apparent from the accompanying drawings, in which like reference numerals generally denote like parts.

[0025] Figure 1 This is a schematic diagram of the IFD-YOLO structure provided in an embodiment of this disclosure; Figure 2 This is a schematic diagram of the RepViTBlock structure provided in the embodiments of this disclosure; Figure 3 This is a schematic diagram of the RepViT backbone network structure provided in an embodiment of this disclosure; Figure 4 This is a schematic diagram of the C3k2_DyGhost structure provided in an embodiment of this disclosure; Figure 5 This is a schematic diagram of the DyGhost Module structure provided in the embodiments of this disclosure; Figure 6 This is a schematic diagram of the DyGhostBottleneck structure provided in an embodiment of this disclosure; Figure 7a A visualization diagram showing the categories and number of targets in the HIT-UAV dataset used in the embodiments of this disclosure; Figure 7b A visualization of the ratio of the height to the width of the bounding box in the original image of the HIT-UAV dataset used in this embodiment of the disclosure; Figure 8 This is a schematic diagram of the PR curve of YOLOv11n on the HIT-UAV dataset provided in this embodiment of the disclosure; Figure 9 This is a schematic diagram of the PR curve of IFD-YOLO on the HIT-UAV dataset provided in this embodiment of the disclosure; Figure 10 A schematic diagram of the PR curve of YOLOv11n on the IRSTD-1k dataset provided in this embodiment of the disclosure; Figure 11This is a schematic diagram of the PR curve of IFD-YOLO on the IRSTD-1k dataset provided in an embodiment of this disclosure. Detailed Implementation

[0026] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0027] Infrared small target detection by unmanned vehicles is often hampered by factors such as small target size, complex background, and limited computational resources. Existing infrared small target detection models generally suffer from insufficient accuracy and excessive model size. To address these issues, this embodiment proposes an improved infrared small target detection algorithm based on YOLOv11n—IFD-YOLO. First, a lightweight backbone network, RepViT, is introduced to reduce parameter redundancy and enhance feature extraction capabilities. Second, a C3k2-DyGhost module is designed to replace the original C3k2 module, significantly reducing computation and parameter count while maintaining the model's feature representation capabilities. Finally, an AF-IoU loss function is designed to dynamically adjust sample weights, improving the localization accuracy of high-quality targets and the overall detection performance of the model. Experimental results on the HIT-UAV public dataset show that IFD-YOLO improves the mAP@50 and mAP@95 metrics by 4.9% and 3.1% respectively compared to the original model, while reducing the number of parameters and computation by 23% and 21% respectively, demonstrating the effectiveness of this method in infrared small target detection.

[0028] The key points of this embodiment are as follows: (a) RepViT backbone reconstruction: The RepViT backbone structure was introduced into the YOLOv11 network. Through the synergy of the MetaFormer framework and the structural reparameterization mechanism, this structure achieves efficient fusion of local spatial information and global contextual relationships in the feature extraction stage, effectively enhancing the feature perception and discrimination performance in infrared small target detection.

[0029] (b) Dynamic Ghost Feature Extraction Module (C3k2_DyGhost): Based on the original C3k2 structure, a novel C3k2_DyGhost feature extraction module is constructed by deeply integrating Ghost feature generation with a dynamic convolution mechanism. This module dynamically and adaptively generates redundant features and adjusts the convolution kernel weights according to the input feature distribution. This not only significantly improves the flexibility and discriminative power of feature representation, but also effectively alleviates the feature degradation problem caused by low target contrast and complex background in infrared small target detection, thereby enhancing the detection accuracy and generalization ability of the model.

[0030] (c) Adaptive Fusion-IoU Loss (AF-IoU): This model integrates an adaptive fusion IoU loss function to replace the traditional bounding box regression loss. This loss function adaptively adjusts the model's focus on prediction boxes of different quality through a dynamic weighting mechanism, enabling a balanced transition between low-quality and high-quality samples during training. In infrared small target detection tasks, this loss function effectively improves the model's regression accuracy and stability for small-scale, low-contrast targets, significantly enhancing overall detection performance.

[0031] This embodiment aims to improve the performance of YOLOv11n in UAV infrared small target detection tasks through targeted improvements. It focuses on addressing the core problems of blurred small target features, strong background interference, and the difficulty in balancing detection accuracy and speed in infrared scenes, proposing a UAV infrared small target detection model called IFD-YOLO. Figure 1 As shown, the improvement of this model focuses on three key modules of YOLOv11n: feature extraction, feature enhancement, and bounding box regression loss. The specific design is as follows: In the feature extraction part, the original basic backbone structure is abandoned, and a lightweight RepViT convolutional neural network is introduced as the new backbone. RepViT enhances the ability to capture subtle features of small infrared targets by separating the MetaFormer-like structure of the Token mixer and the channel mixer, combined with structural reparameterization technology, thus solving the problem of insufficient feature extraction in infrared low-contrast scenes by the traditional backbone. To enhance the feature extraction capability, this study improves the original C3k2 module to the C3k2_DyGhost module. This module integrates a dynamic convolution mechanism in the Ghost structure, so that the convolution kernel weights can adapt to the input features, thereby enhancing the fine-grained representation capability of small infrared targets while maintaining lightweight design. In the bounding box regression loss part, an adaptive fusion loss function (AF-IoU) is used to replace the original IoU-type loss function. AF-IoU dynamically scales the predicted bounding box and ground truth box using Focal Box to adjust the weights of predicted bounding boxes of different quality. It combines an annealing strategy based on the ratio hyperparameter (focusing on low-quality boxes to accelerate convergence in the early stage of training and shifting to high-quality boxes to improve accuracy in the later stage) and optimizes the attention allocation mechanism for infrared small target detection scenarios. This effectively alleviates the problem that traditional IoU loss is insensitive to small target regression under high IoU thresholds and improves bounding box localization accuracy.

[0032] In summary, IFD-YOLO achieves a balance between accuracy and efficiency in UAV infrared small target detection tasks through triple synergistic improvements in the backbone, feature transformation, and loss function, providing a better solution for small target detection in complex infrared scenarios.

[0033] This embodiment discloses an infrared target detection method for unmanned vehicles, which at least partially solves the problems of insufficient accuracy and large model size in the prior art.

[0034] This disclosure provides an infrared target detection method for unmanned vehicles, including: Preprocessing of infrared images collected by unmanned vehicles; The preprocessed image is input into the improved IFD-YOLO detection model for feature extraction and target detection; Based on the output of the improved IFD-YOLO detection model, the location and category information of the target in the infrared image are obtained.

[0035] The improved IFD-YOLO detection model includes: RepViT is used as the backbone network to extract local detail features and global semantic features of the image; Feature extraction is performed using the C3k2_DyGhost module, which integrates dynamic convolution mechanism and Ghost feature generation mechanism. An adaptive fusion IoU loss function is used as the bounding box regression loss function.

[0036] RepViT backbone network: To improve feature extraction capabilities and model inference efficiency in UAV infrared small target detection tasks, this embodiment introduces a lightweight backbone network, RepViT (Re-parameterized VisionTransformer), to replace the original convolutional backbone based on YOLOv11. Infrared images typically exhibit characteristics such as small target size, low signal-to-noise ratio, and complex backgrounds. Traditional convolutional backbones, due to limitations in local receptive fields, struggle to simultaneously preserve fine-grained information and global semantics during downsampling, and their large parameter count hinders real-time deployment on resource-constrained UAV platforms. RepViT integrates the local feature extraction advantages of convolutional neural networks (CNNs) with the global modeling capabilities of visual transformers (ViTs). Through structural re-parameterization, it employs a multi-branch structure during training to enhance feature representation, and merges the branches into a single convolutional path during inference, thereby significantly reducing inference latency and computational complexity while maintaining detection accuracy.

[0037] RepViT's overall structure employs a hierarchical feature extraction strategy, primarily consisting of an input stage (Stem), a multi-stage feature extraction structure (Stage), and a downsampling module (Downsample), achieving progressive modeling from shallow local textures to deep semantic features. In the input stage, the model first performs two 3×3 convolutions and a downsampling operation with a stride of 2 on the infrared image to reduce spatial resolution and extract basic texture information, laying the foundation for subsequent high-level feature extraction. Subsequently, the features undergo an alternating feature extraction and compression process involving multiple Stages and downsampling modules. Each Stage consists of several stacked RepViT Blocks or RepViT SEBlocks. The RepViT Block and RepViT network structures are shown in the figures below. Figure 2 and Figure 3 As shown.

[0038] In each RepViT Block, the network uses a combination of depthwise convolution and pointwise convolution for feature modeling. Specifically, in the i-th RepViT Block of the l-th layer, the input features are denoted as... The Token Mixer module achieves feature mixing and recombination in the spatial dimension through 3×3 depthwise separable convolutions. Its output can be represented as: (1), in, This represents a 3×3 depthwise convolution operation, and BN stands for batch normalization. (·) is the ReLU activation function. Furthermore, a Squeeze-and-Excitation (SE) attention mechanism is used to adaptively adjust channel weights to highlight salient features of small targets in infrared images.

[0039] The Channel Mixer module consists of two 1×1 convolutional layers and a GELU activation function. Functionally, it is equivalent to the Feed-Forward Network (FFN) in the Transformer. It achieves cross-channel feature modeling and nonlinear enhancement through "channel expansion—nonlinear mapping—channel compression," thereby improving the richness and discriminativeness of feature representation. Its calculation process is as follows: (2), in, and These represent the weight matrices of the two 1×1 convolutional layers, (·) is the GELU activation function.

[0040] Finally, for blocks with stride=1, the network introduces residual connections outside the Channel Mixer to preserve the integrity of shallow features and mitigate the vanishing gradient problem. The output after adding the residual connection is: (3), When stride=1, the residual connection ensures that shallow texture information is preserved, alleviates gradient vanishing, and improves feature transfer efficiency.

[0041] Through this hierarchical modeling and multi-scale fusion feature extraction strategy, RepViT can effectively preserve the fine-grained features of small infrared targets during downsampling while capturing global contextual information. This significantly improves the model's detection accuracy and feature robustness while maintaining a lightweight design. Finally, the backbone network outputs multi-scale feature maps, providing high-resolution and semantically complementary feature support for the YOLO detection head, enabling accurate identification and localization of small infrared targets.

[0042] C3k2_DyGhost module: To further enhance the model's feature extraction capabilities in infrared small target detection tasks, this embodiment designs an improved C3k2_DyGhost module based on the original YOLOv11n network's C3k2 module. For example... Figure 4 As shown, this module largely adopts the residual structure of the C3 framework to maintain the continuity of feature flow and the stability of gradient propagation. Its core innovation lies in replacing the internal Bottleneck unit with a DyGhost Bottleneck that integrates the Ghost feature generation mechanism and Dynamic Convolution. Figure 6 As shown, while maintaining lightweight feature extraction capabilities, adaptive modeling of input features is achieved.

[0043] Traditional GhostModules generate inherent features through a small number of standard convolutions, and then use linear transformations (such as channel-wise convolution or depthwise convolution) to generate redundant features, thereby obtaining rich feature representations with low computational cost. However, its convolution kernel weights are fixed, lacking adaptability to changes in input feature distribution, making it difficult to effectively capture weak and small targets in complex infrared backgrounds. Therefore, this embodiment introduces a dynamic convolution mechanism in the Ghost feature generation stage, using multiple sets of learnable convolution kernels to achieve adaptive input feature modeling. The DyGhost module structure is as follows: Figure 5 As shown. Specifically, the DyGhost module presets k sets of convolutional kernel parameters and generates corresponding weighting coefficients through a lightweight attention branch. The weighting coefficients are generated by the Softmax function to ensure the balance and stability of the convolution kernel combination. Let the input features be... The DyGhost module uses multiple sets of learnable convolutional kernels. Generate different feature responses: (4), in, Represents the parameters of the k-th convolutional kernel. The symbol represents the corresponding feature response, and * indicates a convolution operation. To achieve adaptive feature modeling, DyGhost designs a lightweight routing branch that extracts global input information through Global Average Pooling (GAP) and then maps it to a linear mapping function. (·) Generate weighting coefficients : (5), The final output feature is obtained by weighted summation of the results of each convolution group: (6), This mechanism enables the model to adaptively adjust the convolution response based on the distribution of input features, thereby effectively improving its adaptability to diverse targets and complex backgrounds in infrared scenes.

[0044] Next, in the Ghost feature generation stage, the redundant features generated from the main features and the linear features are concatenated: (7), in, This represents the concatenated Ghost feature map. The input features are then fused using batch normalization (BN) and residual connections. (8), When stride=1, the residual term directly connects the input and output, realizing the fusion of shallow and deep features; when stride=2, the input first passes through downsampling convolution to match the spatial dimension.

[0045] like Figure 6 As shown, when stride=1, the main branch consists of two stacked DyGhost modules. Each DyGhost module adopts a standard structure of convolutional layer—batch normalization (BN) layer—ReLU activation function to achieve efficient feature extraction and non-linear representation. The residual branch directly adds the input features, after BN normalization, to the output of the main branch, thereby achieving effective fusion of shallow details and deep semantic features while maintaining computational lightweightness.

[0046] In contrast, when stride=2, the main branch needs to be downsampled using depthwise dynamic convolution (stride=2), halving the feature map resolution. While this process helps expand the receptive field and extract higher-level semantic information, in UAV infrared small target detection tasks, this downsampling leads to a significant loss of fine-grained texture and low-contrast thermal target information, even causing some extremely small targets to completely disappear from the feature map. Therefore, this embodiment uses a stride=1 design in the DyGhost module to maximize the preservation of spatial resolution and detailed features, ensuring the model has stronger response capabilities and localization accuracy when detecting weak, small-scale infrared targets.

[0047] This embodiment's structure maintains computational efficiency while enabling the organic fusion of local details and global semantics during cross-layer flow. The dynamic convolution embedding of the Ghost feature generation unit allows the module to enhance high-frequency texture and edge information while remaining lightweight. This feature is particularly suitable for low-contrast and high-noise scenarios common in infrared small target detection: the DyGhost module can adaptively adjust convolution weights based on the saliency distribution of the input features, thereby highlighting the small target response and suppressing background interference.

[0048] In summary, the C3k2_DyGhost module organically integrates the lightweight approach of GhostModule with the adaptive modeling capabilities of dynamic convolution. It significantly outperforms the original C3k2 structure in terms of feature representation accuracy and robustness, while maintaining a good balance between computational complexity and parameter size. This design provides a more efficient, flexible, and discriminative feature extraction foundation for UAV infrared small target detection tasks, offering more discriminative input information for subsequent feature fusion and detection head prediction.

[0049] In object detection tasks, bounding box regression (BBR) is a crucial step that directly determines detection accuracy, and the design of the loss function plays a central role. Traditional IoU-based losses (such as GIoU, DIoU, and CIoU) primarily characterize the difference between the predicted and ground truth boxes using geometric metrics such as center distance, aspect ratio, and bounding area. Although these methods have improved the gradient vanishing and scale insensitivity problems of the original IoU to some extent, improvements in geometric metrics have gradually reached saturation, and the coupling relationship between different metrics also limits the optimization effect.

[0050] In UAV infrared small target detection scenarios, high-quality bounding boxes are extremely rare, while low-quality bounding boxes dominate gradient updates. This leads to the model overemphasizing low-quality samples, resulting in slow convergence and insufficient localization accuracy. To address this challenge, this embodiment proposes an adaptive fusion loss function, AF-IoU (Adaptive Fusion-IoU Loss), to redefine the weight allocation mechanism for bounding box quality. This allows the model to dynamically adjust its focus on high- and low-quality bounding boxes at different training stages, achieving a better balance between detection accuracy and training efficiency.

[0051] The core idea of ​​AF-IoU goes beyond simply improving geometric metric rules; it achieves a "coarse-to-fine" training process through dynamic weight allocation. Its core mechanisms include a scaling-based Focal Box strategy, annealed attention adjustment, and confidence-weighted approaches. Specifically, AF-IoU indirectly changes the IoU value by scaling the predicted and ground truth boxes, thereby adjusting the loss weights for predicted boxes of different quality. Let the original IoU be: (9), When the predicted bounding boxes are reduced, the intersection area decreases, the IoU decreases, and the loss increases; conversely, when the predicted bounding boxes are enlarged, the IoU increases, and the loss decreases. The scaling factor r determines the model's focus: when r < 1, the model focuses more on high-quality predicted bounding boxes; when r > 1, it focuses more on low-quality predicted bounding boxes. This approach requires no additional geometric calculations; dynamic weight allocation is achieved through scale changes, allowing the model to adaptively focus on more representative samples at different stages of training.

[0052] To balance the convergence speed in the early stages of training with the high-accuracy regression in the later stages, AF-IoU introduces a dynamic annealing strategy based on "box attention." This strategy uses a dynamic hyperparameter `ratio` to control the scaling ratio, which gradually decreases with each training epoch, allowing the model's attention to smoothly transition from low-quality predicted boxes to high-quality predicted boxes. Its variation pattern is similar to that of cosine annealing, and can be expressed as: (10) Where T represents the total number of training epochs. In the early stages of training, a larger ratio causes the model to focus on low-quality predicted boxes to accelerate convergence; while in the later stages of training, as the ratio decreases, the model gradually shifts its attention to high-quality predicted boxes to obtain more accurate bounding box localization results. This "coarse-to-fine" dynamic adjustment method effectively alleviates the problem of sample quality imbalance in object detection.

[0053] Furthermore, AF-IoU borrows from the idea of ​​Focal Loss, further optimizing the loss distribution by introducing a confidence-weighted term. Let the confidence of the predicted bounding box be conf, then the overall form of AF-IoU can be expressed as: (11), in: (12) γ is the dynamic scaling factor, and γ is the exponential factor that controls the weights of hard samples. and These represent the bounding boxes predicted by the model and their corresponding ground truth bounding boxes, respectively. By jointly modeling IoU and confidence, AF-IoU achieves unified optimization of geometric scale, prediction confidence, and sample quality, making the weight allocation of the loss function for high- and low-quality predicted boxes more reasonable at different stages.

[0054] Experimental results show that AF-IoU significantly improves detection accuracy at high IoU thresholds (e.g., mAP@50, mAP@95) on multiple datasets, demonstrating its superiority in high-quality bounding box regression tasks. For UAV infrared small target detection, this loss function not only accelerates model convergence but also effectively improves the localization accuracy and bounding box quality of small targets, thereby enhancing overall detection performance. Since AF-IoU can directly replace the CIoU loss function in the YOLO series without modifying the network structure, its versatility and ease of use provide good room for expansion in subsequent research.

[0055] This embodiment uses the publicly available dataset HIT-UAV to train and evaluate IFD-YOLO. The HIT-UAV dataset was created by a research team at Harbin Institute of Technology (HIT) to provide a reliable data foundation for object detection and tracking in nighttime environments. A significant feature of the HIT-UAV dataset is its focus on nighttime surveillance object detection and tracking, utilizing infrared thermal imaging technology to overcome the limitations of traditional visible light cameras in low-light conditions. This dataset contains 2,898 infrared thermal images extracted from 43,470 frames captured by drones, covering various scenes such as schools, parking lots, roads, and playgrounds, and encompassing different lighting conditions from day to night. This embodiment divides the dataset into a training set (2087 images), a validation set (522 images), and a test set (289 images). The HIT-UAV dataset contains 5 categories and a total of 24,899 annotations. The target category distribution for each instance is as follows: Figure 7a As shown, people, cars, and bicycles constitute the main targets. Figure 7bThe dataset showcases the proportions of targets of different sizes in the images, with most targets being very small, occupying only a few dozen pixels. The lack of color and texture features in the images, coupled with the large number of small targets, effectively tested the algorithm for recognizing small infrared targets.

[0056] This embodiment also uses the publicly available dataset IRSTD-1k to evaluate the performance of IFD-YOLO. This dataset consists of 1001 real-world infrared images taken by infrared cameras, with all infrared small targets in the images manually annotated at the pixel level. All images have a resolution of 512×512. IRSTD-1k includes various types of small targets, such as drones, organisms, ships, and vehicles, which were captured at different locations under long-range imaging conditions. This embodiment divides the dataset into a training set (160 images), a validation set (640 images), and a test set (201 images). This dataset is a single-class dataset; all annotated targets belong to the small target category, containing approximately 10,000 target instances. The dataset covers a wide range of scenes, including backgrounds such as sea surfaces, rivers, fields, mountains, cities, and cloudy environments, characterized by severe clutter and noise interference. Therefore, IRSTD-1k can be used to comprehensively evaluate the performance of infrared small target detection algorithms.

[0057] This embodiment is based on the PyTorch 2.6.0 framework, using YOLOv11n as the baseline model. The input image size is set to 512 pixels. The training parameters were set to 512, with a learning rate of 0.01, a batch size of 16, and a training epoch of 200. SGD stochastic gradient descent was used to optimize the model training. The experimental training parameters are shown in Table 1.

[0058] Table 1. Training Parameters

[0059] To evaluate the detection performance of the improved model, this embodiment uses recall (R), precision (P), mean AP (mAP), number of parameters, and model size as performance evaluation metrics for UAV aerial image detection. mAP@50 represents the average AP of all categories when the IoU threshold is 0.5, and mAP@50:95 means that when the IoU threshold is from 0.5 to 0.95, the mAP value at each threshold is calculated at intervals of 0.05, and the average is taken. The corresponding calculation formulas are shown in equations (13) to (16).

[0060] (13) (14) (15) (16) Where TP (True Positives) represents the number of positive classes predicted as positive, FP (False Positives) represents the number of negative classes predicted as positive, FN (False Negatives) represents the number of negative classes predicted as negative, and P(R) is the precision-recall function.

[0061] To evaluate the usability and effectiveness of the improved model in detecting small infrared targets in UAV scenarios, this embodiment conducted comparative experiments with the baseline model under the same conditions. Both experiments were performed from scratch without pre-training. Experimental results show that, compared with the original model, the improved model achieves varying degrees of improvement in accuracy and recall on the HIT-UAV dataset, with mAP@50 and mAP@95 metrics increasing by 4.9% and 3.1%, respectively. Simultaneously, the model's parameter count and computational cost are reduced by 23% and 20%, respectively. This indicates that the improved model not only has lower computational cost and smaller model size, making it more suitable for resource-constrained scenarios like UAV platforms, but also achieves higher detection accuracy. The experimental results validate the effectiveness of the proposed method. The experimental results are shown in Table 2.

[0062] Table 2. Algorithm comparison with the baseline model on HIT-UAV

[0063] Figure 8 and Figure 9 This paper compares the PR curves of the YOLOv11 base model and the improved IFD-YOLO model in this embodiment on the HIT-UAV dataset. Overall, IFD-YOLO outperforms the original YOLOv11 model in detection performance across all categories. The base model's overall mAP@0.5 is 0.810, while IFD-YOLO improves to 0.859, indicating a significant improvement in target detection accuracy. Specifically, IFD-YOLO achieves significant performance improvements in extremely difficult categories such as "Other Vehicle" and "DontCare," and also shows slight improvements in other categories. The PR curve shifts to the upper right, indicating higher precision and recall. This demonstrates that the improved structure proposed in this embodiment effectively enhances the model's ability to extract and distinguish infrared small target features, thereby improving overall detection performance and robustness.

[0064] In addition, to verify the generalization and applicability of the IFD-YOLO model in the infrared small target detection task, this embodiment conducted a comparative experiment with the benchmark model YOLOv11n on the IRSTD-1K dataset, and the results are shown in Table 3. It can be seen that the improved IFD-YOLO outperforms the original model in all metrics, with precision (P) increasing from 81.8% to 85.4%, recall (R) from 60.7% to 66.2%, mAP@50 from 73.3% to 78.3%, and mAP@95 from 34.3% to 36.4%. These results verify that the proposed improved strategy not only possesses the advantages of effectiveness and lightweight design in the infrared small target detection task, but also demonstrates good generalization and practical applicability.

[0065] Table 3. Algorithm comparison with the baseline model on IRSTD-1k

[0066] Figure 10 and Figure 11 This paper presents a comparison of the PR curves of the YOLOv11 base model and the improved model on the IRSTD-1k dataset. Since this dataset contains only a single target category, it more intuitively reflects the model's performance in detecting small infrared targets. The base model's overall mAP@0.5 is 73.3%, while the improved model improves to 78.3%, indicating a significant improvement in detection accuracy. Specifically, the improved model significantly improves recall while maintaining high precision, with the PR curve shifting to the upper right, indicating that the model can capture more targets while reducing false positives. These results demonstrate that the improved structure used in this embodiment effectively enhances the model's feature extraction and representation capabilities, achieving higher accuracy in complex backgrounds and small target detection scenarios while maintaining lightweight design.

[0067] To fully verify the advancement and effectiveness of the IFD-YOLO model, this embodiment compares it with current mainstream models and other infrared small target detection models on the HIT-UAV dataset. The experimental results are shown in Table 4.

[0068] Table 4. Algorithm Comparison Experiment

[0069] As can be seen from the algorithm comparison results in Table 4, the IFD-YOLO model proposed in this embodiment exhibits significant performance advantages across all metrics. Specifically, IFD-YOLO achieves an mAP@50 of 85.9%, which is 4.9%, 9.0%, and 7.3% higher than mainstream models such as YOLOv11n, YOLOv8n, and RT-DETR, respectively. It also achieves 55.2% on the more stringent mAP@95 metric, indicating that this model has stronger representational capabilities in the accurate localization and boundary regression of small targets. Furthermore, the IFD-YOLO model has only 1.97M parameters and a computational cost (GFLOPs) of only 5.0G, far lower than large models such as RT-DETR, demonstrating excellent lightweight characteristics and inference efficiency.

[0070] In addition, when IFD-YOLO is compared with the current state-of-the-art models YOLO-MBL and YOLO-MARS, IFD-YOLO has the highest average detection accuracy, while also having the lowest number of model parameters and computational cost.

[0071] In summary, IFD-YOLO effectively improves detection accuracy and model efficiency while maintaining low computational complexity, demonstrating good practicality and engineering application value.

[0072] To verify the effectiveness of the improved strategy for infrared small target detection on the HIT-UAV dataset, this embodiment designed seven ablation experiments. Using YOLOv11n as the baseline model, and keeping all environments and parameters consistent, the RepViT backbone network, C3k2_DyGhost, and AF-IoU modules were gradually introduced for comparison. Items marked with "√" in the table represent those added to the experiment, thus verifying the effectiveness of each model. The comparison results are shown in Table 5.

[0073] Table 5. Algorithm Ablation Experiment

[0074] As can be seen from the ablation experiment results in Table 5, all the improved modules proposed in this embodiment have a positive impact on the infrared small target detection performance of UAVs. Using YOLOv11n as the baseline model, its mAP@50 and mAP@95 are 81.0% and 52.1%, respectively. After introducing the RepViT backbone network, mAP@50 increases to 83.3%, indicating that RepViT, while maintaining lightweight design, can enhance the network's ability to jointly model local details and global semantics, thereby improving feature representation. Based on the YOLOv11n baseline model, replacing the C3k2 module with C3k2_DyGhost increases mAP@50 to 82.6%, demonstrating that this module effectively enhances fine-grained feature extraction and cross-layer information fusion capabilities by utilizing dynamic feature extraction and a lightweight convolutional structure, making it particularly effective for the identification of infrared small targets. Furthermore, by introducing the RepViT backbone network and replacing the original bounding box regression loss with the adaptive fusion loss function (AF-IoU), the model's mAP@50 and mAP@95 reached 84.2% and 54.5%, respectively, validating the improvement effect of this loss function on boundary localization and target matching accuracy. Notably, when the three improvement strategies are synergistically fused, the constructed IFD-YOLO model achieves 85.9% and 55.2% mAP@50 and mAP@95, respectively, representing improvements of 4.9% and 3.1% compared to the baseline model, with only 2.0M parameters and 5.0G of computation.

[0075] This performance improvement demonstrates the significant complementarity and synergy among the modules: RepViT enhances the representational ability of feature extraction, C3k2_DyGhost improves the preservation and fusion of detailed features, and AF-IoU further optimizes the regression accuracy of the detection boxes. The combined effect of these three modules enables the model to achieve a good balance between detection accuracy, lightweight design, and convergence stability, fully demonstrating the rationality and effectiveness of the improvement strategy in this embodiment.

[0076] This embodiment introduces the RepViT backbone network. This backbone enhances the network's representational capabilities while maintaining its lightweight nature through a convolutional structure re-parameterization mechanism. This enables the model to extract richer and more stable target features from infrared weakly textured backgrounds, thereby significantly improving the perception capability of small targets.

[0077] The improved model did not exhibit false detection issues due to the introduction of the C3k2_DyGhost module. This module utilizes adaptive dynamic convolution and the Ghost feature generation mechanism to achieve efficient feature representation and redundancy suppression. The model can dynamically adjust the convolution kernel weights based on the context of the input image, thereby enhancing its ability to distinguish between complex background regions and real targets, effectively reducing the false detection rate.

[0078] Finally, the images processed by IFD-YOLO did not exhibit overlapping detection boxes, thanks to the AF-IoU loss function used in this embodiment. This loss function introduces a unified IoU constraint mechanism during the localization accuracy optimization process, simultaneously incorporating factors such as regression error, scale mismatch, and angle offset into the optimization objective. This makes the model more balanced and stable during bounding box regression, thereby improving the accuracy and consistency of the detection boxes.

[0079] In summary, the improved model shows significant improvement in issues such as missed detections, false detections, and overlapping detection boxes, verifying the effectiveness of the proposed method in infrared small target detection tasks from the perspective of UAVs.

Claims

1. A method for detecting infrared targets in unmanned vehicles, characterized in that, include: Preprocessing of infrared images collected by unmanned vehicles; The preprocessed image is then input into the improved IFD-YOLO detection model for feature extraction and target detection. Based on the output of the improved IFD-YOLO detection model, the location and category information of the target in the infrared image are obtained; The improved IFD-YOLO detection model includes: RepViT is used as the backbone network to extract local detail features and global semantic features of the image; The processing procedure of the RepViT Block includes: The Token Mixer module performs spatial feature mixing and recombination on the input features. The Token Mixer module is implemented using a 3×3 depthwise separable convolution. The output of the Token Mixer module is as follows: ， in, This represents a 3×3 depthwise convolution operation, and BN stands for batch normalization. It is the ReLU activation function. For input features.

2. The infrared target detection method for unmanned vehicles according to claim 1, characterized in that, The improved IFD-YOLO detection model also includes: Feature extraction is performed using the C3k2_DyGhost module, which integrates dynamic convolution mechanism and Ghost feature generation mechanism. An adaptive fusion IoU loss function is used as the bounding box regression loss function.

3. The infrared target detection method for unmanned vehicles according to claim 1, characterized in that, The RepViT backbone network uses a structural reparameterization mechanism. During the training phase, the RepViT backbone network adopts a multi-branch structure to enhance feature representation, and during the inference phase, the RepViT backbone network merges the branches into a single convolutional path. The RepViT backbone network includes an input stage, multiple feature extraction stages, and a downsampling module. The feature extraction stage contains multiple RepViT Blocks or RepViT SEBlocks.

4. The infrared target detection method for unmanned vehicles according to claim 3, characterized in that, In the RepViT Block, the network uses a combination of channel-wise convolution and point-wise convolution for feature modeling.

5. The infrared target detection method for unmanned vehicles according to claim 3, characterized in that, The processing of the RepViT Block also includes: The Channel Mixer module performs cross-channel feature modeling and nonlinear enhancement on the output of the Token Mixer. The Channel Mixer module consists of two 1×1 convolutional layers and the GELU activation function. When the step size is 1, a residual connection is introduced between the output and input features of the Channel Mixer.

6. The infrared target detection method for unmanned vehicles according to claim 2, characterized in that, The C3k2_DyGhost module includes the DyGhost module, which generates features through the following steps: K sets of convolutional kernel parameters are preset, and corresponding weighting coefficients are generated based on the input features through a lightweight routing branch; The input features are convolved using the K sets of convolution kernels to obtain K sets of feature responses; The dynamic convolution output features are obtained by weighting and summing the K sets of feature responses based on weighted coefficients. The Ghost feature is obtained by concatenating the dynamic convolution output feature with the redundant feature generated by its linear transformation. The Ghost features are batch normalized and then fused with the input features via residual connections to obtain the output features.

7. The infrared target detection method for unmanned vehicles according to claim 6, characterized in that, The preset K sets of convolutional kernel parameters are used to generate corresponding weighting coefficients based on the input features through a lightweight routing branch, including: The lightweight routing branch extracts global information of the input features through global average pooling (GAP), generates weighted coefficients through a linear mapping function, and then normalizes the weighted coefficients through the Softmax function.

8. The infrared target detection method for unmanned vehicles according to claim 2, characterized in that, The adoption of an adaptive fusion IoU loss function as the bounding box regression loss function includes: Scale the predicted bounding box and the ground truth bounding box, and calculate the scaled IoU value. ; Calculate the loss value based on the scaled IoU value and the confidence score (conf) of the predicted bounding box. ,in An exponential factor for controlling the weights of difficult samples.

9. The infrared target detection method for unmanned vehicles according to claim 8, characterized in that, The method of using an adaptive fusion IoU loss function as the bounding box regression loss function also includes: A dynamic annealing strategy is introduced, in which the scaling ratio is controlled by the hyperparameter ratio to change with the training rounds. In the early stage of training, the hyperparameter ratio is larger, so that the model focuses on low-quality prediction boxes to accelerate convergence. In the later stage of training, the hyperparameter ratio is reduced, so that the model focuses on high-quality prediction boxes to improve localization accuracy.

10. The infrared target detection method for unmanned vehicles according to claim 9, characterized in that, The formula for calculating the hyperparameter ratio is: Where T is the total number of training rounds, The time unit refers to the process of the entire training dataset undergoing one forward and backward propagation through the neural network.

Images