The application range of unmanned aerial vehicles （UAVs） is constantly expanding，encompassing areas such as military reconnaissance，outdoor photography，power line inspection and other fields. However，at the same time，it has also given rise to a series of social issues. These include concerns about privacy infringement due to UAVs being used for illicit filming and the potential threat to national security posed by the military application of UAVs. Therefore，the research of anti-UAV technology has important practical significance. Infrared UAV target detection technology is a technique that uses infrared imaging to continuously monitor UAVs. It enables UAV target detection based on infrared radiation，and also has obvious advantages in low-light conditions^［1］. In recent years，this technology has become an important research direction and provides an effective complement to radar target detection technology.

However，infrared UAV target detection faces challenges. The distance between the UAV and the sensor makes small infrared targets lack distinctive texture and shape features，hampering the detection. Additionally，background clutter and noise，like clouds and buildings，can cause confusion with obstacles^［2］. Infrared images have high noise，poor spatial resolution，and are sensitive to environmental temperature changes. These challenges add complexity to infrared UAV target detection. Infrared small UAVs are shown in Fig. 1.

In recent years，there has been a continuous emergence of object detection methods based on deep learning，which has achieved impressive detection performance. These methods can be categorized into two types based on how they handle input images. The first type is the two-stage detection method，such as the region-based R-CNN and its variants^［3］. The second type is the one-stage detection methods，including RetinaNet^［4］，SSD^［5］，and YOLO series^［6］. During the flight of UAVs，real-time transmission of infrared images is required for the infrared cameras. In scenarios that demand high real-time performance，methods of the YOLO series，known for their fast speed and high accuracy，have been widely adopted. Among them，YOLOv5 stands out as an advanced detector with strong real-time processing performance and low hardware computing requirements，allowing for easy deployment on mobile devices. Therefore，using YOLOv5 can significantly enhance the detection accuracy and real-time performance of infrared small UAV^［7］.

This paper proposes a UAV detection method based on an improved YOLOv5s model to address the challenges of detecting small targets. The original YOLOv5 structure only includes three feature detection heads，which are not effective in extracting the feature information of small target UAVs captured by infrared cameras at long distances. To address this issue，this paper adds a feature detection head suitable for detecting small targets by YOLOv5. Additionally，the Intersection over Union （IoU） in the original YOLOv5 model is not a good metric for small target detection tasks. Therefore，this paper replaces it with a more suitable metric for small targets called the Normalized Gaussian Wasserstein Distance （NWD） ^［8］. This metric calculates the similarity between bounding boxes using the Gaussian distribution of their corresponding boxes.

YOLOv5 can be categorized into five architectures based on the depth and width of the model：YOLOv5n，YOLOv5s，YOLOv5m，YOLOv5l，and YOLOv5x. Among these，to balance detection speed and accuracy，we choose to improve YOLOv5s. The YOLOv5 network structure consists of three main components. CSP-Darknet53 serves as the backbone feature extraction network，extracting features from the input image. CSP-Darknet53 is an improved version of Darknet53 from YOLOv3，utilizing the Cross-Stage Partial （CSP） network strategy to reduce parameters and computation，thus enhancing the inference speed. In the middle part，YOLOv5 combines two modules：SPPF and PANet ^［9］. The SPPF module increases receptive fields and diversifies the feature pyramid，while the PANet module achieves bottom-up and top-down feature fusion，thereby improving the object detection capability. In the head part，YOLOv5 adopts the head structure of YOLOv4. This structure outputs predictions such as class probabilities，confidence scores，and bounding boxes on the feature maps.

The YOLOv5 model uses a backbone network that undergoes five downsampling stages，producing five feature maps （P1-P5） with resolutions of 1/2，1/4，1/8，1/16，and 1/32 of the input image size，respectively. The neck network combines multi-scale features in a top-down and bottom-up manner without changing the feature map sizes. The detection head operates on the P3-P5 feature maps for object detection. This design is based on the relationship between the feature layer size and the receptive field size in YOLOv5. The receptive field refers to the size of the input image region corresponding to each output unit in a convolutional neural network. A larger receptive field captures more object features，making it suitable for detecting larger objects. On the contrary，a smaller receptive field can only capture a limited number of object features，making it suitable for detecting smaller objects. A smaller receptive field implies that each pixel in the feature map is influenced by a smaller region in the original image. This enables more precise localization of object positions and boundaries，unaffected by irrelevant regions. Additionally，a smaller receptive field corresponds to a larger feature map size，preserving more spatial information and avoiding the loss of fine details on small objects. Therefore，a smaller receptive field is better suited for detecting smaller objects.

We have added a new detection head for small objects after the P2 feature layer in the YOLOv5 model. The detection head operates at a resolution of 160×160 pixels in the P2 layer，which corresponds to two downsampling operations in the backbone network. Each pixel in the P2 layer has a receptive field of 10×10 pixels，which is the smallest receptive field among the P2-P5 feature extraction layers. Additionally，we have assigned different loss function weights to the P2-P5 feature layers based on the target sizes. Specifically，we have assigned a weight of 4 to the feature layers for P2 and P3，a weight of 1 to the feature layer for P4，and a weight of 0.4 to the feature layer for P5. The purpose of this weighting scheme is to enhance the focus on small and tiny objects while reducing overfitting to large objects. The weighted formula for the object loss function is shown in Eq. （1）. Although adding the new detection head increases the computational and memory overhead of the model，it significantly improves the detection performance for small objects. The improved YOLOv5s network architecture is shown in Fig. 2. Small target detection head is shown in Fig. 3.

In YOLOv5，IoU is used as an indicator to measure the degree of matching between predicted bounding boxes and real bounding boxes. It is obtained by calculating the ratio of the intersection area and the union area of the two. However，in the UAV images obtained by infrared devices，the overlapping part of the bounding boxes of small targets is often very small，which will result in lower IoU values. As shown in Fig. 4，IoU is very sensitive to the scale of small targets. For the infrared small UAV target with a size of 6×6 pixels，only a small position deviation can cause IoU to decrease from 0.53 to 0.06，thereby affecting the accuracy of label allocation and reducing the performance of detection^［8］. So，for small target objects，IoU does not measure their matching degree effectively.

We used a metric known as Normalized Gaussian Wasserstein Distance named NWD，which is suitable for small object detection. It is insensitive to the scale of targets，allowing for better assessment of similarities between small objects. Specifically，this method models the object bounding boxes as two-dimensional Gaussian distributions and calculates the NWD between the predicted and the ground truth distributions，as shown in Eq. （2）. Modeling the bounding boxes with Gaussian distributions enables the representation of the object's position as the mean point and its shape as the standard deviation of the Gaussian distribution. This approach has the advantage of better describing the spatial distribution of the objects，not solely relying on their geometric shapes. NWD takes into account the geometric features and spatial distributions between two bounding boxes，even when they do not overlap，by considering their geometric relationships^［10］. This measurement approach avoids the sensitive issues of IoU with small objects，making it more suitable for quantifying small object matches.

where N_A and N_B are Gaussian distributions modeled by A=（cx_A，cy_A，w_A， h_A） and B=（cx_B，cy_B，w_B，h_B），$W_{\mathrm{2}}^{\mathrm{2}}$（N_A，N_B） is the distance measurement，C is the constant closely related to the dataset.

This article is based on the experimental analysis of selected subsets from the 3rd Anti-UAV Workshop & Challenge^［11］. The dataset consists of a total of 9841 small target drone images captured under thermal infrared，involving complex environmental factors such as dynamic backgrounds，complex motions，scale changes，and small targets. The labels in this dataset only include the category of drones. It is noteworthy that the Signal-to-Noise Ratio （SNR） of this dataset is 12.615 dB. The dataset is divided into training and testing sets in a 7：3 ratio. The distribution of the dataset is shown in Fig. 5.

To verify the performance of the model，this article selects Average Precision （AP） to evaluate the model performance，where the AP@0.5 and the AP@0.5：0.95 represent the detection accuracy of two models under different IoU thresholds. The AP@0.5 is the average accuracy of a certain category when IoU is 0.5. The average accuracy is based on different confidence levels，including the curve area of Precision （P） and Recall （R）. And AP@0.5：0.95 is the average accuracy of a certain category when IoU is taken every 0.05 from 0.5 to 0.95. This indicator requires a higher degree of overlap in the target box. The False Alarm Rate （FAR） is depicted in Eq. （6），the Miss Rate （MR） is represented by Eq. （7），with the Average Precision （AP） illustrated in Eq. （8）.

where TP means true positive，TN means true negative，FP means false positive，and FN means false negative.

In this work，all experiments were conducted on the Ubuntu 22.04 operating system，with 128 GB of RAM and an Intel i9-13900K processor. The system was equipped with an NVIDIA RTX 3090 Ti graphics card with 24 GB of video memory； the deep learning framework used was Pytorch 1.12.1； and the programming language was Python 3.10.

The optimization algorithm used for model training was Stochastic Gradient Descent（SGD）. The initial learning rate was 0.01，the momentum was 0.937，and the weight decay coefficient was 0.0005. In addition，the model was trained for 300 epochs，the batch size of the dataset was set to 64.

The experimental results regarding the allocation of different loss function weights for different feature layers are shown in Table 1. From the ablation experimental results in Table 2，it can be observed that adding the NWD metric with a coefficient of 0.5 results in a 2.7% improvement in the AP@0.5 compared to the original model. Furthermore，when using the complete NWD metric，the AP@0.5 is improved by 5.2%. Moreover，introducing a small object detection head with a resolution of 160×160 pixels from the P2 layer also leads to a 3.7% increase in the AP@0.5 compared to the baseline results. Finally，by combining the complete NWD loss function with the small object detection head，the overall model achieves a 7.2% improvement. The improved model also shows a 3.9% increase in the AP@0.5：0.95 compared to the original model.

The performance comparison chart of AP is shown in Fig. 6. The partial experimental result is shown in Fig. 7. To validate the effectiveness of the improved YOLOv5s in this paper，a comparison was made with various mainstream detection networks using the official sub-dataset from the 3rd Anti-UAV Workshop & Challenge. The comparison results are shown in Table 3.

Our algorithm is deployed on the BM1684X TPU，and the specific flow is shown in Fig. 8. Generally speaking，deep learning-based algorithms have two main results in the quantization process：int8 and FP16. According to the characteristics of infrared small targets，learning-based target detection algorithms will have a significant accuracy decline in the quantization process. The experimental results show that when the algorithm accuracy is quantized to FP16，the target detection efficiency is the highest，while the quantization to int8 has the weakest effect to the extent that no valid data can be counted. The deployment experimental results are shown in Table 4.

To address the challenge of small drone detection in infrared devices，this paper proposes a light weight detection model. The model introduces a small object feature extraction layer at the P2 level of the backbone network，connecting it to a high-resolution detection head，thereby enhancing the network's capability to perceive small objects. Additionally，the paper adopts the NWD metric to replace the original IoU-based metric，as the NWD metric provides better measurements for small object instances and improves the model's detection accuracy. The paper conducts several comparative experiments on the partial sub-dataset provided by the 3rd Anti-UAV Workshop & Challenge. The results demonstrate that the proposed model outperforms other mainstream detection models in terms of both AP@0.5 and AP@0.5：0.95 evaluation metrics，validating the effectiveness of the proposed approach. Furthermore，the proposed method maintains high performance levels concerning parameter count，computational complexity，and model weight file size.