Abstract
While the proliferation of Unmanned Aerial Vehicle (UAV) technology has driven significant industrial progress, unauthorized low-altitude flights have posed increasingly severe challenges to public safety and airspace order. Within anti-UAV defense frameworks, infrared imaging sensors have emerged as a core technology for target surveillance in complex environments, owing to their all-weather detection capabilities and excellent anti-interference characteristics. However, constrained by infrared imaging mechanisms and environmental noise, infrared images suffer from inherent defects such as low signal-to-noise ratios (SNR), weak contrast, and a lack of detailed textures. Especially when UAVs are at medium-to-long distances or against cluttered backgrounds, targets often appear as blurry thermal radiation patches. This prevents existing algorithms from maintaining stable detection rates and makes it difficult to accurately distinguish specific UAV configurations from limited pixels. Precise identification of UAV types is critical; it not only assists the anti-UAV system in assessing potential threat levels but also provides a vital decision-making basis for deploying targeted electronic jamming or physical interception. To overcome these limitations and meet the demand for "detection as identification" in anti-UAV systems, this paper proposes a UAV target detection and classification method in infrared images based on an improved YOLO11. This method primarily enhances detection performance through two stages of optimization. First, a targeted preprocessing mechanism is introduced at the data input end, utilizing the multi-scale Top-Hat operator to effectively suppress background interference noise and enhance the local contrast of the target area. This significantly improves image clarity and feature saliency, providing a high-quality data foundation for subsequent network discrimination. Second, the core architecture of the YOLO11 algorithm is optimized by embedding Deformable Convolution (DCN) modules into the feature network. This focuses on enhancing the model's ability to extract and analyze infrared targets and key structural contours, allowing it to acutely capture the subtle feature differences necessary for distinguishing various UAV types. Experimental verification on a multi-scenario infrared dataset demonstrates the effectiveness of the proposed method. Results show that the algorithm maintains a high success rate in complex detection tasks while effectively reducing the risks of false alarms and missed detections. More importantly, it breaks through the limitations of traditional infrared detection that only "bounds" the target, successfully achieving precise identification of specific UAV types (e.g., fixed-wing, quadrotor, etc.). This research significantly improves the comprehensive performance of infrared detection systems in object discovery and fine-grained classification, demonstrating promising engineering application prospects.