SPD-YOLOv8: an small-size object detection model of UAV imagery in complex scene

Traditional camera sensors rely on human observation. However, in complex scenes, people often experience fatigue when observing objects of various sizes. Moreover, human cognitive abilities have inherent limitations, leading to potential judgment errors. To overcome these challenges, object recognition technology, a pivotal means of categorizing objects captured by camera sensors is introduced. This paper presents a specialized small-size object detection algorithm designed for unique scenarios. This algorithm offers distinct advantages, including enhanced accuracy in detecting small-size objects and improved detection performance for objects of various sizes. The main innovations in this paper include the following three points. Firstly, summarizing a small-size object detection layer and reconfiguring both the feature extraction network and the feature fusion network to enhance the effectiveness of capturing small-size objects. Secondly, SPD-Conv modules are introduced to replace stride convolutions and pooling layers to improve the detection accuracy of small objects. Finally, employing the MPDIoU loss function to enhance the precision of bounding box fitting. In our experiments. We utilized authoritative official datasets. The experimental results on the Visdrone dataset demonstrate a 10.9% increase in mAP0.5 and a 9.3% increase in mAP0.5:0.95 rates compared to the original YOLOv8s. This model not only meets the accuracy requirements but also accounts for the lightweight demands when deployed on embedded devices.