Small object detection algorithm based on improved YOLOv8s for UAV viewpoints

Object detection under the UAV viewpoint exists problems such as small percentage of object images, dense distribution, uneven categories, as well as limited model size due to hardware environment constraints leading to low detection accuracy. To solve the above problems, this paper selects the YOLOv8s model with smaller model parameters and improves its performance in detecting small object surfaces, and suggests incorporating the EIem module into the backbone network to improve the model's capability in capturing features like boundary contours.. Secondly, in order to construct an improved model based on multiple attention mechanisms, this paper proposes the idea of large-scale separable convolutional attention, and improves the spatial pyramid pooling layer to enhance the fusion between the features at each level, and optimizes the neck network by adding multiple feature layers to enrich the semantic information of small objects. Experiments demonstrate that the algorithm outperforms YOLOv8s by improving mAP50, mAP50:95 on VisDrone dataset by 2.6%, and 1.9%, respectively, compared with YOLOv8s. it can be effectively applied to the task of detecting small objects in UAV viewpoint.