Advances in UAV Object Detection: An Improved YOLOv8 Model for Superior Small Object Accuracy

In the Unmanned Aerial Vehicle (UAV) object detection task, small objects constitute a significant portion and are set against a complex environmental background, making feature extraction challenging. Classical detection algorithms exhibit limitations in accurately detecting small objects and fail to meet the accuracy requirements of small objects detection. Therefore, an improved YOLOv8 model for small object detection is proposed. Firstly, this paper introduces a spatial depth conversion convolution module to minimize the loss of small object information typically seen with conventional strided convolutions. This modification helps mitigate feature loss that could otherwise degrade detection performance. Additionally, this paper incorporates an attention mechanism into the backbone network to better concentrate on small objects, thereby improving detection and localization performance. To further optimize performance, a new bounding box regression loss Wise-IoU (WIoU) v3 is adopted. This choice not only enhances the model's generalization capability but also expedites its convergence. Moreover, this paper has designed a new spatial pyramid pooling layer that preserves the texture features of objects while maintaining operational speed. Experimental results are promising, showing that the new method achieves a 47.1% mAP50, which represents a 14.3% improvement over the baseline model. Importantly, these advancements are achieved with only a minimal increase in parameter count.