An Optimized Deep Learning Model for Tiny Object Detection in UAV Imaging

Abstract

Recent advancements in deep learning models-based Unnamed Aerial Vehicle (UAV) object detection technologies have garnered significant interest in smart cities. The tiny object detection task is still a crucial challenge in research due to variant image resolution and training sample size. The primary aim of this paper is to realize precision and model generalization in multi-scale object detection tasks. An optimization of the YOLOv8 (You Only Look Once version 8) deep learning model was carried out to detect 16 classes of tiny objects in UAV imagery data with the assistance of the transfer learning technique. The optimization method's workflow consists of two main procedures; the first procedure aimed to fine-tune the YOLOv8 model's hyperparameters and adopted the Rectified Linear Unit (ReLU) activation function in the model architecture instead of the Sigmoid Linear Unit (SiLU) activation function for feature map generation. Afterword, the fine-tuned YOLOv8 model is optimized further by an open-source optimization workspace. Open Visual Inference & Neural Network Optimization (OpenVINO) to accelerate the training/inference performance along with getting more accurate detection of tiny objects in the input imagery samples. The proposed framework's performance evaluation was conducted using the Dataset for Object Detection in Aerial Images DOTA-v1.5 dataset. The DOTA dataset has been augmented and balanced to generate a customized dataset to mitigate the effect overfitting problem and get better detection accuracy. The results of the experiment showed a significant improvement in small object detection, achieving a 23.67% increase in inference speed while maintaining a higher detection accuracy.