Compressed YOLOv5 for Oriented Object Detection with Integrated Network Slimming and Knowledge Distillation

Abstract

In recent years, object detection has been expanded to drone scenes, where remote sensing images contain a greater variety and arbitrary-oriented targets. In order to solve the problem of detection difficulty and computational intensity for remote sensing images, oriented object detection is needed and the network model is expected to be deployed on resource-limited devices. This paper proposes a lightweight object detection method for oriented object detection by leveraging and compressing YOLOv5 network model. We integrate the fine-tuning stage in network slimming with knowledge distillation to enhance the accuracy of the detection model and save training time by transferring the important feature information to the student network. Loss function is redesigned by combining Theta loss with other detection and distillation losses to make the compression model more accurate. Extensive experiments are conducted to verify the effectiveness of our proposed method on the remote sensing public dataset DOTA. The compressed model achieves an accuracy of 76.18% on the DOTA dataset, 1.7% increase compared to the original YOLOv5 model. The FLOPs are decreased by 37.0%, the number of parameters is decreased by 58.9%, the weight file size is decreased by 57.6%, and the inference time is decreased by 17.4%.