Ship Detection in Optical Remote Sensing Images Using YOLOv4 and Tiny YOLOv4

With the advances in remote sensing domain, images with higher spatial and spectral resolution are obtained from increasing number of sensors, and they have been employed in more research fields, including object detection and tracking. In particular, the detection of marine vehicles has a significant role in civil and military applications. However, due to the varying type, size, posture, and complex background of the ships to be detected, ship target detection is still considered as a challenging task. Deep learning techniques with their wide-spread use in computer vision applications have been successfully applied to object detection problems that is important to monitor marine traffic and ensure maritime safety. In this study, a freely available aerial image dataset is utilized to train and test the two popular single-stage object detection models, namely YOLOv4 and Tiny YOLOv4, based on the "You Only Look Once" approach. Produced results were analyzed using conventional accuracy metrics, and average prediction times were also compared. The trained models were evaluated on different ship images and detections were performed. As a result of the study, mean average precision (mAP) values of 80.82% and 62.30% were obtained using YOLOv4 and Tiny YOLOv4 architectures, respectively. This indicates major performance difference between YOLOv4 and Tiny YOLOv4 models for ship detection studies.