Transfer Learning for Maritime Vessel Detection using Deep Neural Networks

Abstract

Reliable vessel detection can improve safety and security in maritime environment. Recently, application of Deep Learning (DL)-based detectors have become popular in autonomous vehicles. The aim of this paper is to study how much a pretrained DL model on a domain-specific marine data can improve the performance of the detectors for vessel detection? To this end, we trained state-of-the-art DL-based detectors (Faster R-CNN [1], R-FCN [2] and SSD [3]) using an open source generic object detection COCO dataset [4] and a marine SeaShips dataset [5]. The performance of these detectors are explored based on different feature extractors. Moreover, we investigate the effect of object size on the detection accuracy. To obtain results, we collected a real marine dataset by a sensor system onboard a vessel in the Finnish archipelago. This system is used for developing autonomous vessels, and records data in a range of operation and climatic conditions. The experimental results show that Faster R-CNN with ResNet101 achieves the highest object detection accuracy with mean average precision of 75.2%.