Enhancing Performance of IR Ship Detection with Baseline AI Models over a new Benchmark

Abstract

Classical vision applications mainly extract features from IR ship images using saliency or region based segmentation algorithms. However due to the complex scenes having intensity in-homogeneity, sea clutter and noise, success of these algorithms is limited. In order to bridge this gap, in this paper we focused on automatic IR ship detection and semantic segmentation problem with the perspective of deep learning based algorithms. At first we have enhanced the previously published IRShips1 dataset to IRShips2 by increasing 18 scenes to newly labeled 178 diverse set of scenes. Then for the ship semantic segmentation task we have trained a UNet architecture with a ResNet based backbone. We have used transfer learning and data augmentation to overcome the limitations of small size of dataset. Experimental results show significant improvement in both efficiency (FPS) and accuracy (IoU and F-Score) as compared to the existing GBVS based algorithm on both IRShips1 and IRShips2 datasets. Finally we have also labeled our dataset for bounding box based detection and showed benchmark results (mAP) of two state of the art object detectors i.e.yoloV4 and yoloV5. We hope that our new IRShips2 dataset along with results of proposed baseline AI models will serve a useful benchmark for the research community.