Inshore Ship Detection Based on Convolutional Neural Network in Optical Satellite Images

Fei Wu, Zhiqiang Zhou*, Bo Wang, Jinlei Ma

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review

55 Citations (Scopus)

Abstract

In this paper, we present a novel inshore ship detection method based on convolutional neural network (CNN). Different from current inshore ship detection methods that need complex shape and texture analysis or sea and land segmentation, our method starts from a global search for the relatively distinct ship head with an efficient classification network. This can help to obtain the location of possible ship heads as well as the rough ship directions, which are beneficial to generate smaller and more precise candidate regions of ship targets. Compared with other region proposal methods, our method can produce a rather smaller set of proposals. Next, iterative bounding-box regression and classification are unified into a multitask network, which is constructed and trained specially by considering the practical condition of the inshore ships in remote sensing images. At last, nonmaximum suppression is applied to eliminate duplicate detections. Experiments on optical satellite images demonstrate the effectiveness and robustness of the proposed method for inshore ship detection.

Original languageEnglish
Article number8490712
Pages (from-to)4005-4015
Number of pages11
JournalIEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Volume11
Issue number11
DOIs
Publication statusPublished - Nov 2018

Keywords

  • Convolutional neural network (CNN)
  • inshore ship detection
  • iterative bounding-box regression
  • optical satellite images

Fingerprint

Dive into the research topics of 'Inshore Ship Detection Based on Convolutional Neural Network in Optical Satellite Images'. Together they form a unique fingerprint.

Cite this