Ship Segmentation in Aerial Images for Maritime Surveillance

January 29, 2021

In this work, we propose a new method capable of real-time ship segmentation during maritime surveillance missions using onboard cameras. To perform segmentation, we implement a two stages cascade model with a detection part followed by a segmentation stage. In the first stage, we select the most likely regions to contain a ship, and then we segment those regions to identify the targets. Including a detection stage, we narrow the image region to be segmented, which significantly improves the overall image processing time. Plus, it decreases the number of miss detection and improves overall segmentation performance. We apply this method in the Seagull dataset for the first time. Additionally, we study the viability of using post-processing techniques, like conditional random fields, to improve the segmentation performance. We notice that it does not improve the performance when we use the cascade model, but it increases the segmentation score when we apply segmentation to the whole image. We train the model with aerial ship images from detection and segmentation maritime datasets. However, we have to label the detection images to be used in the segmentation stage. We test the cascade model on the Airbus ship detection challenge, and it produces fast and accurate maritime ship segmentation results capable of real-time segmentation.

Authors in the community:

• 

  C

  Carlos David Coito Pires

  ist425321

Supervisors of this institution:

• 

  A

  Alexandre José Malheiro Bernardino

  ist13761

• 

  B

  Bruno Duarte Damas

  ist45387

electrical-engineering-electronic-engineering-information-engineering - Electrical engineering, electronic engineering, information engineering

eng - English

Instituto Superior Técnico