Session: 07-03-01 Arctic Frontier Regions

Paper Number: 123922

123922 - Ibnet: A Comprehensive Iceberg Semantic Segmentation Network With Multi-Scale Integration and Attention Mechanisms 

Accurate semantic segmentation of iceberg images plays a pivotal role in polar ship navigation and ecological surveys. With the challenges posed by extended observation cycles and signal instability due to satellite remote sensing, as well as signal delays affecting real-time analysis, a cognitively enabled algorithm for ship-mounted iceberg recognition using optical sensors is needed to enhance real-time risk perception in polar navigation channels. However, the presence of adverse weather conditions and limited resources, combined with the similarity in textures among different ice images, presents significant challenges, potentially leading to navigation errors.This paper aims to address these challenges by developing a precise pixel-level iceberg image segmentation algorithm, IBNet, designed for challenging sea ice optical vision conditions. We propose a novel feature extraction network that leverages multi-scale features and a double-layer attention mechanism within the pyramid pooling module layer for weighted feature fusion. This learning-based approach takes into account six distinct iceberg backgrounds in the polar region scene and enhances the reliability of the semantic segmentation algorithm through data augmentation. Through a comprehensive comparison of various performance metrics, including Pixel Accuracy (PA) and Mean Intersection over Union (MIoU), the IBNet algorithm demonstrates significant advantages in semantic segmentation of icebergs in the challenging polar environment.This algorithm will find applications in subsequent multi-source information fusion, dynamic navigation situational awareness, and the development of a polar navigation VR system.

Presenting Author: Wangyuan Zhao 哈尔滨工程大学

Presenting Author Biography: Wangyuan Zhao received the B.S. degree in engineering from Harbin Engineering University in 2018. He is currently pursuing studies in ship and marine structure design and manufacturing, with a specialization in the intelligent application of optical sensors in underwater and surface environments. His research areas encompass the utilization of deep learning techniques for optical image analysis in underwater robotics and perception and parameter measurement in highly complex underwater environments.

Authors:

Fenglei Han Harbin Engineering University
Wangyuan Zhao Harbin Engineering University
Yanzhuo Xue Harbin Engineering University
Yuliang Wu Harbin Engineering University
Xiao Peng Harbin Engineering University
Jiawei Zhang Harbin Engineering University