Semi-Supervised Salient Object Detection for Side-Scan Sonar Images via Entropy-Based Uncertainty and Contrastive Learning

Recently, the field of salient object detection (SOD) in underwater sonar images has gradually gained attention. Benefiting from the advancements in deep learning technologies and insights from specific tasks, semi-supervised learning has demonstrated promising results for salient object detection i...

Full description

Saved in:
Bibliographic Details
Main Authors: Shenao Yuan, Zhen Wang, Fu-Lin He, Shan-Wen Zhang
Format: Article
Language:English
Published: IEEE 2025-01-01
Series:IEEE Access
Subjects:
Online Access:https://ieeexplore.ieee.org/document/11059963/
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Recently, the field of salient object detection (SOD) in underwater sonar images has gradually gained attention. Benefiting from the advancements in deep learning technologies and insights from specific tasks, semi-supervised learning has demonstrated promising results for salient object detection in side-scan sonar (SSS) images. However, current methods often face the following challenges: 1) scarce data annotation; 2) complex and dynamically changing underwater environments; 3) difficulty in distinguishing similar target and background features; 4) unclear boundaries with high uncertainty. To address these challenges, we propose a semi-supervised SOD method for SSS image based on entropy uncertainty contrastive learning (EUCL). Our method aims to improve the quality of pseudo-labels and the overall performance of the model using a small amount of labeled data. Specifically, we design an uncertainty map evaluation module that effectively extracts global contextual information through an entropy-based partition loss and a contrastive learning mechanism, integrated with the uncertainty map to handle the complex underwater background. Moreover, we introduce an uncertainty weighting mechanism to enhance the detection performance of salient objects in different regions. Comparative experiments on the underwater scene SSS dataset against seven state-of-the-art semi-supervised methods demonstrate that our method achieves outstanding performance, with the mean absolute error (MAE) of 7.89% and an accuracy (Acc) of 94.69%. Implementation codes will be available on <uri>https://github.com/darkseid-arch/EUCL</uri>
ISSN:2169-3536