DNTFE-Net: Distant Neighboring-Temporal Feature Enhancement Network for side scan sonar small object detection

Abstract

Listen

Translate article

Side-scan sonar small target detection becomes fundamental work for side-scan sonar applications which holds vital importance for marine engineering, maritime military, and so on. However, existing algorithms suffer from performance degradation and poor portability when sonar data is limited. We find that benefit from the strong temporal correlation of sonar images, background information from the long-term images and foreground information from the short-term images can be used to enhance small targets in the current image. To fully utilize these two cues, this paper presents the Distant Neighboring-Temporal Feature Enhancement Network (DNTFE-Net) for detecting small objects in sonar images. The algorithm is based on the signal acquisition process of sonar devices, which is considered a continuous time sequence. Firstly, a flexible foreground feature alignment module is developed to align the foreground information between neighboring signals and the current signal so that foreground information can be used to enhance small target features. Then, we propose a background comparison module for selecting distant-temporal signals that provide the most contextual information in order to improve the ability of migrating to different domains. Finally, a foreground background feature amalgamation module is designed to fuse neighboring and distant-temporal signals for enhancing small target features. Furthermore, real sonar data is collected to compose the real continuous sonar dataset. Our DNTFE-Net exhibits superior recognition performance on the real sonar dataset compared to other popular methods, with an improvement of at least 27%. Our algorithms are publicly available at https://github.com/zbyhnu/DNTFE-Net.git.