.png){kind=icon}
Return
EFLSDet: Efficient forward-looking sonar object detector
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2026.134025.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
•
A forward-looking sonar object detector EFLSDet with 0.39M parameters, is proposed.
•
RSEC blocks inject physical priors via reparameterization for efficient extraction.
•
LRFWT module isolates high-frequency noise and refines global target features.
•
Real-world AUV deployment achieves 83.94 FPS, balancing accuracy and efficiency.
Abstract
Object detection in forward-looking sonar (FLS) images is a core perception task for autonomous underwater vehicles (AUVs). Prior studies have often treated FLS images as conventional grayscale images, achieving performance gains through generic feature representation improvements while overlooking the intrinsic characteristics specific to sonar targets. To alleviate this issue, we first analyze two characteristics of sonar images: richer low-level visual information and more important global representation. Then, based on these characteristics, we propose an Efficient Forward-Looking Sonar Object Detector, named EFLSDet, which contains two novel components: a set of Reparameterized Semantic Enhancement Convolution (RSEC) blocks and a Large Receptive Field Wavelet Transform (LRFWT) module. The RSEC blocks achieve adaptive enhancement of various view-specific semantic information by initializing convolution kernels with different handcrafted operators and combining reparameterization techniques to maintain the computational efficiency of a single convolution kernel. The LRFWT module achieves global feature modeling and refinement through large convolution kernels and wavelet transforms, respectively. Comprehensive experiments on public and custom-built datasets demonstrate that EFLSDet achieves state-of-the-art performance and efficiency. Furthermore, we deployed EFLSDet in real-world field environments, where it runs at 83.94FPS with only 0.39M parameters on an NVIDIA Jetson AGX Xavier, proving its competitive efficiency for resource-constrained underwater exploration. The code will be made publicly available upon publication.
Keywords:
Forward looking sonar
Autonomous underwater vehicles
Object detection
Structural reparameterization
Wavelet module
Journal
IF:
7.5
Papers:
3.0W
Citations:
10.2W
Organization
Cited Papers
.png){kind=icon}
ShareSaveRepCrack: An efficient pavement crack segmentation method based on structural re-parameterization
PREAI
ShareSave
ShareSaveWTAPNet: Wavelet Transform-Based Augmented Perception Network for Infrared Small-Target Detection
PREAI
ShareSave
View more