Abstract
In underwater target detection tasks, side-scan sonar (SSS) is currently the most widely used imaging tool. However, due to the inherent imaging mechanism of sonar and the complexity of the underwater environment, SSS images often suffer from blurred target boundaries and strong noise interference, which significantly increases the difficulty of high-precision underwater target recognition. We propose a Lightweight Edge–Frequency Driven Real-Time Detection Transformer (LEF-RT-DETR) framework, aiming to improve both the detection accuracy and real-time performance for target detection in SSS images. Specifically, to address the distortion of target edge contours in SSS images, we design a Gaussian-Edge Enhancement Module (GEEM) by integrating Gaussian smoothing with edge extraction to enhance the model’s capability of perceiving target edge features. Simultaneously, to address severe noise interference in SSS images, we introduce the Multi-Scale Frequency-Spatial Denoising Block (MFDB), a feature fusion module integrating spatial-domain and frequency-domain information to improve the model’s ability to distinguish noise from targets. Finally, we propose the Partial Convolution with Efficient Channel Attention (PCCA), which reduces redundant channel computations and enhances inter-channel interactions via channel attention, thereby reducing model complexity while preserving high detection accuracy. Experimental results on a self-constructed dataset demonstrate that LEF-RT-DETR achieves improvements of 4.3% in Average Precision (AP) and 5.3% in Average Precision at an Intersection over Union (IoU) threshold of 0.50 (AP50) compared with the Real-Time Detection Transformer (RT-DETR), while reducing the number of parameters and computational cost by approximately 24% and 18%, respectively. This significantly enhances model accuracy and efficiency, providing robust support for real-time underwater target detection tasks.