Lightweight Edge–Frequency Driven Real-Time Detection Transformer for side-scan sonar target detection

The persistent challenge of accurately and efficiently detecting objects underwater using side-scan sonar (SSS) continues to drive innovation within the ocean intelligence space. SSS, while the dominant imaging tool, produces images plagued by noise and blurred boundaries, hindering reliable target recognition. This new research, introducing the Lightweight Edge–Frequency Driven Real-Time Detection Transformer (LEF-RT-DETR) framework, represents a significant step forward in addressing these limitations. The development builds upon existing advancements in transformer-based detection, as explored in related work like Improved Transformer-based detection of underwater plastic debris in complex environments, demonstrating a growing trend towards utilizing these powerful architectures for increasingly complex underwater scenarios. The focus on both accuracy and real-time performance is particularly noteworthy, recognizing the practical constraints of operational deployments.

The LEF-RT-DETR framework’s ingenuity lies in its modular design, specifically tailored to the characteristics of SSS imagery. The Gaussian-Edge Enhancement Module (GEEM) directly tackles the issue of blurred target edges by integrating Gaussian smoothing and edge extraction, allowing the model to better perceive these critical features. Similarly, the Multi-Scale Frequency-Spatial Denoising Block (MFDB) effectively mitigates noise interference by fusing spatial and frequency domain information. This approach is a compelling example of leveraging domain-specific knowledge to enhance model performance. Furthermore, the Partial Convolution with Efficient Channel Attention (PCCA) optimizes computational efficiency without sacrificing accuracy, a crucial consideration for resource-constrained platforms. The reported 24% reduction in parameters and 18% reduction in computational cost compared to RT-DETR, alongside a 4.3% improvement in Average Precision (AP) and 5.3% improvement in AP50, highlights the framework’s efficacy. This resonates with broader efforts to improve the efficiency of ocean monitoring systems, as discussed in IntroductionUnderwater plastic debris detection remains challenging in visually cluttered environments because underwate, where computational resources are often a limiting factor.

The development of LEF-RT-DETR has broader implications for numerous applications, ranging from maritime security and underwater infrastructure inspection to marine archaeology and environmental monitoring. Reliable real-time target detection in challenging underwater environments is essential for autonomous underwater vehicles (AUVs) and remotely operated vehicles (ROVs) to perform their tasks effectively and safely. The framework’s ability to operate efficiently could facilitate wider deployment of these technologies, enabling more comprehensive data collection and analysis. The use of a self-constructed dataset, while allowing for focused evaluation, also suggests an opportunity for future research involving larger, more diverse datasets to further validate the framework’s robustness across different operational conditions and sonar systems. The emphasis on empirical validation – clearly demonstrated through the quantitative results – aligns with the core values of scientific authority and measurable impact that define our approach to ocean intelligence.

Looking ahead, a key question is how LEF-RT-DETR can be integrated with other sensor modalities, such as acoustic Doppler current profilers (ADCPs) or multi-beam echo sounders, to create a more holistic understanding of the underwater environment. The development of integrated data ecosystems, as we advocate for, will be crucial for unlocking the full potential of ocean data. Furthermore, exploring the framework’s adaptability to different sonar frequencies and imaging geometries will be important for expanding its applicability across a wider range of underwater scenarios. The ongoing refinement of these deep learning models, combined with advances in hardware and data processing capabilities, promises to revolutionize our ability to explore, understand, and ultimately protect our oceans.