Scholarlink

<p>Underwater object detection suffers from uneven illumination, low contrast, color distortion, and scattering, while limited onboard computing resources demand a balance between accuracy and speed. However, existing lightweight detectors often sacrifice detection precision in challenging underwater conditions. To address this, we propose a lightweight enhancement of the Real-Time DEtection TRansformer (RT-DETR) tailored for underwater environments. First, a lightweight backbone with enhanced global perception is introduced, leveraging the Cross Stage Partial (CSP) concept to eliminate redundant residual convolutional operations. This is combined with a Single-Head Self-Attention (SHSA) mechanism to strengthen global feature modeling, thereby improving detection accuracy for low-contrast targets. Second, a Feature Selection (FS) module based on Channel-wise Adaptive Attention (CAA) is proposed to filter multi-level features extracted by the backbone within a broader receptive field. In addition, a Select Feature Fusion (SFF) module is introduced for bottom-up feature fusion, simplifying the neck’s feature fusion structure and enhancing real-time performance in underwater object detection. Moreover, a Self-Feature Knowledge Distillation (Self-FKD) strategy is employed to enhance detection performance without increasing inference cost, thereby mitigating the degradation of features caused by underwater image quality. Experimental results demonstrate that the proposed model achieves a superior balance between performance and model size compared to several state-of-the-art object detection models across multiple underwater detection datasets, making it well-suited for real-time underwater object detection scenarios.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Lightweight real-time detection transformer with Single-Head Self-Attention and Feature Selection for underwater object detection

  • 期刊论文

  • 发表日期：

    2025年8月26日

查看全文
  • Jiawei Peng,
  • Zhiyu Zhou,
  • Haiyan Wang

摘要

Underwater object detection suffers from uneven illumination, low contrast, color distortion, and scattering, while limited onboard computing resources demand a balance between accuracy and speed. However, existing lightweight detectors often sacrifice detection precision in challenging underwater conditions. To address this, we propose a lightweight enhancement of the Real-Time DEtection TRansformer (RT-DETR) tailored for underwater environments. First, a lightweight backbone with enhanced global perception is introduced, leveraging the Cross Stage Partial (CSP) concept to eliminate redundant residual convolutional operations. This is combined with a Single-Head Self-Attention (SHSA) mechanism to strengthen global feature modeling, thereby improving detection accuracy for low-contrast targets. Second, a Feature Selection (FS) module based on Channel-wise Adaptive Attention (CAA) is proposed to filter multi-level features extracted by the backbone within a broader receptive field. In addition, a Select Feature Fusion (SFF) module is introduced for bottom-up feature fusion, simplifying the neck’s feature fusion structure and enhancing real-time performance in underwater object detection. Moreover, a Self-Feature Knowledge Distillation (Self-FKD) strategy is employed to enhance detection performance without increasing inference cost, thereby mitigating the degradation of features caused by underwater image quality. Experimental results demonstrate that the proposed model achieves a superior balance between performance and model size compared to several state-of-the-art object detection models across multiple underwater detection datasets, making it well-suited for real-time underwater object detection scenarios.