Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (13): 200-208.DOI: 10.3778/j.issn.1002-8331.2312-0153

• Graphics and Image Processing • Previous Articles     Next Articles

RepViTS-YOLOX:Underwater Blurred and Occluded Target Detection Method

TAO Yang, ZHU Teng, ZHONG Bangqian, ZHOU Kun, ZHOU Liqun   

  1. School of Communications and Information Engineering, Chongqing University of Posts and Telecommunications, Chongqing 400065, China
  • Online:2024-07-01 Published:2024-07-01

RepViTS-YOLOX:水下模糊及遮挡目标检测方法

陶洋,朱腾,钟邦乾,周昆,周立群   

  1. 重庆邮电大学 通信与信息工程学院,重庆 400065

Abstract: The RepViTS-YOLOX method is proposed to address target blurring and occlusion in underwater target detection, improving upon the YOLOX framework. This approach employs RepViTS for feature extraction and incorporates structural reparameterization to enhance underwater target feature extraction and model inference speed. It features a spatial and channel reconstruction convolution (SCConv) module to better focus on blurred targets. The feature fusion network uses cross-scale and multi-scale fusion to better interpret occluded target characteristics. Additionally, a task-specific context decoupling head (TSCODE) is introduced for more accurate localization and classification of occluded targets. Experiments on the RUOD dataset show that RepViTS-YOLOX achieves 85.7% detection accuracy, surpassing YOLOX by 3.8 percentage points. These findings indicate that the method effectively improves the detection of blurred and occluded underwater targets, thus enhancing the precision of underwater target detection.

Key words: YOLOX, target detection, structural re-parameterization, decouple detection head, attention mechanism

摘要: 针对水下目标检测中的目标模糊和遮挡问题,提出基于YOLOX改进的RepViTS-YOLOX水下目标检测方法。该方法采用RepViTS作为特征提取网络并通过结构重参数化,有效提升了对水下目标的特征提取能力和模型推理速度。引入空间和通道重构(spatial and channel reconstruction convolution,SCConv)模块,增强网络对模糊目标的关注。改进特征融合网络,通过跨尺度连接和多尺度融合,加强不同层次特征间的信息交流,使模型更好理解遮挡目标特征。针对定位和分类任务对特征的不同需求,引入上下文解耦头(task-specific context decoupling head,TSCODE),对遮挡目标更精准地定位和分类。实验结果证明,RepViTS-YOLOX方法在RUOD数据集上取得了85.7%的检测效果,较YOLOX提高了3.8个百分点。检测结果显示,该方法可以有效改善水下模糊和遮挡目标的检测情况,提高水下目标检测精度。

关键词: YOLOX, 目标检测, 结构重参数化, 解耦检测头, 注意力机制