Scholarlink

<p>To address the issues of insufficient feature extraction and low tracking accuracy in underwater dynamic target tracking, this paper presents an underwater dynamic target tracking algorithm based on CWSCNet-RCAM (Channel-Weighted Skip Connection Network with Residual-Convolution Attention Module). This approach integrates the channel-weighted skip connection network (CWSCNet) and the residual convolution attention module (RCAM), constructing a “encoding-fusion-attention-decoding” four-level architecture. Initially, this architecture achieves feature enhancement, heterogeneity elimination, and interference suppression of underwater degraded images through the collaborative effect of channel-weighted skip connections and RCAM, thereby improving tracking accuracy and robustness end-to-end. Subsequently, CWSCNet adopts two skip connection strategies, namely channel-weighted summation or channel-weighted concatenation, to fuse low-level and high-level feature maps: the summation operation is suitable for scenarios with large-scale differences in features but complementary semantics, while the concatenation operation is suitable for scenarios that need to retain complete information at multiple levels. The choice of these two strategies is based on the degree of feature heterogeneity and the need for information richness and eliminates feature heterogeneity by minimizing the numerical differences between feature maps, thereby improving the accuracy of feature extraction. Finally, by introducing RCAM, combined with residual networks, channel attention, and spatial attention modules, it effectively suppresses the influence of background interference, target occlusion, and deformation on tracking, further enhancing the feature extraction ability. To verify the effectiveness of the suggested approach, experiments are conducted on the UOT32 datasets and indoor and outdoor underwater scenes. The experimental results show that, compared with SiamFC, SiamRPN, ATOM, DiMP18, DiMP50, SiamRPN++, PrDiMP18, and PrDiMP50 algorithms, the tracking accuracy of the suggested algorithm has increased by 33.1%, 24.2%, 20.9%, 15.4%, 13.1%, 10.2%, 12.5%, and 10.8%, respectively. This fully validates the effectiveness and real-time performance of the employed approach.</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

Underwater dynamic target tracking algorithm based on CWSCNet-RCAM

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Yu Zhang,
  • Xuyang Zhang,
  • Jining Peng

摘要

To address the issues of insufficient feature extraction and low tracking accuracy in underwater dynamic target tracking, this paper presents an underwater dynamic target tracking algorithm based on CWSCNet-RCAM (Channel-Weighted Skip Connection Network with Residual-Convolution Attention Module). This approach integrates the channel-weighted skip connection network (CWSCNet) and the residual convolution attention module (RCAM), constructing a “encoding-fusion-attention-decoding” four-level architecture. Initially, this architecture achieves feature enhancement, heterogeneity elimination, and interference suppression of underwater degraded images through the collaborative effect of channel-weighted skip connections and RCAM, thereby improving tracking accuracy and robustness end-to-end. Subsequently, CWSCNet adopts two skip connection strategies, namely channel-weighted summation or channel-weighted concatenation, to fuse low-level and high-level feature maps: the summation operation is suitable for scenarios with large-scale differences in features but complementary semantics, while the concatenation operation is suitable for scenarios that need to retain complete information at multiple levels. The choice of these two strategies is based on the degree of feature heterogeneity and the need for information richness and eliminates feature heterogeneity by minimizing the numerical differences between feature maps, thereby improving the accuracy of feature extraction. Finally, by introducing RCAM, combined with residual networks, channel attention, and spatial attention modules, it effectively suppresses the influence of background interference, target occlusion, and deformation on tracking, further enhancing the feature extraction ability. To verify the effectiveness of the suggested approach, experiments are conducted on the UOT32 datasets and indoor and outdoor underwater scenes. The experimental results show that, compared with SiamFC, SiamRPN, ATOM, DiMP18, DiMP50, SiamRPN++, PrDiMP18, and PrDiMP50 algorithms, the tracking accuracy of the suggested algorithm has increased by 33.1%, 24.2%, 20.9%, 15.4%, 13.1%, 10.2%, 12.5%, and 10.8%, respectively. This fully validates the effectiveness and real-time performance of the employed approach.