Scholarlink

<p>Accurate underwater object detection plays an important role in intelligent aquaculture systems, including fish monitoring, feeding management, and automated underwater inspection. However, underwater images in aquaculture environments often suffer from blurriness, low contrast, and degraded structural details due to light attenuation, color distortion, noise interference, and water scattering. These degradations significantly reduce the effectiveness of traditional detection algorithms and increase the risk of false positives and false negatives in complex aquatic environments. To address these challenges, this study presents an optimized underwater object detection framework, DIR-YOLOv11, based on YOLOv11n. Primarily, the DualConv module replaces the original C3K2 structure to ensure more continuous feature representation and improve the extraction of discriminative target features. Second, an iEMA attention mechanism is incorporated into the detection layer for small objects to strengthen the responsiveness of the model to key features in dense and occluded scenarios. Furthermore, the RFAConv dynamic convolution substituted the conventional convolution in the neck network, adaptively modeling receptive fields to better capture local regions and deformed targets. Ultimately, the Shape-GIoU bounding box regression loss is introduced to mitigate gradient discontinuity under conditions without overlap and enhance the modeling of target shape consistency. Experimental results on the URPC2018 and DUO underwater datasets demonstrate that DIR-YOLOv11 achieves 80.5% mAP50 and 45.5% mAP50:95 on URPC2018, improving the YOLOv11n baseline by 3.6% and 3.4%, respectively. On the DUO dataset, DIR-YOLOv11 reaches 83.3% mAP50 and 62.1% mAP50:95, corresponding to improvements of 3.1% and 2.7%, respectively, over the baseline. Meanwhile, the model achieves an inference speed of 135 FPS, demonstrating strong potential for real-time deployment in aquaculture monitoring and underwater inspection systems.</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

DIR-YOLOv11: an underwater target detection algorithm based on feature enhancement and dynamic receptive field convolution

  • 期刊论文

  • 发表日期：

    2026年7月7日

查看全文
  • Jiao Wang,
  • Juncheng Guo

摘要

Accurate underwater object detection plays an important role in intelligent aquaculture systems, including fish monitoring, feeding management, and automated underwater inspection. However, underwater images in aquaculture environments often suffer from blurriness, low contrast, and degraded structural details due to light attenuation, color distortion, noise interference, and water scattering. These degradations significantly reduce the effectiveness of traditional detection algorithms and increase the risk of false positives and false negatives in complex aquatic environments. To address these challenges, this study presents an optimized underwater object detection framework, DIR-YOLOv11, based on YOLOv11n. Primarily, the DualConv module replaces the original C3K2 structure to ensure more continuous feature representation and improve the extraction of discriminative target features. Second, an iEMA attention mechanism is incorporated into the detection layer for small objects to strengthen the responsiveness of the model to key features in dense and occluded scenarios. Furthermore, the RFAConv dynamic convolution substituted the conventional convolution in the neck network, adaptively modeling receptive fields to better capture local regions and deformed targets. Ultimately, the Shape-GIoU bounding box regression loss is introduced to mitigate gradient discontinuity under conditions without overlap and enhance the modeling of target shape consistency. Experimental results on the URPC2018 and DUO underwater datasets demonstrate that DIR-YOLOv11 achieves 80.5% mAP50 and 45.5% mAP50:95 on URPC2018, improving the YOLOv11n baseline by 3.6% and 3.4%, respectively. On the DUO dataset, DIR-YOLOv11 reaches 83.3% mAP50 and 62.1% mAP50:95, corresponding to improvements of 3.1% and 2.7%, respectively, over the baseline. Meanwhile, the model achieves an inference speed of 135 FPS, demonstrating strong potential for real-time deployment in aquaculture monitoring and underwater inspection systems.