Scholarlink

<p>In the scenario of water surface garbage detection, the model must accurately detect different types of objects and be able to respond continuously within a short time frame, enabling timely retrieval by the surface cleaning robot. Therefore, this paper proposes a surface garbage detection model named Precise and Real-time YOLO (PAR-YOLO), with a focus on real-time performance and detection accuracy. Firstly, to reduce model computation and improve detection efficiency, the Ghost Bottleneck module is designed and utilized in the backbone section as a replacement for the traditional Bottleneck module. Secondly, in order to effectively reduce the interference of factors such as water ripples, lighting variations, or reflections on object feature recognition, we have designed a Noise Suppression Module (NSM) and integrated it into the neck section. Lastly, to enhance the model’s attention to challenging samples and improve detection accuracy, the Varifocal Loss function is employed in the head section. Experimental results demonstrate that the PAR-YOLO model achieves a Frames Per Second (FPS) of 238, with a mean average precision (mAP) of 85.53%, 47.3%, and 28.5% on our self-made water surface garbage dataset, the Flow public water surface garbage dataset and the Pascal VOC2007 dataset, respectively. Compared to other comparative models, our model achieves the best results.</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

PAR-YOLO: a precise and real-time YOLO water surface garbage detection model

  • 期刊论文

  • 发表日期：

    2025年1月2日

查看全文
  • Ning Li,
  • Mingliang Wang,
  • He Huang,
  • Bo Li,
  • Baohua Yuan,
  • Shoukun Xu

摘要

In the scenario of water surface garbage detection, the model must accurately detect different types of objects and be able to respond continuously within a short time frame, enabling timely retrieval by the surface cleaning robot. Therefore, this paper proposes a surface garbage detection model named Precise and Real-time YOLO (PAR-YOLO), with a focus on real-time performance and detection accuracy. Firstly, to reduce model computation and improve detection efficiency, the Ghost Bottleneck module is designed and utilized in the backbone section as a replacement for the traditional Bottleneck module. Secondly, in order to effectively reduce the interference of factors such as water ripples, lighting variations, or reflections on object feature recognition, we have designed a Noise Suppression Module (NSM) and integrated it into the neck section. Lastly, to enhance the model’s attention to challenging samples and improve detection accuracy, the Varifocal Loss function is employed in the head section. Experimental results demonstrate that the PAR-YOLO model achieves a Frames Per Second (FPS) of 238, with a mean average precision (mAP) of 85.53%, 47.3%, and 28.5% on our self-made water surface garbage dataset, the Flow public water surface garbage dataset and the Pascal VOC2007 dataset, respectively. Compared to other comparative models, our model achieves the best results.