Scholarlink

<p>The alarming issue of plastic waste polluting natural inland aquatic environments necessitates the development of automated trash detection techniques that work effectively in complex and dynamic inland water environments. Several deep learning techniques have been developed for detecting floating trash but suffer due to objects’ tiny size, occlusions, complex backgrounds, etc. Hence, the paper proposes a novel waste detection technique, YOLO-Ecofocus, employing YOLOv9-gelan-c architecture, augmented with an efficient attention mechanism and parallel Shifted Window (Swin) transformer block for detecting small floating plastic waste bottles. The work derives the convolutional channel and spatial attention network (CSA-Net) and strategically integrates it at different architectural locations along with varied color features to specifically focus on intrinsic characteristics of small, littered, transparent and partially submerged bottles. The Swin transformer block uniquely processes features through a parallel path alongside the RepNCSPELAN block to concatenate multi-scale and attention features crucial for small waste bottle detection. The comparisons of the proposed work against 14 baseline YOLO variants and 16 state-of-the-art existing techniques indicate the proposed technique’s outperforming results with the highest precision, mean Average Precision (mAP@50) and mean of mAP (mmAP@50:95) of <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(84.9\%\)</EquationSource> </InlineEquation>, <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(86.9\%\)</EquationSource> </InlineEquation>, and <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(49.1\%\)</EquationSource> </InlineEquation>. The higher mAP@50 and mmAP@50:95 values indicate significant improvements in emphasizing the model’s potential to detect small and transparent waste bottles in challenging real-time environmental monitoring, thereby contributing to United Nations’ Sustainable Goals.</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

YOLO-EcoFocus: small waste bottle detection in inland water bodies focusing on sustainable environmental cleaning using USVs

  • 期刊论文

  • 发表日期：

    2025年11月21日

查看全文
  • B. Saranya Devi,
  • Rimjhim Padam Singh,
  • Angelina George,
  • S. Anirudh,
  • P. Radha Nishant

摘要

The alarming issue of plastic waste polluting natural inland aquatic environments necessitates the development of automated trash detection techniques that work effectively in complex and dynamic inland water environments. Several deep learning techniques have been developed for detecting floating trash but suffer due to objects’ tiny size, occlusions, complex backgrounds, etc. Hence, the paper proposes a novel waste detection technique, YOLO-Ecofocus, employing YOLOv9-gelan-c architecture, augmented with an efficient attention mechanism and parallel Shifted Window (Swin) transformer block for detecting small floating plastic waste bottles. The work derives the convolutional channel and spatial attention network (CSA-Net) and strategically integrates it at different architectural locations along with varied color features to specifically focus on intrinsic characteristics of small, littered, transparent and partially submerged bottles. The Swin transformer block uniquely processes features through a parallel path alongside the RepNCSPELAN block to concatenate multi-scale and attention features crucial for small waste bottle detection. The comparisons of the proposed work against 14 baseline YOLO variants and 16 state-of-the-art existing techniques indicate the proposed technique’s outperforming results with the highest precision, mean Average Precision (mAP@50) and mean of mAP (mmAP@50:95) of \(84.9\%\) , \(86.9\%\) , and \(49.1\%\) . The higher mAP@50 and mmAP@50:95 values indicate significant improvements in emphasizing the model’s potential to detect small and transparent waste bottles in challenging real-time environmental monitoring, thereby contributing to United Nations’ Sustainable Goals.