Scholarlink

<p>This study presents the AI-powered Waste Sorting Apparatus (AI-WSA), which integrates deep learning with advanced electromechanical design and system-level interfacing to automate the segregation of plastic, paper, and metal waste. Addressing the global waste crisis, the system employs a two-phase transfer learning protocol to evaluate three pre-trained architectures—NASNet-Mobile, EfficientNet-B0, and MobileNetV3-Large—on an NVIDIA Jetson Nano. EfficientNet-B0 achieved the highest performance, with a classification accuracy of 91.6%, balanced precision of 91.67% and recall of 91.62%, and inference speeds under 200&#xa0;ms, while MobileNetV3-Large delivered excellent throughput with 21–24 items/min, but with minimal accuracy loss. Although NASNet-Mobile exhibited stable performance with 89.6% validation accuracy, it suffered from excessive latency. The AI-WSA features a compact, corrosion-resistant design that operates at <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11227_2025_7931_Article_IEq1.gif" Format="GIF" Height="15" Rendition="HTML" Resolution="72" Type="Linedraw" Width="19" /> </InlineMediaObject> <EquationSource Format="TEX">\(\le\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>≤</mo> </math></EquationSource> </InlineEquation>5&#xa0;W with ISO 10218-2-compliant throughput. Grad-CAM analysis confirmed the models’ reliance on specific key discriminative features, despite ongoing difficulties in plastic classification. This work demonstrates the feasibility of edge AI in resource-constrained settings and offers a scalable framework for intelligent recycling infrastructure that supports circular economy principles and advances Sustainable Development Goal (SDG 12.5), which focuses on waste reduction through recycling and reuse. By leveraging an edge GPU (NVIDIA Jetson Nano) and NVIDIA’s TensorRT optimizations, the proposed system achieves sub-200&#xa0;ms inference latency within a 5-W power envelope. This demonstrates supercomputing-grade performance on a low-power platform, enabling real-time waste segregation through high-performance parallel processing.</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

Deep transfer learning for sustainable waste management: Real-time waste segregation apparatus using a two-phase CNN framework

  • 期刊论文

  • 发表日期：

    2025年10月6日

查看全文
  • Natheer Almtireen,
  • Abdelrazzaq A. Abuhejleh,
  • Mutaz Ryalat,
  • Hisham Elmoaqet,
  • Ghaith Al-refai

摘要

This study presents the AI-powered Waste Sorting Apparatus (AI-WSA), which integrates deep learning with advanced electromechanical design and system-level interfacing to automate the segregation of plastic, paper, and metal waste. Addressing the global waste crisis, the system employs a two-phase transfer learning protocol to evaluate three pre-trained architectures—NASNet-Mobile, EfficientNet-B0, and MobileNetV3-Large—on an NVIDIA Jetson Nano. EfficientNet-B0 achieved the highest performance, with a classification accuracy of 91.6%, balanced precision of 91.67% and recall of 91.62%, and inference speeds under 200 ms, while MobileNetV3-Large delivered excellent throughput with 21–24 items/min, but with minimal accuracy loss. Although NASNet-Mobile exhibited stable performance with 89.6% validation accuracy, it suffered from excessive latency. The AI-WSA features a compact, corrosion-resistant design that operates at \(\le\) ≤ 5 W with ISO 10218-2-compliant throughput. Grad-CAM analysis confirmed the models’ reliance on specific key discriminative features, despite ongoing difficulties in plastic classification. This work demonstrates the feasibility of edge AI in resource-constrained settings and offers a scalable framework for intelligent recycling infrastructure that supports circular economy principles and advances Sustainable Development Goal (SDG 12.5), which focuses on waste reduction through recycling and reuse. By leveraging an edge GPU (NVIDIA Jetson Nano) and NVIDIA’s TensorRT optimizations, the proposed system achieves sub-200 ms inference latency within a 5-W power envelope. This demonstrates supercomputing-grade performance on a low-power platform, enabling real-time waste segregation through high-performance parallel processing.