Scholarlink

<p>Accurate and timely determination of solid waste composition, categories, and dimensions is vital for effective waste management. However, significant challenges in identifying and sorting diverse, contaminated, and stacked solid waste persist. This study integrates hyperspectral imaging (HSI) with the improved lightweight convolutional neural network (CNN), FreeNet, to enhance waste identification accuracy and efficiency. To facilitate real-world industrial applications, the model architecture was optimized to support efficient batch training on large-scale datasets. The 10-fold cross-validation strategy and grid search strategy were further employed to ensure model stability, and data augmentation was used to enhance hyperspectral image utilization. After training on a self-generated dataset with 10 typical solid wastes (specifically including wood, paper, fabric, plastic, rubber, leather, sponge filler, concrete aggregate, tile, and metal), the model demonstrated superior performance. Experimental results indicate that FreeNet achieves an ultra-fast inference speed of 4&#xa0;ms per image, which is significantly faster than the 523&#xa0;ms of the baseline model, extreme gradient boosting (XGBoost). In terms of classification accuracy, FreeNet achieved a weighted F1-score of 94.04%, outperforming XGBoost (92.42%). Furthermore, FreeNet attained a weighted intersection over union (IoU) score of 89.37% (compared to XGBoost’s 86.79%), showcasing its superior capability for pixel-wise representation of waste areas. Therefore, the fusion of HSI and lightweight CNN holds promising potential for real-time waste identification, providing a robust solution for addressing waste pollution challenges through advanced technologies. This manuscript represents a pioneering application of integrating advanced HSI with lightweight CNN in the field of intelligent solid waste characterization, which holds substantial promise for advancing the field of waste management toward intelligent and efficient waste sorting and treatment strategies. Furthermore, this work also contributes to helping related researchers understand how HSI can be used to measure and examine waste pollutants non-destructively in different scenes and can provide in-situ and real-time intelligent identification.</p> Graphical Abstract <p></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

A rapid method for pixel-wise categorization and quantification of typical solid wastes using hyperspectral imaging and lightweight convolutional neural network

  • 期刊论文

  • 发表日期：

    2026年6月17日

查看全文
  • Tingwei Wu,
  • Pinjing He,
  • Dongying Lan,
  • Fan Lü,
  • Hua Zhang

摘要

Accurate and timely determination of solid waste composition, categories, and dimensions is vital for effective waste management. However, significant challenges in identifying and sorting diverse, contaminated, and stacked solid waste persist. This study integrates hyperspectral imaging (HSI) with the improved lightweight convolutional neural network (CNN), FreeNet, to enhance waste identification accuracy and efficiency. To facilitate real-world industrial applications, the model architecture was optimized to support efficient batch training on large-scale datasets. The 10-fold cross-validation strategy and grid search strategy were further employed to ensure model stability, and data augmentation was used to enhance hyperspectral image utilization. After training on a self-generated dataset with 10 typical solid wastes (specifically including wood, paper, fabric, plastic, rubber, leather, sponge filler, concrete aggregate, tile, and metal), the model demonstrated superior performance. Experimental results indicate that FreeNet achieves an ultra-fast inference speed of 4 ms per image, which is significantly faster than the 523 ms of the baseline model, extreme gradient boosting (XGBoost). In terms of classification accuracy, FreeNet achieved a weighted F1-score of 94.04%, outperforming XGBoost (92.42%). Furthermore, FreeNet attained a weighted intersection over union (IoU) score of 89.37% (compared to XGBoost’s 86.79%), showcasing its superior capability for pixel-wise representation of waste areas. Therefore, the fusion of HSI and lightweight CNN holds promising potential for real-time waste identification, providing a robust solution for addressing waste pollution challenges through advanced technologies. This manuscript represents a pioneering application of integrating advanced HSI with lightweight CNN in the field of intelligent solid waste characterization, which holds substantial promise for advancing the field of waste management toward intelligent and efficient waste sorting and treatment strategies. Furthermore, this work also contributes to helping related researchers understand how HSI can be used to measure and examine waste pollutants non-destructively in different scenes and can provide in-situ and real-time intelligent identification.

Graphical Abstract