Scholarlink

<p>Modern waste management faces challenges making recycling processes more efficient. Precise sorting is an essential part and could be optimized utilizing artificial intelligence (AI). However, the implementation of such AI-based object recognition systems is challenging due to the lack of representative training data, the acquisition of which is often practically difficult or typically associated with high labor and costs. This article presents a&#xa0;novel procedural alpha-matting based dataset generation approach that creates new compositions and labels simultaneously using limited real data. Unlike conventional cut-and-paste methods that rely on hard pasting or edge blending, the presented approach takes the realistic representation of transparent objects such as plastics into account using alpha estimates. To demonstrate the applicability of the approach, RGB images from a&#xa0;typical sorting scenario with a&#xa0;camera mounted orthogonally above the conveyor belt are used, in which various objects should be automatically recognized using AI. The experiments conducted show that modern AI models trained with purely generated data achieve very good localization performance on real data. Overall, this work demonstrates the potential of alpha-matting based synthetic data generation as a&#xa0;cost-effective solution for acquiring additional data to improve waste management systems and highlights the challenges in recognizing different types of plastics in sorting applications.</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

Synthetische Generierung von Trainingsdaten für kontrollierbares Lernen von Bilderkennungsverfahren im Recycling

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Sebastian Schreiner,
  • Daniel Scheuchenstuhl,
  • Bernd Hoedl,
  • Julian Aberger,
  • Thomas Pock

摘要

Modern waste management faces challenges making recycling processes more efficient. Precise sorting is an essential part and could be optimized utilizing artificial intelligence (AI). However, the implementation of such AI-based object recognition systems is challenging due to the lack of representative training data, the acquisition of which is often practically difficult or typically associated with high labor and costs. This article presents a novel procedural alpha-matting based dataset generation approach that creates new compositions and labels simultaneously using limited real data. Unlike conventional cut-and-paste methods that rely on hard pasting or edge blending, the presented approach takes the realistic representation of transparent objects such as plastics into account using alpha estimates. To demonstrate the applicability of the approach, RGB images from a typical sorting scenario with a camera mounted orthogonally above the conveyor belt are used, in which various objects should be automatically recognized using AI. The experiments conducted show that modern AI models trained with purely generated data achieve very good localization performance on real data. Overall, this work demonstrates the potential of alpha-matting based synthetic data generation as a cost-effective solution for acquiring additional data to improve waste management systems and highlights the challenges in recognizing different types of plastics in sorting applications.