Every year human discharges about 350 million tons of plastic waste into the environment and can be projected to triple in 2060 without any attempts to change situation. From 1970 to 2019, an estimation of 130 million tons of plastic waste was accumulated into the rivers, lakes and sea, while only 27 % is recycled and utilized. Moreover, waste treatment plants in most places around the world are using out-of-date technology, may pose a threat to the health of the workers. Therefore, it is essential to modernize these systems for protecting human health. This paper proposes fine-tuning DETR, which applies Artificial Intelligent in plastic waste sorting system. Consequently, this study analyzed the applicability of fine-tuning DETR in the domain of plastic waste categorization and its potential drawbacks. For fair experiment and evaluation, model candidates were trained and evaluated on an industrial plastic waste dataset. The fine-tuning DETR outperformed other candidates in the context of critical indicators, from accuracy (25.1 mAP), processing speed (28 FPS) to computational cost (GFLOPs 86). Furthermore, fine-tuning DETR possesses the capability of autonomous operation without requiring human intervention, distinguishing this candidate from other prevalent algorithms. Our research demonstrates that, fine-tuning DETR specifically and Transformer-based algorithms in general, are entirely suitable and hold significant potential for large-scale application in holistic plastic waste sorting systems.

中文翻译：

---

微调 DETR：迈向塑料垃圾分类系统的整体流程

人类每年向环境排放约 3.5 亿吨塑料垃圾，如果不尝试改变现状，预计到 2060 年将增加两倍。从1970年到2019年，估计有1.3亿吨塑料垃圾被累积到河流、湖泊和海洋中，而只有27%被回收利用。而且，世界上大多数地方的废物处理厂都使用过时的技术，可能对工人的健康构成威胁。因此，必须对这些系统进行现代化改造，以保护人类健康。本文提出微调DETR，将人工智能应用于塑料垃圾分类系统。因此，本研究分析了微调 DETR 在塑料废物分类领域的适用性及其潜在缺点。为了公平的实验和评估，候选模型在工业塑料废物数据集上进行了培训和评估。微调 DETR 在关键指标方面优于其他候选方案，从准确性 (25.1 mAP)、处理速度 (28 FPS) 到计算成本 (GFLOPs 86)。此外，微调DETR具有无需人工干预的自主操作能力，这使该候选算法与其他流行算法区分开来。我们的研究表明，具体微调 DETR 和一般基于 Transformer 的算法是完全适合的，并且在整体塑料废物分类系统中具有大规模应用的巨大潜力。