[Paper Review] Virtual Mines -- Component-level recycling of printed circuit boards using deep learning

Muhammad Mohsin, Stefano Rovetta+2 more|arXiv (Cornell University)|Jun 24, 2024

Recycling and Waste Management TechniquesEnvironmental Science3 citations

TL;DR

This paper proposes a deep learning-based system using YOLOv5 to enable component-level recycling of printed circuit boards (PCBs) by detecting electronic components in a custom V-PCB dataset. The model achieves 82% precision and 55% recall, demonstrating effective real-time detection for urban mining in circular economy applications.

ABSTRACT

This contribution gives an overview of an ongoing project using machine learning and computer vision components for improving the electronic waste recycling process. In circular economy, the "virtual mines" concept refers to production cycles where interesting raw materials are reclaimed in an efficient and cost-effective manner from end-of-life items. In particular, the growth of e-waste, due to the increasingly shorter life cycle of hi-tech goods, is a global problem. In this paper, we describe a pipeline based on deep learning model to recycle printed circuit boards at the component level. A pre-trained YOLOv5 model is used to analyze the results of the locally developed dataset. With a different distribution of class instances, YOLOv5 managed to achieve satisfactory precision and recall, with the ability to optimize with large component instances.

Motivation & Objective

• To address the growing challenge of electronic waste (e-waste) by enabling high-efficiency, cost-effective recovery of critical raw materials (CRMs) from end-of-life PCBs.
• To develop a computer vision pipeline that supports component-level recycling of PCBs, reducing reliance on traditional mining and minimizing environmental impact.
• To create a local, high-quality V-PCB dataset annotated with eight electronic component classes for training and evaluating deep learning models.
• To optimize object detection performance on resource-constrained hardware (Jetson Nano) for real-time deployment in recycling facilities.
• To improve model generalization through data augmentation, especially for underrepresented component classes in the dataset.

Proposed method

• A custom V-PCB dataset of PCB images was collected from recycling plants and preprocessed using OpenCV for image enhancement, contrast adjustment, and region-of-interest extraction.
• Eight electronic component classes (e.g., capacitors, resistors, ICs) were annotated using Label Studio to create a supervised learning dataset.
• A pre-trained YOLOv5 object detection model was fine-tuned on the V-PCB dataset using a high-performance GPU server, with training conducted at batch sizes of 4 and 8.
• Model inference was deployed on an NVIDIA Jetson Nano for real-time performance evaluation on edge devices.
• Data augmentation techniques were applied to balance class distribution and improve model robustness for low-frequency component types.
• Performance was evaluated using precision, recall, mean average precision (mAP), and loss metrics (box, obj, cls) at both bounding box and class levels.

Experimental results

Research questions

• RQ1 Can a deep learning-based object detection model achieve high precision and recall for detecting diverse electronic components on PCBs in a real-world recycling context?
• RQ2 How does class imbalance in the V-PCB dataset affect the performance of YOLOv5, and can data augmentation improve detection for underrepresented components?
• RQ3 To what extent can a YOLOv5 model trained on a custom dataset generalize and perform efficiently on edge hardware like the Jetson Nano for real-time PCB component detection?
• RQ4 What is the impact of different batch sizes on the convergence and performance of the YOLOv5 model during training on the V-PCB dataset?
• RQ5 Can the proposed pipeline support scalable, cost-effective component-level recycling of PCBs, contributing to the circular economy and urban mining goals?

Key findings

• The YOLOv5 model achieved a precision of 82% and recall of 55% on the V-PCB dataset, indicating strong detection performance for most component classes.
• Components with higher instance counts in the training data showed better detection performance, highlighting the importance of data balance.
• The model demonstrated feasible real-time inference on the Jetson Nano, a low-power edge device, supporting deployment in field recycling environments.
• mAP (mean average precision) was 0.56, with mAP-95 at 0.27, indicating moderate overall detection accuracy across all classes.
• Data augmentation significantly improved model generalization, especially for rare component types with low initial instance counts.
• The system successfully demonstrated the feasibility of using deep learning for automated, component-level recycling of PCBs, supporting the concept of 'virtual mines' in the circular economy.