Enhanced Waste Segmentation

The rise in global waste production presents a significant threat to both people and the environment. To combat this issue, it is necessary to implement resource-friendly waste separation and efficient recycling of recyclable materials. The utilization of convolutional neural networks shows great potential, particularly for separating various types of waste. However, recent developments indicate that state-of-the-art object segmentation models are insufficient in dealing with the complexity and diversity of recycling waste. This thesis presents a low-cost electric actuator system (EAS) in the form of a vibration plate. Additionally, a waste segmentation dataset is created to train, validate, and test two Mask R-CNN and two YOLOv8 segmentation models with varying specifications. The dataset comprises 947 manually annotated images, with 22,630 annotated objects belonging to cardboard, hard plastic, metal, and soft plastic classes. It was augmented using various techniques and enriched with complexity, resulting in 5,716 images and 137,324 annotated instances. The vibration generated by the EAS was effective in separating waste objects of different materials to a significant extent. This resulted in an increase in the mean average precision for segmentation mask (mAP mask) of a Mask R-CNN model with R101-FPN-3x backbone from 26.5% to 60.9%. The other models also demonstrated a significant improvement in mAP mask: Mask R-CNN R50-FPN-1x (24.4% to 58.8%), pre-trained YOLOv8 nano (24.9% to 50.6%) and non-pre-trained YOLOv8 nano (20.2% to 51.1%). It has been confirmed that the use of EAS improves the results of convolutional neural networks. In the future, this system could be used as a sub-process in automated waste separation. The GitHub repository for the project can be found here.全球垃圾产量的激增对人类与环境均构成严重威胁。为应对这一难题，亟需推行资源友好型垃圾分类，并对可回收材料开展高效回收利用。卷积神经网络（convolutional neural network）的应用展现出巨大潜力，尤其适用于多品类垃圾的分类识别。然而现有研究表明，当前顶尖的目标分割模型在应对回收垃圾的复杂性与多样性时仍存在不足。本论文提出了一种低成本的电动执行器系统（electric actuator system, EAS），其形态为振动盘。此外，本研究构建了一套垃圾分割数据集，用于训练、验证与测试两款规格各异的Mask R-CNN模型与两款YOLOv8分割模型。该数据集初始包含947张人工标注图像，共计22630个标注对象，涵盖硬纸板、硬质塑料、金属与软塑料四个类别。通过多种数据增强技术对数据集进行扩充并提升其复杂度后，最终得到5716张图像与137324个标注实例。EAS产生的振动可在较大程度上有效分离不同材质的垃圾对象，使得搭载R101-FPN-3x骨干网络的Mask R-CNN模型的分割掩码平均精度均值（mean average precision for segmentation mask, mAP mask）从26.5%提升至60.9%。其余模型的mAP mask也均实现显著提升：Mask R-CNN R50-FPN-1x模型从24.4%提升至58.8%，预训练YOLOv8 nano模型从24.9%提升至50.6%，非预训练YOLOv8 nano模型从20.2%提升至51.1%。实验证实，引入EAS可有效提升卷积神经网络的分割效果。未来，该系统可作为自动化垃圾分类流程中的一个子工序加以应用。本项目的GitHub代码仓库可在此处获取。