A UV Marking and Deep Learning System for Mitigating Textile Environmental Impact

Overview

The garment industry is one of the world’s largest carbon and waste polluters. In the next decade, this industry is expected to produce 150 billion garments/year, while currently recycling ~1%. Garment landfills are growing large enough to be seen from spa ce, while water consumption side effects threaten the environment and human health. The circular economy for textiles is hampered by two challenges – automated fabric sorting and automated tracing. Without automatic fabric identification – scalable recycling measures cannot be put into effect. Without traceability, governments cannot enforce recycling laws and incentives. We propose a solution that leverages low -cost hardware along with deep learning models to create a traceability system. We develop a fabr ic identification component using microscope images classified by Convolutional Neural Networks (MobileNetV2, ResNet101, ResNet50 and VCG16), and a fabric tracing component that marks fabrics with a code visible only under UV light, using YOLOv8 object detection to remain effective in the presence of unique fabric challenges such as creasing and light refraction. We present experiments using several state -of-the-art algorithms and hyper-parameter tuning. Our results show over 90% accuracy for fabric identification across 14 classes and over 0.98 mAP for fabric tracing for new fabrics and over 0.93 mAP after wash cycle. We also demonstrate a prototype robotic arm to automate the fabric marker application. Finally, we provide three datasets for future research . This solution can help create a traceability system that can be implemented worldwide for a textile circular economy. SEL Fusion System: Multisource Digital Biometrics and Stimuli for Early ASD Screening Jingjing Liang The Harker School, San Jose, CA 1 in 36 children in the USA are impacted by autism spectrum disorder (ASD). Less than half receive early intervention and support during this essential neurodevelopmental window. The current diagnosis process is costly, lengthy, and subject to interpretation bias. The SEL Fusion System focuses on identifying objective and computable digital biometrics and designing educational stimuli to provide an accessible and effective ASD early screening system. Through a child-friendly web application utilizing webgazer.js, the SEL Fusion System uses four classes of stimuli: videos, still pictures with audio narration, picture prompted storytelling activities, and games to collect two types of digital biometrics: eye gaze and audio data. Through this child- friendly web site, 1014 experiment data -sets were collected from 108 participants (ASD n=30; General Population (GP) n=78) with computer built-in webcams and microphones. The eye gaze data was converted into heat maps and trained with a VGG16 model. The audio data was processed with Mel Spectrogram feature extraction and then trained with the ECAPA -TDNN model. Through multisource biometrics data collection, processing, and model training, the SEL Fusion System achieved eye-gaze/audio accuracies of 77.6% / 88.6% and stim uli-specific accuracy of 95.5%. The biometrics data under each stimuli class was further evaluated for a wide spectrum of ASD characteristics. The SEL Fusion System is the first system to collect multisource digital biometrics data with multiclass stimuli in non -lab environments and provide accessible early ASD screening and developmental monitoring with minimal cost and lessened stigma.

Competition history

• JSHS 2024 Category not listed

Resources

• National JSHS abstract book ↗

Related projects

Closest projects by meaning, across every fair and year in the corpus.