1 of 1

Quality Control of Textile Garments Through Computer Vision

Agustin Martíenz García Paul Antonio Quintal Peralta

Universidad Anahuac Mayab

Introduction and Justification

Manual quality control in the textile industry is often inefficient and prone to human error due to operator fatigue, leading to financial losses and inconsistent product quality. This project addresses the need for an automated solution at "La Rana Uniformes," implementing a cost-effective

computer vision system using the NVIDIA Jetson Nano and YOLOv8.

This system detects logo defects in real-time and automatically rejects defective garments without pausing production, offering a scalable solution for SMEs

General Objective

Evaluate the feasibility of implementing a low-cost industrial vision system to automate the quality inspection of screen-printed logos, thereby reducing defect rates and operational costs

Specific Objectives

Design a vision system to evaluate logo quality parameters.
Integrate an automatic rejection mechanism for defective products.
Develop a user-friendly interface for operator monitoring .

Métodos

Methods The project followed a structured engineering workflow divided into four stages:

Hardware Architecture: Selected the NVIDIA Jetson Nano for Edge AI processing and designed a custom PETG mount to protect the Arducam 64MP camera from high oven temperatures .
Data Pipeline: Collected and annotated a dataset of over 300 images using Roboflow to train a specialized YOLOv5 object detection model .
Hybrid Logic: Enhanced the AI model with "Common Sense" algorithms (ROI filters, size gating) to eliminate false positives and ensure reliability .
Integration: Deployed a software ecosystem ("AI Factory" & "Production App") integrated with GPIO actuators for the automatic physical rejection of defective garments .

Results and Discussion

The final prototype demonstrated exceptional stability, achieving continuous operation at approximately 25 FPS and 60°C with zero system crashes, thanks to a 2-thread architecture optimization.

Regarding accuracy, the implementation of a "Hybrid AI" (combining AI with logic filters) significantly improved detection rates from 90% in the baseline model to over 99% in the final system, successfully eliminating false positives such as "ghost" detections.

Furthermore, the system ensured full traceability by generating unique session folders, saving dual evidence photos (Crop/Context) for every defect, and logging all events in a CSV file .

Conclusions

The project successfully automated the quality control process, offering a reliable, non-intrusive solution. The integration of YOLO v8 with algorithmic logic proved superior to standard models. The system is modular, cost-effective, and ready for industrial deployment, successfully meeting the objective of reducing human error and operating costs for "La Rana Uniformes."

Acknowledgments

We thank La Rana Uniformes for granting access to their facilities and Ms. Nidelvia Natali Cruz Huicab for her guidance. Special thanks to Professor Geovanny Giorgana Macedo for his initial support.