AI for Fabric Defect Detection

Academic AI Case Study

Paradda Keeter | June 23, 2026

Introduction

This project applies no-code image classification to textile quality control by labeling fabric images as normal or defective. Defects such as stains, holes, tears, and weave irregularities increase rework, scrap, and customer complaints if they are not caught early. A computer vision model can support faster first-pass inspection and help quality teams focus on uncertain cases. This use case is strategically important because quality problems affect cost, efficiency, and brand reputation. If a company detects visible defects earlier, it can reduce waste, improve consistency, and support a better customer experience.

Dataset Description

The model used two classes: Normal fabric and Defective fabric. The project draws on the AITEX Fabric Image Database, a public textile inspection dataset with 140 defect-free images and 105 defect images. Version 1 was trained with 25 images per class. Version 2 expanded the training set to 40 images per class to give the model more variety in texture and defect appearance. Following the assignment instructions, I used uploaded images rather than the webcam.

Results and Analysis

The results show that adding more relevant examples improved performance. Version 1 reached 92% accuracy, while version 2 reached 100% accuracy on the test set. What worked best was keeping the task simple with two classes and using obvious defects. A limitation is that perfect accuracy on a small, controlled test set does not guarantee the same result in real production conditions.

Version 1

Version 2

Business Case and ROI Estimate

This model could reduce inspection bottlenecks and catch visible defects before fabric moves to later production stages. Earlier detection lowers scrap, rework, and complaint risk while helping quality-control teams inspect more consistently.

As an illustrative estimate, assume a facility processes 100,000 yards of fabric per month, defect-related loss affects 2% of output, and the average loss is $1.50 per yard. That equals about $3,000 in monthly defect cost. If AI-assisted screening reduces those losses by 20%, the savings would be about $600 per month, or $7,200 per year, before softer benefits such as faster inspection and stronger brand trust.

Ethics and Scalability

The main risks are false negatives, false positives, and dataset bias. A false negative could allow defective fabric to pass inspection, while a false positive could trigger unnecessary waste or re-inspection. Bias may appear if the training images overrepresent certain colors, textures, or defect types. Companies should also avoid uploading confidential, proprietary, or personally identifiable data into external tools.

To scale this solution, a company would need a larger and more representative image dataset, standardized image capture, stronger model validation, and integration with quality-control workflows. Enterprise deployment would also require image storage, monitoring for model drift, retraining, and human review for uncertain predictions.

Key Learnings

This lab showed that even a no-code AI platform can support a meaningful business application of computer vision. I learned that data quality, class balance, and image consistency strongly affect performance, and that adding more relevant images can improve accuracy. From a management perspective, AI creates value when it is tied to a real business problem, measured against outcomes, and deployed responsibly.

Source note: AITEX Fabric Image Database (official dataset source). Google Teachable Machine was used as the no-code image classification platform.