Mathematical Modeling and Fractal Geometry for Microtexture Fabric Analysis

TMLR Paper5886 Authors

Abstract: Automated microtexture analysis of textile materials is critical for scalable fabric characterization in industrial quality control and high-throughput processing. We introduce a reproducible pipeline that employs Raspberry Pi microscope imaging, robust preprocessing with augmentation and imaging condition simulation, and encodes each fabric sample as a 40-dimensional feature vector. This vector captures statistical, edge, Haralick/GLCM, LBP, fractal, wavelet, Tamura, and morphological descriptors, supplemented by fractal fitting overlays that yield interpretable surface roughness and complexity maps. We release an open dataset of 20 fabric types with 500 high-resolution images, paired feature vectors, raw microscopy data, and fractal overlay visualizations. Experimental results show strong 20-class fabric classification performance with hybrid features improving macro F1 by 7% over handcrafted-only baselines, and improved unsupervised defect detection: Isolation Forest achieves ROC AUC = 0.780 with precision = 0.820, recall = 0.750, and F1 = 0.780, balancing false positives and detection rate. Our work provides a transparent, extensible framework for computational materials science, AI-driven quality control, and educational use in automated textile analysis. Code and dataset link removed for anonymity.

Submission Length: Regular submission (no more than 12 pages of main content)

Changes Since Last Submission:

- Explained statistics that were provided, and detailed entropy calculations
- Included results for "deep-features-only" and "handcrafted-only" setups
- Enlarged fonts on some figures for better visibility
- Revised tables to show unique contributions per feature group (added means, std devs, confidence intervals, and significance tests from multiple runs with different seeds)
- Included results for all mentioned models like XGBoost, LightGBM, and etc.
- Made figure/table captions standalone and descriptive
- Fixed any inconsistencies with number of images and classes
- Changed wording of writing in some areas to focus on appeal to TMLR audience in applied ML and materials science

Assigned Action Editor: ~Chinmay_Hegde1

Submission Number: 5886