This study addresses the challenges of automated fabric inspection, specifically complex background textures and limited annotated samples. It aims to develop a lightweight, robust defect inspection framework that fuses complementary hand-crafted features. The goal is to offer a high-accuracy alternative to resource-heavy deep learning models, suitable for industrial environments with computational and data constraints.
The proposed framework applies local binary pattern (LBP) transformation to enhanced images, extracting gray level co-occurrence matrix (GLCM) statistics and histogram of oriented gradients (HOG) descriptors. These features are fused, compressed via principal component analysis (PCA), and classified using a support vector machine (SVM). For localization, GLCM features are clustered by K-means and refined by a Markov Random Field (MRF) model. A six-class dataset was constructed for validation.
The method achieved 94.94% classification accuracy and 95.15% defect detection accuracy. It significantly outperformed single-feature baselines and conventional classifiers. Crucially, the framework demonstrated superior performance compared to representative deep segmentation networks when operating under small-sample conditions and comparable runtime constraints, validating its efficacy for real-world industrial application.
While effective, reliance on hand-crafted features may require specific parameter tuning for substantially different textile textures. The study implies that hybridizing traditional computer vision with shallow learning remains a vital research direction, particularly for scenarios where the massive annotated datasets required for deep learning are unavailable.
The proposed lightweight framework allows for deployment on hardware with limited computational power, reducing implementation costs. Its ability to perform accurately with limited annotated samples significantly lowers the barrier to entry for textile manufacturers, minimizing the manual effort required for data labeling and system setup.
Automating defect detection reduces the physical strain and visual fatigue associated with manual inspection, promoting better working conditions. Additionally, higher consistency in quality control minimizes material waste in textile production, contributing to more sustainable manufacturing practices.
This paper presents a novel fusion of LBP, GLCM and HOG features combined with MRF-refined localization. It effectively bridges the gap between weak single-feature methods and resource-heavy deep networks, offering a unique, efficient solution for high-precision inspection in data-constrained settings.
