Two-Stage Preprocessing Approach for Background Normalization and Defect Segmentation on Denim Fabric Image Analysis

The textile industry is one of the oldest and largest industries in the world. Detecting defects early in manufacturing enables cost reduction through fast intervention and fabric correction. Vision-based software can automatically identify defects in fabrics during preor post-production. However, this detection is challenging due to denim fabric color and texture variations, making some defects less visible than others, such as double thread and cut weft. A robust preprocessing approach can reveal intrinsic features of the data. Therefore, this paper investigates background normalization and defect segmentation to highlight features and allow better differentiation during image analysis. We propose a two-stage pipeline using a combination of filters to reduce background information and merge it with the output of a defect segmentation process based on thresholding and morphological operations. To test this hypothesis, we extracted 100 features from the processed and unprocessed images and benchmarked them using several machine learning algorithms, including a deep learning model. With the processed data, we achieved up to a 7% increase in accuracy for the evaluated metrics compared to the unprocessed experiments. This study demonstrates the feasibility of detecting defective regions in denim fabric using computer vision techniques.