PIXELSEGNET-CNN: A HYBRID DEEP LEARNING FRAMEWORK WITH OPTIMIZED PRE-PROCESSING FOR HIGH-PRECISION TEXTILE DEFECT SEGMENTATION

Automated textile defect detection is essential for quality control in manufacturing; however, it presents challenges due to variability in defects, intricate textures, and the need for real-time processing. This research presents PixelSegNet-CNN, an innovative deep learning framework that combines advanced pre-processing techniques (Gaussian Blur, Gabor filters, Local Binary Pattern) with lightweight encoder-decoder architecture for pixel-wise defect segmentation. The model was assessed using three benchmark datasets such as TILDA 400, FABRIC STAIN, and AITEX and demonstrated state-of-the-art performance, exceeding established methods (U-Net, ResNet, CNN, Mask R-CNN) in terms of accuracy (91.0%), precision (83.0%), recall (85.0%), F1-score (84.0%), IoU (80.0%), and Dice coefficient (88.0%). Pre-processing techniques improve defect visibility and reduce noise, while the efficient design of PixelSegNet-CNN allows for accurate localisation of defects such as holes, stains, and weave irregularities. The framework exhibits robustness across various fabric types and defect categories, providing a scalable solution for industrial automated inspection systems.