AI Driven Hybrid Descriptor and Classifier for Robust Textile Fabric Defect Detection
The accurate identification of defects in textile fabrics remains one of the most crucial issues in automated quality checking because identical textures, ambiguous defect patterns, and varying illumination levels in factories pose challenges. Although deep convolutional neural networks (CNNs) have proven to be good in visual recognition processes, it can be very weak in identifying fine-grained direction and micro-texture anomalies that are characteristic of woven fabrics. In order to deal with this shortcoming, this research paper introduces AI-based hybrid descriptor-classifier model that combines both manually created texture features with deep hierarchical feature learning with the goal of identifying textile defects effectively. More precisely, the multi-scale Gabor filters are used to retrieve orientation- and frequency-sensitive responses, whereas Local Binary Pattern (LBP) operators are used to acquire micro-texture variations of a localized nature. These handcrafted feature maps are combined with the middle-level CNN representations via channel-wise concatenation and maximized multi-scale feature aggregation to strengthen the discriminative ability of small and complicated defects. A balanced and augmented industrial fabric dataset is used to train the model in different lighting and texture conditions. Large experiments show that the proposed hybrid architecture has Precision of 95.4%, Recall of 93.8%, F1-score of 94.6%, mAP at 0.5 of 96.3% and IoU of 90.7% better than CNN-only and state-of-the-art object detection baselines. The system has real-time performance of 58 FPS, making it appropriate in the high-speed manufacturing systems. The investigation of ablation justifies that the combination of Gabor and LBP descriptors is much more effective in terms of the localization accuracy and robustness, especially with regard to very delicate and tiny defects. The suggested framework will create a successful collaboration between deep learning and classical descriptions of texture, which will give a high-performance and scalable solution to future generation automated textile inspection systems.