Rakibul Islam
Contact
ACM Digital Library · 2024 First Author Peer-Reviewed Proceeding Published: 2024

Ensemble Deep Learning for Real-Time Textile Fabric Defect Detection and Classification Using YOLOv8 and CNNs

Rakibul Islam1, Co-Authors1
1Department of Computer Science & Engineering, United International University (UIU), Dhaka, Bangladesh
Cite Paper

Abstract

Automated quality assurance in high-speed textile manufacturing demands both millisecond-level defect localization and fine-grained multi-class classification. Traditional manual inspection suffers from human visual fatigue and misses up to 30% of subtle structural weave anomalies, while single-backbone object detectors struggle to distinguish visually similar defect topologies such as broken picks, warp float, stain, and hole formations under variable loom illumination.

In this paper, we propose a two-stage Hybrid YOLOv8 and Ensemble CNN Framework for real-time fabric defect inspection. First, an anchor-free YOLOv8 detector rapidly localizes candidate defect regions of interest (ROIs) on streaming loom frames. Second, cropped defect patches are routed through a weighted soft-voting ensemble of specialized Convolutional Neural Networks (ResNet-50, DenseNet-121, and MobileNetV3) to refine fine-grained defect categorization. Experimental evaluation demonstrates a 96.4% classification F1-score and 94.1% mAP@50 at 48 FPS, outperforming standalone detection baselines while remaining deployable on industrial edge GPUs.

System Architecture Two-Stage Detection & Ensemble Verification
Loom Frame CLAHE + Resize YOLOv8 ROI Localization Bounding Boxes ResNet-50 Residual Features DenseNet-121 Dense Feature Reuse MobileNetV3 Fast Depthwise Conv Soft Voting Defect Class & Confidence
Figure 1: Proposed two-stage inspection pipeline. YOLOv8 isolates candidate defect bounding boxes in real time, and a 3-branch CNN ensemble classifies the cropped ROI via probability-weighted soft voting.

Key Scientific Contributions

  • Decoupled Localization & Ensemble Verification: Mitigates single-detector classification confusion on micro-textures by pairing anchor-free YOLOv8 ROI proposals with a diverse 3-backbone CNN ensemble.
  • Illumination-Robust Preprocessing Pipeline: Incorporates Contrast Limited Adaptive Histogram Equalization (CLAHE) and mosaic augmentation to handle uneven factory floor lighting and class imbalance.
  • Real-Time Industrial Throughput: Achieves 48 FPS end-to-end inference speed with a 4.2% F1-score gain over standalone YOLOv8 classification heads.

Quantitative Benchmark Results

Test Split Evaluation
Architecture mAP@50 Precision Recall F1-Score Speed (FPS)
Faster R-CNN (ResNet-50) 87.4% 86.9% 85.2% 86.0% 19 FPS
Standalone YOLOv8s 91.8% 92.5% 91.9% 92.2% 64 FPS
YOLOv8 + Single DenseNet-121 93.2% 94.1% 93.6% 93.8% 53 FPS
Proposed (YOLOv8 + Ensemble CNNs) 94.1% 96.8% 96.0% 96.4% 48 FPS

Related Publications by Rakibul Islam

View All 9 Papers
IEEE ICEEICT 2024 First Author

Deep Multi-Modal Approach for Protein Function Prediction and Classification

ICTIS 2025 Co-Author

TIG-UNet++: A Thresholded Input Guided Modified UNet++ for Separating Lung Region in Chest X-Rays

Manuscript in Prep · 2026 First Author

Attention-Based Graph Similarity Learning Using Graph Convolutional Networks