A Lightweight Approach to Comprehensive Fabric Anomaly Detection Modeling.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Materials Science

Background:

• Fabric anomaly detection is crucial for quality control.
• Existing methods often require high computational resources, limiting real-time application.
• Lightweight and efficient models are needed for deployment on edge devices.

Purpose of the Study:

• To propose a lightweight network, GH-YOLOx, for efficient fabric anomaly detection.
• To reduce computational resource consumption while maintaining high detection accuracy.
• To enable real-time fabric anomaly detection on mobile and embedded devices.

Main Methods:

• Integration of ghost convolutions and a hierarchical GHNetV2 backbone for capturing multi-scale features.
• Implementation of GhostConv, dynamic convolutions, feature fusion modules, and shared group convolution head.
• Application of lamp pruning for inference acceleration and channel-wise knowledge distillation for accuracy enhancement.

Main Results:

• GH-YOLOx significantly reduces the number of parameters compared to existing lightweight models.
• The proposed network achieves a higher detection rate in fabric anomaly detection tasks.
• Experimental results validate the effectiveness and efficiency of the GH-YOLOx model.

Conclusions:

• GH-YOLOx provides a practical and efficient solution for real-time fabric anomaly detection.
• The lightweight design makes it suitable for deployment on resource-constrained mobile and embedded systems.
• This approach addresses the challenge of high computational costs in fabric inspection.