ZFD-Net: Zinc flower defect detection model of galvanized steel surface based on improved YOLOV5

Area of Science:

• Materials Science
• Computer Vision
• Industrial Automation

Background:

• Distinguishing zinc flower defects from galvanized sheet backgrounds is challenging due to complex industrial environments.
• High production line speeds necessitate real-time defect detection methods with improved accuracy and speed.
• Existing detection techniques struggle to meet the demands for real-time, accurate zinc flower defect identification.

Purpose of the Study:

• To develop an advanced deep learning model, ZFD-Net, for efficient and accurate zinc flower defect detection on galvanized sheets.
• To address the limitations of current methods in terms of speed and accuracy for real-time industrial applications.
• To create a comprehensive dataset for zinc flower defect detection to facilitate future research.

Main Methods:

• Proposed ZFD-Net model based on an improved YOLOv5 architecture.
• Integrated a novel Cross Stage Partial Transformer (CSTR) module for enhanced global feature extraction.
• Employed a Bi-directional Feature Pyramid Network (Bi-FPN) for multi-scale defect detail fusion.
• Introduced a Cross ResNet SIMAM FasterNet (CRSFN) module to optimize reasoning speed and detection accuracy.
• Constructed a high-quality, publicly unavailable dataset for zinc flower defect detection.

Main Results:

• ZFD-Net demonstrated superior performance compared to state-of-the-art methods on the self-built dataset.
• The CSTR module improved receptive field and global feature extraction capabilities.
• Bi-FPN effectively fused defect details across different scales.
• The CRSFN module enhanced detection speed while maintaining high accuracy.
• The newly created dataset provides a valuable resource for zinc flower defect research.

Conclusions:

• ZFD-Net offers a significant advancement in real-time zinc flower defect detection for galvanized sheets.
• The proposed model architecture and modules effectively address the challenges of accuracy and speed in industrial settings.
• The development of a dedicated dataset overcomes a critical limitation in the field, enabling further progress.