A visual defect detection for optics lens based on the YOLOv5 -C3CA-SPPF network model

Area of Science:

• Optics and Photonics
• Computer Vision
• Machine Learning

Background:

• Optical lens defects degrade performance and scattering properties.
• Current machine vision struggles with low-contrast, multi-scale lens defects.

Purpose of the Study:

• To develop an advanced deep learning model for precise optical lens defect detection.
• To improve the accuracy and efficiency of automated optical lens quality control.

Main Methods:

• An improved YOLOv5 network (YOLOv5-C3CA-SPPF) integrating coordinate attention (CA) and CSPNet (C3) modules.
• Incorporation of a SPPF features fusion module into the network's neck.
• Creation of a comprehensive dataset with 3800 images, featuring over 600 samples per defect type.

Main Results:

• The YOLOv5-C3CA-SPPF model achieved a mean average precision (mAP) of 97.1%.
• The model demonstrated a detection speed of 41 frames per second (FPS).
• Outperformed traditional algorithms in accuracy and speed for detecting surface and internal lens defects.

Conclusions:

• The YOLOv5-C3CA-SPPF model offers superior accuracy and speed for optical lens defect identification.
• The model exhibits strong generalizability and robustness, suitable for online quality assurance.
• This advancement supports consistent quality in finished optical products through automated detection.