An Efficient Printing Defect Detection Based on YOLOv5-DCN-LSK.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Manufacturing Quality Control

Background:

• Inkjet printing labels are susceptible to defects impacting product information readability.
• Existing deep learning defect detection systems struggle with the subtle and varied shapes of these printing defects, limiting accuracy and speed.

Purpose of the Study:

• To develop an improved deep learning model for high-accuracy and rapid detection of printing defects in inkjet labels.
• To enhance the detection capabilities for narrow, elongated, and small printing defects.

Main Methods:

• An improved YOLOv5 network architecture was proposed, incorporating C3-DCN modules for enhanced detection of elongated defects.
• Large Selective Kernel (LSK) and RepConv modules were integrated into the feature fusion network.
• A combined loss function using Normalized Gaussian Wasserstein Distance (NWD) and Efficient IoU (EIoU) was employed to improve small target detection.
• Model pruning techniques were applied to reduce model size and increase detection speed.

Main Results:

• The improved YOLOv5 model achieved a mean Average Precision (mAP@0.5) of 0.741.
• The enhanced model demonstrated a detection speed of 323.2 Frames Per Second (FPS).
• Performance metrics showed a 2.7% increase in mAP and a 20.8% increase in FPS compared to the original YOLOv5.

Conclusions:

• The proposed enhanced YOLOv5 model effectively addresses the challenges of detecting subtle and varied printing defects.
• The method meets the stringent requirements for high precision and high efficiency in industrial printing defect detection.
• This advancement contributes to improved quality control in inkjet label production.