A review on modern defect detection models using DCNNs - Deep convolutional neural networks.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning models have surpassed traditional computer vision algorithms in accuracy and processing time.
• The success of deep learning in benchmarks like ImageNet has spurred research and applications, including visual defect detection.
• Advancements in hardware (GPUs, CPUs) and software frameworks have accelerated deep learning model performance.

Purpose of the Study:

• To provide a structured overview of popular object detection models (CNNs, YOLO, SSD, cascaded architectures) for defect detection.
• To analyze the advantages and disadvantages of these models in the context of defect detection.
• To summarize techniques for model compression and acceleration, enabling portable deep learning solutions.

Main Methods:

• Review of existing object detection architectures: Region-based CNNs, YOLO, SSD, and cascaded models.
• Analysis of model compression and acceleration techniques.
• Experimental repurposing of YOLOv4 for industrial cable defect detection.

Main Results:

• Popular object detection models are readily adaptable for defect detection tasks.
• The YOLOv4 model was successfully trained and repurposed for industrial cable detection within hours.
• Computational requirements can be met by general-purpose or high-performance desktop setups.

Conclusions:

• Deep learning-based defect detection can significantly enhance manufacturing productivity.
• The adaptability and performance of models like YOLOv4 lower the barrier for businesses to implement advanced visual inspection.
• Accessible computing resources facilitate the adoption of deep learning for defect detection across industries.