3D Convolutional Neural Networks Initialized from Pretrained 2D Convolutional Neural Networks for Classification of...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning excels in 2D image processing using vision sensors.
• Emerging 3D sensors offer new perception possibilities.
• Current 3D convolutional neural networks (CNNs) lag behind 2D versions and other 3D deep learning techniques.

Purpose of the Study:

• To enhance 3D deep learning performance by transferring pretrained weights from 2D networks to their 3D counterparts.
• To evaluate different 3D CNN architectures (VGG16, ResNet, Inception ResNet, EfficientNet) in an industrial object recognition setting.

Main Methods:

• Weight transfer from established 2D networks to 3D CNNs.
• Analysis of various 3D CNN architectures including VGG16, ResNet, Inception ResNet, and EfficientNet.
• Performance comparison based on recognition accuracy in an industrial context.

Main Results:

• The EfficientNetB0 architecture, utilizing extrusion, achieved the highest accuracy of 0.9217, comparable to state-of-the-art methods.
• The weight transfer approach boosted the Inception ResNet 3D version's accuracy by up to 18% compared to the baseline 3D method.
• Different 3D CNN combinations were analyzed for recognition accuracy.

Conclusions:

• Transferring pretrained 2D weights is an effective strategy to improve 3D deep learning accuracy.
• EfficientNetB0 demonstrates strong performance in 3D object recognition tasks via weight transfer.
• This method offers a viable path to bridge the performance gap between 2D and 3D deep learning models.