Combine EfficientNet and CNN for 3D model classification.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Analysis

Background:

• The proliferation of 3D models necessitates effective classification and retrieval methods.
• Existing 3D model classification techniques often overlook crucial contour information, leading to suboptimal accuracy.
• Complexity and irregularity of 3D models present significant challenges for accurate classification.

Purpose of the Study:

• To enhance the accuracy of 3D model classification.
• To develop a novel method integrating both view and shape features for improved discriminative power.
• To address the limitations of methods focusing solely on local 2D features.

Main Methods:

• A hybrid approach combining EfficientNet for view feature extraction and Convolutional Neural Networks (CNNs) for shape feature extraction.
• Projection of 3D models into 2D views from multiple angles.
• Utilizing shape descriptors such as D1, D2, D3, Zernike moments, and Fourier descriptors.
• Combining extracted view and shape features for classification using a softmax function.

Main Results:

• The proposed method demonstrated superior performance compared to existing approaches on the ModelNet 10 dataset.
• Integration of view and shape features significantly improved classification accuracy.
• The method effectively captures both global and local characteristics of 3D models.

Conclusions:

• The proposed EfficientNet and CNN-based method offers a robust solution for accurate 3D model classification.
• Combining multi-view features with detailed shape descriptors enhances the discriminative capability for complex 3D objects.
• This approach holds promise for applications in mechanical design, education, and medicine.