Convolution-Based Encoding of Depth Images for Transfer Learning in RGB-D Scene Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Indoor environment classification is challenging.
• RGB-D data integration for scene understanding is an active research area.
• Existing methods like HHA encoding are computationally expensive.

Purpose of the Study:

• To propose a computationally efficient encoding method for RGB-D data.
• To improve multimodal transfer learning for scene classification.
• To address class imbalance in image datasets.

Main Methods:

• Developed a novel, computationally efficient encoding technique for depth images.
• Integrated the encoding method with a VGG16 network for multimodal transfer learning.
• Applied feature-level synthetic minority oversampling technique (SMOTE) to handle class imbalance.

Main Results:

• The proposed encoding method performs comparably or better than existing approaches.
• Achieved scene classification accuracy on par with state-of-the-art architectures.
• Demonstrated the effectiveness of the efficient encoding for real-time applications.

Conclusions:

• The novel encoding method offers an efficient alternative for RGB-D scene classification.
• This approach enhances multimodal transfer learning performance.
• The method contributes to more effective and efficient indoor environment understanding.