Volumetric Feature Extraction from 2D Images Using Cubixels
Sanad Aburass, Maha Abu Rumman · 2024
This paper introduces innovative volumetric feature extraction methods for 2D images using Cubixels and the Volume of the Void (VoV) metric. Cubixels transform traditional pixels into three-dimensional entities, capturing more detailed spatial and color relationships by assigning RGB values to the dimensions of width, length, and height. The VoV metric quantifies spatial gaps between adjacent Cubixels, providing valuable insights into textural characteristics and edge acuity. We developed two methods: the first focuses on calculating VoV between Cubixels, while the second integrates gradient and curvature features to enhance geometric feature extraction. These methods were applied to the CIFAR-10 and MNIST datasets and evaluated using four convolutional neural network (CNN) models incorporating residual connections, spatial dropout, and attention mechanisms. Our experimental results demonstrate significant improvements in model performance, including higher validation accuracy and reduced overfitting, as indicated by the Overfitting Index. This study underscores the potential of volumetric feature extraction in advancing image processing and computer vision, paving the way for more sophisticated and context-aware AI applications.