Comparison of Feature Extraction Methods and Deep Learning Framework for Depth Map Recognition

摘要

In this paper a comparison between three feature extraction methods (Fourier Transform, Radon Transform, Canny Edge Filter) and Convolutional Neural Network is presented. These methods are tested on set of depth maps. The Microsoft Kinect camera is used for capturing the images. For the image classification the Support Vector Machine with Radial Basis Function kernel was used. The experimental results from each tested method are stored in confusion matrix. Each row in this matrix represents actual class of tested data and each column represents predicted class. The quality of the Convolutional Neural Networks features has been compared with traditional methods of feature extraction. From the experimental results, we have shown that the Convolutional Neural Network based deep learning framework achieve better classification performance than test feature extraction methods (Fourier Transform, Radon Transform, Canny Edge Filter).