Comparison of Convolutional Neural Networks with Feature Extraction Methods for Depth Map Classification
Peter Sýkora, Patrik Kamencay, Róbert Hudec, Miroslav Benčo, Martin Šinko · Advances in Military Technology · 2020
In this paper a comparison between feature extraction methods (Radon Cosine Method, Canny Contour Method, Fourier Transform, SIFT descriptor and Hough Lines Method) and Convolutional Neural Networks (proposed CNN and pre-trained AlexNet) is presented. For evaluation of these methods depth maps were used. The tested data were obtained by Microsoft Kinect camera (IR depth sensor). The feature vectors were classified by the Support Vector Machine (SVM). The confusion matrix for evaluation of experimental results was used. The row of confusion matrix represents target class of tested data and the column represents predicted class. From the experimental results is evident that, the best results were achieved by proposed CNN (97.4%). The feature extraction methods reached up to 91.9% (Radon Cosine Method). The pre-trained AlexNet scored 93.7%.