Scene Classification: A Comprehensive Study Combining Local and Global Descriptors

Abstract

In this paper, local region characteristics and overall structure of scene images are used for scene classification by combining different local and global descriptors. For this purpose, GIST, Histogram of Oriented Gradients (HOG), dense Scale-Invariant Feature Transform (SIFT), dense Speed-Up Robust Features (SURF), Daisy and Local Binary Patterns (LBP) features are classified individually and jointly with Support Vector Machine (SVM) by using different sizes of training sets. Evaluation tests were conducted on Places15, MIT indoor, SUN397 and Places365 datasets. Most used machine learning algorithms in scene classification literature -SVM with RBF and linear kernels, K-Nearest Neighbors and Random Forest- were evaluated on Places15 dataset for comparison. Besides accuracy, recall and precision, processing time for testing with SVM was measured individually and jointly for a deeper evaluation of the features.