RNN-LSTM based indoor scene classification with HoG features

Abstract

The machine learning and artificial intelligence models have evolved with unenvisaged swiftness over the past decade. The machine learning and artificial intelligence models work on the basis of the mathematical models, which are capable of processing the financial, image, video, audio, audio-visual and several other forms of data. In this paper, the work has been carried upon the bulk image data representing the variety of indoor scenes. The goal is to detect the type of indoor scene, which can be utilized by variety of artificial intelligent application for various purposes. The indoor scene recognition has been performed over the computer vision & pattern recognition (CVPR), 2009 dataset, which is consisted of 15620 images. The deep learning mechanism known as recurrent neural network (RNN) has been incorporated for the classification of the indoor scene data over the histogram of oriented gradient (HoG) based features. Specifically, the forget gates based recurrent mode called Long short-term memory (LSTM) has been incorporated for the classification of the indoor scene data. The performance of the proposed model has been analyzed over the CVPR09 dataset using 50 and 100 randomly drawn test cases. The proposed model has been found 92% accurate in comparison with SVM (88%), KNN (80%) and Naïve Bayes (86%). The F1-measure based performance assessment also proves the robustness of LSTM based model with 96% accuracy over 92% of Naïve Bayes, 93% of SVM and 88% of KNN.