Single-plane scene classification using deep convolution features

Abstract

Scene classification is considered as one of the challenging tasks of computer vision. Due to the availability of powerful graphics processing unit and millions of images, deep learning techniques such as convolutional neural networks (CNNs) have become popular in the image classification. This paper proposes the use of a pre-trained CNN model known as Places CNN, which was trained on scene-centric images. In this work, the pre-trained CNN is used as a feature extractor. The features are then used as input data for support vector machines (SVMs). The effect of grayscale images on the performance of pre-trained CNN-based scene classification system is analyzed by means of classification accuracy and equal error rate. The dataset used for this purpose is Oliva Torralba (OT) scene dataset, which consists of eight classes. The classification experiments are conducted using the feature vector from the ‘fc7’ layer of the CNN model for RGB, RGB2Gray, and SVD decolorized images. The classification experiment was also done using a dimensionality reduction technique known as principal component analysis (PCA) on the feature vector. The results obtained from classification experiments show that RGB2Gray and SVD decolorized images were able to give results similar to that of RGB images. The grayscale images were able to retain the required shape and texture information from the original RGB images and were also sufficient to categorize the classes of scene images.