A Comparison between Image Metadata and Visual Features to Classify Social Media Images

Deepank Agarwal, Chirag Bhatia, Mithun Sai, Preety Singh, Rohit Hinduja

Abstract: Image classification is needed to build image search and recommender systems. This paper applies classification techniques on two properties of the social media images: metadata of the images and their pixel intensities. Naive Bayes is applied on the tags and captions extracted from the images while a Convolutional Neural Network is trained for classifying images based on pixel values. Our dataset consists of more than 33,000 Flickr images segregated into 18 classes. Naive Bayes is applied by converting the metadata into a feature vectors. The Convolutional Neural Network is made up of three convolutional layers and three fully connected layers. On comparison, it was observed that analyzing the metadata associated with an image rather than the intensities provided better classification.