An efficient image categorization approach using deep belief network

Image categorization is the process of categorizing the images into its respective class or bins. It is still challenging problem in computer vision key area. The existing methodologies for image categorization like semantic modelling approaches, neural network approaches does not provides an accurate solution. This is due to inefficient feature extraction and their processing. Deep Learning is a new and emerging subfield in machine learning research. It consist of a set of algorithms that attempt to learn features using multiple layers where each layer represents higher level of abstraction of lower layers. The acceptance of deep learning concept is due to unsupervised and scalable nature of it. Manifold regularized deep learning architecture extracts the structural information of the input image which leads to better classification. Primarily low level features of an input image is obtained by SURF algorithm. This is inputted into DBN architecture to formulate high level features. The SVM classifier and RLSC classifier used separately to classify the input image set into its category. It is observed that SVM provides better accuracy than RLSC classifier.