Content Based Image Retrieval using Convolutional Neural Network and Improved Sparse Representation

Abstract

Content based image retrieval is a system that takes an image as an input and provides a set of similar images to the input as output in an order of matched similarity. Features matching among images is a vague topic. It depends what features are taken in consideration and to what level features are being matched. Images having feature similarities with other images might differ in terms of semantic analysis of those images. This thesis presents an image retrieval scheme, using convolutional neural network and improved sparse representation. Features are obtained in the form of feature matrix from convolutional neural network. These features are then processed applying improved sparse representation. Some works have been done in sparse representation of images. Sparse representation aims to represent the query sample in terms of weighted sum of training samples. Improved sparse representation uses the virtual training samples generated from original training samples, extracts the features and the model is trained using those features in batch. Sparse code for the training images are obtained first. In similar fashion, query image is passed through a convolutional neural network to extract the feature. Using the features extracted and the dictionary model constructed, sparse code for the query image is calculated. These sparse code then compared to find the similarity index between query image and the training images. This methodology is different from the earlier methodology where similar images are predicted only from the sparse code matrix of the query image. The performance of the content based image retrieval scheme is targeted to increase with the application of convolutional neural network and improved sparse representation in different image repositories.