Year of Graduation
Sparsification of Neural Networks for Feature Space Dimensionality Reduction in Image Retrieval Problem
Applied Mathematics and Information Science
Deep convolutional neural networks can be successfully used to build image descriptors in the image retrieval problem. However, the proposed methods can hardly be used in various applications because of high dimensionality of final image descriptors that results in expensive computations. At the same time there are several approaches aimed at inducing structured sparsity to neural networks that results in more compact architecture without significant loss in performance. In this paper we explore properties of structured sparsity methods in image retrieval. We compare performance of proposed methods with existing ones. Our approach may reduce the required amount of computational resources thus making advanced approaches more applicable for real-life tasks.