An Improved Convolutional Neural Network for Classification of Small Patches of Granite Tiles

摘要

In some of the industries, textural classification is one of the most important and challenging problem. Among all, the stone industries has to deal with such issues more often. Many a times the buyer receives different rocks due to the fact that the visual appearance of some of the rocks is so similar that it creates confusion. This paper proposes a resolution invariant Convolutional Neural Network (CNN) architecture by improving different aspects of the network, including designing of layer, loss functions, activation function, regularization and optimization to extract the intrinsic features from the small granite image patches. These extracted features will help in patches classification and will make the proposed network capable of proving itselfin uncontrolled environmental conditions and will also resolveanykind of misunderstanding. The network is trained from the scratch and has outperformed the existing first data driven technique with a well-known data set.