Convolutional Neural Networks and Texture Classification

Abstract

Convolutional neural networks (CNN) model is an instrumental computational model not only in computer vision but also in many image and video applications. Similar toCognitron and Neocognitron, CNN can automatically learn the features of data with the multiple layers of neurons in the network. There are several different versions of the CNN which have been reported in the literature. If an original image texture is fed into the CNN, it will be called an image-based CNN. A major problem with the image-based CNNs is that the number of training images is very demanding for the good generalization of the network due to the rotation and scaling change in images. An alternative method is to divide an image into many small patches for the CNN training. This is very similar to the patches used in the K-views model. In this chapter, we will briefly explain the image-based CNN and patch-based CNN for image texture classification. The LeNet-5neural network architecture will be used as a basic CNN model. CNN is useful not only in the image recognition but also in the textural feature representation. Texture features, which are automatically learned and extracted from a massive amount of images using the CNN, become the focus of developing feature extraction methods.