Research On Texture Image Classification Based On Deep Learning

Texture is an important visual clue in images which is found everywhere in nature.Texture images are influenced by various factors such as illumination,noise,and scale,which makes it difficult to extract discriminative texture features and achieve accurate classification.This challenging problem has become a hot research topic in the computer vision community in recent years.With the development of deep learning,the powerful end-to-end modeling capacity of deep neural networks can achieve better performance than classical methods for many tasks.From the point of view of defect detection,this article first introduces the research status of texture image classification;then introducing the texture feature representation in classical methods and some "shallow" classifiers model;further introduces the basic knowledge of deep learning and carry out texture image spatial correlation modeling and classification based on deep convolutional neural network;texture image classification based on multi-directional spatial correlation feature learning,and texture image classification methods based on global and local correlation.The main contents are as follows:(1)Put forward the method of modeling spatial correlations and classification by using deep convolutional neural networks for texture images.First of all,the convolutional layers in base network can learn base feature representations for different kinds of edges,contours and objects,etc.,and then the LSTM units can model the spatial correlations in both horizontal and vertical directions by leveraging the long-term memory mechanism of the Long and Short Time Memory Network(LSTM).Finally,using the learned spatial correlation features to construct a classifier for classification,and the performance of the method is verified on two standard texture datasets.(2)Put forward the method of texture image classification based on multi-directional and spatial correlation texture feature learning.Firstly,the multi-directional end-to-end network architecture for texture image classification is constructed.Then,the fully connected feature embedding layer is constructed to transform the learned features from LSTM units further.Finally,a fully connected linear classifier is constructed to make category prediction based on the above feature embeddings,and the performance of the method is verified on two standard texture datasets.(3)Put forward the method of texture image classification based on global and local correlation.Input to local modules and global modules,which is based on the deep convolutional features from the base network.Then,extract texture features are Concat by the above two modules.Finally,the fully connected feature embedding layer is constructed to transform the learned features from local modules and global modules further;Finally,the final classification is based on the above embedded coding features,and the performance of the method is verified on two standard texture datasets.