Research On Texture Image Automatic Classification Algorithms

Texture classification is an important research topic in the field of computer vision and pattern recognition.Its key point is how to extract a suitable texture character vector and construct a fast and stable classifier.This thesis focuses on the later.The traditional classifier,such as K-means clustering,support vector machines and artificial neural networks,always have some disadvantages like low accuracy,high computational efficiency and the lack of theoretical support.As a new type of single-hidden-layer feed-forward neural networks(SLFNs),Extreme learning machine(ELM)randomly chooses the input weights,analytically determines the output weights of SLFNs.Different from the traditional classifiers,ELM tends to provide the best generalization performance at extremely fast learning speed.Based on several mature feature extraction algorithms,this thesis analyzes and studies several texture classifiers in detail.The main contents and contributions are as follows:(1)The advantages of extreme learning machine in the application of texture classification are studied.This method tends to provide the best generalization performance at extremely fast learning speed.Based on several texture feature extraction algorithms,such as wavelet packet decomposition,gray level co-occurrence matrix,gray gradient co-occurrence matrix,statistical geometrical features,Gabor wavelet and dual tree complex wavelet,ELM is trained as a classifier to realize the automatic classification of texture images.(2)An improved texture classification algorithm based on genetic algorithm is proposed.As a kind of random search algorithm,ant colony algorithm has the advantages of global optimization,parallel computing and strong robustness.By combining the ant colony algorithm with ELM,the random parameters in the network can be taken as the ants,thus the ant foraging process is a parameter optimization process.According to the pheromone on each path,the ants can obtain the next path.Keep updating the pheromone,the shortest path can be obtained,as well as the optimal parameters.The simulation results show that the method can achieve higher classification results.(3)Due to the unstable output of traditional texture classification methods based on ELM,a new texture classification approach is presented.In order to improve the generalization ability and the robustness of learning model,this thesis improves the traditional dynamical model,and then realizes the optimal fusion of multiple ELMs with the iteration of linear and local attractor.Experimental results demonstrate that the proposed approach can significantly improve the stability,as well as the classification accuracy,and achieve ideal texture classification effect.