Texture Image Segmentation On Gabor Wavelet Transform And SVM

As an important aspect of digital image process and pattern recognition, texture segmentation has always been one of the hottest and most difficult study topics.Texture segmentation involves accurately partitioning an image into sections according to the textured regions or by recognizing the borders between different textures in the image.Texture feature extraction is crucial in the texture segmentation.In this paper,we made an investigation into texture feature extraction and classification based on Spectrum and its application in field pest image classification,a new method which combined Gabor wavelets,AdaBoost learning algorithm and SVM is proposed.The main work is as follows:(1)Considering that Gabor wavelet transform is similar with human vision characteristic,it was used to extract texture feature of images and then this method was analyzed in detail.(2)Through the weaklearning process,AdaBoost learning algorithm considers the linear classify capability of every feature in each feature vector completely,and then the key features are extracted,so in this thesis it is used after Gabor wavelet filtering to reduce the dimension of Gabor features and improve the performance of segmentation.(3)Support vector machine is already the best classifier in pattern recognition area.In this thesis, the GA based RBF kernel function is used to reduce the amount of work by human.In this thesis,all experiments are implemented based on field pest images.Gabor +BP, Gabor+SVM and Gabor+AdaBoost+SVM are implemented to segmente the field pest images.The experiment results report that Gabor feature has so many dimentions,if we then use SVM directly could ruduce the performance of segmentation.So AdaBoost learning algorithm is used after Gabor wavelet filtering to reduce the dimension of Gabor features.when used the AdaBoost algorithm,the training process take a long time,but in the study process we can used the result directly,and then we can get higher speed of segmentation.The method proposed in this paper gets the better performance and the accuracy is above 90%,and also obtains good experimental performance.Programs were all compiled in the WinXP by Matlab 6.1.