Image Texture Analysis And Classification Study

Texture is a very important research work in the fields of pattern recognition and computer vision, and has an extensive application background in science research and engineering technology. Extracting the texture features of texture image is the foundation of texture analysis. In the past decades, many works have been done to extract the texture feature from texture images. A lot of feature descriptors have been suggested. However, most these texture feature descriptors are limited in simulations, they have hardly been used in the real world applications. For instance, the gray statistical feature, filter bank feature and the geometrical structure feature make assumption that the texture images are perpendicular to the camera. That is to say these methods fails to consider the changes of images brought by camera pose, such as, image scaling, nonrigid deformation and affine transformation etc. In classification of texture images, although we can accquire the training images by sampling different camera poses, we still run the risk of that the training samples are deficient. Actually, it is impossible to collect the training samples of all camera poses, so the texture features obtained by the traditional methods have limit value of applications in practice. Therefore, to find the proper texture descriptor is still a problem needed to be further investigated.In this paper, we develop the robust texture descriptor of texture image. The proposed texture feature descriptor is based on the DoG keypoints, which defines the local invariant image patches. Since the local image patches are stable when geometric deformations happen, so the proposed texture feature descriptor is quite promising. Particularly, in this paper, we propose the eDoG detector to model the stable spacial co-occurrence pairs of keypoints. As a result, we find the texture textons covering the co-occurrence pairs, which is stable because it reflects a closed local structure in images. In addition, a new MG-SIFT descriptor is proposed to describe the keypoint. We evaluate the proposed texture descriptor extracted by the DoG and eDoG detectors using two popular texture databases, the competitive results are achieved. The experimental results demonstrate that the proposed texture descriptor is suitable for both the geometric-transformed database and the common front-viewed database. So we think this texture descriptor can be used in the real world applications.Pattern classification is the stage after the feature extraction. Even using the same pattern feature, the different algorithms of pattern recognition may lead to different classification results. For instance, the proper classification algorithm can give better recognition rate while the unsuitable methods may not work well. Therefore, in this paper, we proposed two kinds of new method for pattern classification. The first one is the improved version of the nearest neighbor algorithm (kNN) and the other is the dynamtic threshold training algorithm for artificial neural networks(ANN).Firstly, we analysis the drawbacks of the traditional kNN algorithm. When the class conditional distributions are overlapping, the noisy samples cause the classifier to overfit to the training set, leading to poor generalization performance. The noises means all training patterns falling on the wrong side of the decision boundary. In classification, they make no contribution to the classification problem but harm the performance. In this paper, we propose a new classification method, which performs classification task based on the local probabilistic centers of each class. This method works through reducing the error-prone samples and restricting their influence regions. Concretely, this method classifies the query sample by using two measures respectively, of which one is the distance between the query and the local categorical probabilistic centers, and the other is the computed posterior probability difference of query and the nearest categorical center. Although both measures are effective, the experiments show the second one achieves the smaller classification error. Meanwhile, the theory analysis of suggested methods are investigated and some experiments are conducted on the basis of both constructed datasets and real world datasets. The investigation results show that this method improves the classification performance of the nearest neighbor algorithm substantially. Specailly, we use this improved kNN algorithm to classify the texture image, and the better results are achieved.Secondly, the neural network is also a popular tool for pattern classification. However, when training set is unbalanced, the conventional least square error(BP) training strategy is less effective to train neural network(NN) for classification because it often lead the NN to overcompensate for the dominant group. Therefore, in this paper a dynamic threshold learning algorithm (DTLA) is proposed as the substitute for the conventional LSE algorithm. This method uses multiple dynamic threshold parameters to gradually remove some training patterns that can be classified correctly by current Radial Basis Function (RBF) network out of the training set during training process, which changes the unbalanced training problem into a balanced training problem and improves the classification rate of the small group. Moreover, we use the dynamical threshold learning algorithm to classify the remote sensing images, when the unbalanced level of classes is high, a good effect is obtained. The experimental results shows: the DTLA method has two advantages: (1) this method can improve the classification rate of small class without sacrificing too much performance of large class. (2)even the dataset is highly unbalanced, this training algorithm can work well. In this sense, it is a stable training algorithm.