Research On Methods Of Remote Sensing Image Texture Retrieval Based On Hidden Markov Tree Model And Rotation Invariance

Due to the rapid advanced in remote sensing technology, large quantities of high spatial resolution remote sensing images are becoming available, which has not only brought new favorable opportunities for finer level earth observation, but also raised higher requirements for efficiently organizing, managing, browsing and searching large amount of image data. Researches in this paper are based on the textural information of high spatial resolution remote sensing images and intend to put forward some new thoughts on the method of rotation invariant remote sensing texture image retrieval using the hidden Markov tree model. The main content includes:(1)Summarize the rationale of wavelet transform, framelet transform and steerable pyramid transform as well as their characteristics in frequency domain and probability distribution of the coefficients. Respectively introduce the techniques of multiscale processing of two dimensional images using these three transform and analyze the advantages and disadvantages of each of them.(2)Enhance the capacity of the traditional 2-D wavelet domain hidden Markov tree(HMT) model in capturing the cross-orientation dependency of wavelet coefficients by means of modeling the coefficients with the vector model of HMT. The texture feature extraction of images are performed by estimating the corresponding model parameters using the expectation maximization algorithm. A recursive “upward” algorithm is adopted to fast approximate the Kullback-Leibler distance between trained HMTs, by which the similarity between the texture features can be measured. Then the texture image retrieval based on the wavelet domain HMT model can be conducted.(3)For the fact that the wavelet domain HMT is not rotation-invariant, two improved methods are presented from different perspectives: One is to add a texture orientation estimating step in the original scheme and rotate the query image and candidate images to an identical direction according to their own dominant orientation. The wavelet domain HMT is applied to the rotated images to extract the texture features by which the similarity between the query image and candidate images are measured. Another improved scheme is adopting the steerable pyramid transform to execute the multiscale decomposition on images. The HMT model is built on the obtained multiscale oriented bandpass coefficients of each image, and the rotation invariant texture features can be captured through an adjustment of the model parameters. Then the rotation invariant texture image retrieval can be realized.(4)A prototype system of remote sensing texture image retrieval and the supported graphical user interface are designed according to the above mentioned schemes. The Quin-tree decomposition is adopted for the organization of large remote sensing images. Four different datasets are used to compare the retrieval performance of the improved schemes with the texture image retrieval method based on wavelet domain HMT. The results show that the two schemes presented in this paper can effectively overcome the inhibition of the orientation difference on the texture retrieval accuracy and consequently outperformed the wavelet domain HMT-based scheme in the retrieval of anisotropic remote sensing texture images.