Research On Remote Sensing Image Processing Applications Based On Complex Wavelet Transform

Image processing of remote sensing images (including visible images and Syn-thetic Aperture Radar (SAR) images) has been playing an increasingly important role in emerging applications recently, such as environmental monitoring, ship target recog-nition, flood extent mapping, coastline extraction, etc. Edge detection is a fundamental issue of image processing and computer vision. Edge information is used to acquire the contour of target or region, and includes the important features of target or region. So edge detection is the base work of target detection and image segmentation. There is usually serious speckle noise in SAR images, while serious gaussian white noise in visible images. The noise make the automatic interpretation of remote sensing images difficult. Therefore image denoising becomes an essential step for the application of remote sensing images. Water/Land segmentation is an efficient tool in applications such as ship target recognition, flood extent mapping and oil pollution monitoring. It also attracted a lot of attention. Therefore, how to design high-performance and high-efficiency algorithms of these research fields has attracted a lot of attention.Because of the difference between remote sensing images and natural scene im-ages, existing algorithms of natural scene images cannot be simply applied for remote sensing images. At the same time, due to the instrument or environment factors, many interferences will exist on the remote sensing images, resulting in the image processing more diffcult. Currently the algorithms for visible images have been widely researched, but the algorithms for SAR images acquires less attention, especially SAR images in complicated environment. In this thesis we discuss the edge detection algorithm of SAR images, image denoising algorithm of visible/SAR images, water/land segmentation al-gorithm of SAR/visible images. The main research work is concluded as follows:1. Due to the multi-scale nature of SAR images and multi-orientation nature of edges of objects, we design a new SAR image edge detection algorithm by combining the oriented gradient of histograms method and the dual tree complex wavelet transform (DT-CWT). We calculate the oriented gradient matrix of histograms for each wavelet subband in the complex wavelet domain, where the parameter is determined by the orientation of the complex wavelet subband. We acquire the global oriented gradient matrix of histograms with the orientation selectivity and multi-scale nature of complex wavelet transform and extract multi-orientation edges. Non-maximum suppression method is employed to extract salients, the the gradient direction parameter which needs is also determined with the orientation of the complex wavelet subband. The experiments on real SAR images with dif-ferent features show that our algorithm can extract significant edges effectively, and is robust to speckle noise, intensity inhomogeneity, radar backscattering and complex boundaries.2. Wavelet threshold shrinkage and nonlinear diffusion equation have been wide-ly used in the field of image denoising. They are effectively combined recently. This thesis proves that2-D anisotropic diffusion equation is equal to the filter form of discrete wavelet threshold shrinkage. Then the equivalence is further extend-ed to describe the relationship between DT-CWT and2-D anisotropic complex diffusion equation. Based on this equivalence we design a new SAR/visible re-mote sensing image denoising algorithm by combining DT-CWT and anisotropic complex diffusion equation. We re-write the wavelet subband coefficients into complex form and then apply the anisotropic complex diffusion function to ac-quire the denoising result. We also investigate the influence of different complex diffusion functions.3. We improve the image denoising algorithm above, replacing the DT-CWT with the double-density dual-tree complex wavelet transform (DDDT-CWT) because of its properties of specified vanishing moments and of short support and of bet-ter denoising capability. Based on the interscale continuum of complex wavelet subband coefficients, we calculate the coefficient diffusion weight by using the current coefficient and its parent coefficient at the same time as the bi-variate shrinkage function. Two different noise threshold estimation methods are applied corresponding to visible/SAR images.We also design a new threshold estimation method for speckle noise in real SAR images. Simulation and real experimental results confirm the effectiveness of proposed algorithm.4. Pixel intensity difference is one of the main differences between water and land region. However, due to the reality of SAR remote sensing images with intensi-ty inhomogeneity and noise phenomena, the performance of existing water/land segmentation algorithm built on pixel intensity difference is limited. To solve this problem, we propose an improved Chan-Vese model for water/land segmentation and a new segmentation algorithm based on this model. The model considers the fact that land region is characterized by many edges while almost no edges are found over water. At first, edge information is extracted in complex wavelet transform domain by modifying the edge detection algorithm above. After that, together with constraints of edge information and pixel intensity information, the new model splits water and land region for SAR images. The new segmenta-tion algorithm overcomes the shortcoming that the traditional Chan-Vese model is sensitive to the initial contour and improves the result accuracy. This algorithm is also applied in visible images and acquires satisfactory results.5. In practice the target area may be monitored for long-term, it will lead to multi-temporal remote sensing images. From an information theory perspective it in-dicates more redundant information, which is helpful for more accurate result of segmentation water and land region. To solve this problem, the proposed model above is extended to the vector-valued image. We treat multi-temporal remote sensing images as a vector-valued image, each one as a channel of the vector-valued image. Then a new water/land segmentation algorithm for multi-temporal remote sensing images is designed based on the extended model.