Heechan Park, Affine Symmetry and Applications in Image Processings
Natural images have a great deal of self-similarity in a sense that a part of an image can replace another part with a slight deformation. We employ affine symmetry to address the self-relationship, i.e. rotation, scaling, shearing and translation. In this report, various applications are presented exploiting affine symmetry in images.
Firstly, an image coding technique is discussed that uses affine symmetric redundancy between blocks. This poses an interesting approach as opposed to the recent directional Wavelet based trend. Experiments show visually acceptable picture quality at low bitrates.
Secondly, a segmentation algorithm is presented that analyses images in a similar manner to the image coding application. The algorithm uses a directional shape extracted from the local Fourier spectrum. From the obtained shape, a feature is extracted using the affine-invariant Fourier descriptor. Experimental evaluation on structural textures shows encouraging results and application on natural images demonstrates identification of texture objects.
Thirdly, a denoising technique that combines Independent Component Analysis (ICA) and the Multiresolution Fourier Transform (MFT) is presented. This technique inherits the ability to find bases adaptively from given data using ICA and the computational efficiency of the MFT. Another denoising approach is demonstrated that utilises the shape information from the segmentation algorithm as a thresholding mechanism. The experimental results are promising even compared with the recent directional transform such as the Curvelet.
Lastly, future research directions are briefly mentioned.