[00391] Recent Advances in Multiscale Transforms for Image Analysis
Session Time & Room : 4D (Aug.24, 15:30-17:10) @A502
Type : Proposal of Minisymposium
Abstract : This minisymposium will bring together researchers working on multiscale image transforms beyond wavelets and discuss deeper connections between harmonic analysis and image analysis. We plan to discuss various methods to decompose an image into "predictable" local segments and their residuals that allow efficient and sparse image approximation and associated tools based on new types of directional wavelets and monogenic signal representations. The key idea here is how to predict main features, e.g., dominant orientation information in texture images, in each local segment in such a way that the unpredictable portion in that segment is easily compressible or remains as noise.
[01735] Multiscale monogenic image representations using Poisson kernels
Format : Talk at Waseda University
Author(s) :
Brian Knight (University of California, Davis)
Naoki Saito (University of California, Davis)
Abstract : By viewing a 1D signal as the boundary value of a harmonic function in the unit disc in $\mathbb{C}$, one can obtain a multiscale analytic signal representation by supplementing its conjugate counterpart. This is done by the Poisson/Cauchy integral formula. We generalize this for a 2D image by sandwiching it by quaternionic Poisson/Cauchy kernels. This leads to a natural multiscale monogenic image representation. We also plan to discuss its application in oriented texture image analysis.
[02090] Image Interpolation Technique by the PCA of the Gradient Distribution
Format : Talk at Waseda University
Author(s) :
Masaki Morita (Meijo University)
Yuto Kimura (Meijo University)
Katsu Yamatani (Meijo University)
Masayoshi Nakagawa (Meijo University)
Abstract : We propose a method to reconstruct local image patches with gradient data and boundary information. In this talk, we present an image interpolation technique based on the principal component analysis of the gradient distribution of image intensities. Our numerical experiments show superiority of our proposed method over previous method based on the interpolation using harmonic functions.
[02079] Improvement of coding procedures for Haar transform-based lossy image compression
Format : Talk at Waseda University
Author(s) :
Keita Ashizawa (Shizuoka Institute of Science and Technology)
Katsu Yamatani (Meijo University)
Abstract : We discuss an improved version of our Multi−neighbor Predictors and Residual Orthogonal Transformations, MPROT, which was a Haar transform-based lossy image compression method without generating mosquito noise. Due to the slow decay of the Haar coefficients, however, the PSNR values of certain test images compressed by the MPROT were not entirely satisfactory. Our new coding scheme takes advantage of redundancy of the MPROT coefficients and improves high-resolution image compression quality.
[02086] Edge enhancement with directional wavelet transform
Format : Talk at Waseda University
Author(s) :
Kensuke Fujinoki (Kanagawa University)
Keita Ashizawa (National Institute of Technology, Maizuru College)
Abstract : We introduce a two-dimensional directional discrete wavelet transform that can decompose an image into twelve multiscale nearly isotropic directional edge components. The transform is designed in fully discrete setting and therefore is easy to implement in the spatial domain. Experimental results for image edge detection and enhancement show that both global and local edge structures of images are successfully represented.
