The problem of extraction of the image objects features by means of using the inverse filters (IF) is considered. The IF are formed by the inversion of the matrix composed of correlation vectors of a set of objects templates examples. The inversion is made with the help of singular value decomposition. Three approaches to regularization and its impact on IF recognition properties are also considered. There was defined the functional that specifies minimal mutual relations between functions of the filters to obtain optimal separation of the features. A training process is used in order to obtain filters with high recognition performance.
Recognition of textured objects is a typical problem in computer vision and pattern recognition. Usually are meeting two variations of this problem. The first is the recognition of a signed texture. The second is objects of interest recognition in image with textured background. We offer an approach for solving of the both problems with using of the inverse resonant filtration. A textured object or background dynamic space can be approximated by a set of principal eigenvectors in the form of resonantharmonic functions called as eigen harmonics (EH). The Inverse ResonantFilter (IRF) which is resonant in respect to EH series eliminates textured background and forms indicator signal of certain species. The IRF is founded on the approximation and extrapolation of the texture template signal by series of EH. Two methods of 2-dimensional (2D) EH parameter estimation are considered. The methods give estimates robust tobreaks and noise peakspresentedin textured image signal. The first method is based on the linear symmetry model and can be presented as a double linear prediction model. Additionally the condition of the unitary symmetry is used to provide stationarity and periodicity of the model. The second method is based on 2D correlation matrix splitting with projection into the subspace of principal harmonic components. Implementation of the IRF is considered in spatial and spectral domains. Discrete Fourier transform with eigen kernel (DFTEK) was used for design and realization of the high order IRF. The DFTEK has fractionally fast transform algorithm. Aligning image fragment phases improved inverse resonant filtration in the spectral domain. It was shown that the optimal approach to image filtration consists of an initial fast filtration in the spectral domain, followed by postfiltration of the image zones containing anomalous background variations using IRF in the spatial domain. It was shown that the IRF is invariant to shift transform. It can be invariant to affine and scale transform if instead initial texture and image are using their invariant pattern in the form of image surface geometry characteristics.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.