We describe an application of the clustering filter to image sequence preprocessing for MPEG encoding. Since noise is unavoidable and compression is lossy, we propose and demonstrate that one can obtain better picture quality if one takes out noise before the video signal is fed into an encoder. This is where our contribution differs from previous work on preprocessing. We test our schemes on digital video in the CCIR 601 format converted from film using a standard telecine process. The results show that the motion estimation is more effective and motion-compensated difference signal is easier to encode, resulting in higher quality at a fixed bit rate or lower bit rate at a fixed quality.

We derive and demonstrate a nonlinear scale-space filter and its application in generating a nonlinear multiresolution system. For each datum in a signal, a neighborhood of weighted data is used for clustering. The cluster center becomes the filter output. The filter is governed by a single scale parameter that dictates the spatial extent of nearby data used for clustering. This, together with the local characteristic of the signal, determines the scale parameter in the output space, which dictates the influences of these data on the output. This filter is thus adaptive and data driven. It provides a mechanism for (a) removing impulsive noise, (b) improved smoothing of nonimpulsive noise, and (c) preserving edges. Comparisons with Gaussian scale-space filtering and median filters are made using real images. Using the architecture of the Laplacian pyramid and this nonlinear filter for interpolation, we construct a nonlinear multiresolution system that has two features: (1) edges are well preserved at low resolutions, and (2) difference signals are small and spatially localized. This filter implicitly presents a new mechanism for detecting discontinuities differing from techniques based on local gradients and line processes. This work shows that scale-space filtering, nonlinear filtering, and scale-space clustering are closely related and provides a framework within which further image processing, image coding, and computer vision problems can be investigated.

Demonstrates the use of saddle-node dynamics for edge-preserving and scale-space filtering. The filter is derived from the maximum entropy principle and an analogy with statistical physics. The filter is governed by a single scale parameter which dictates the spatial extent of nearby data used for determining the output. This filter is thus highly adaptive and data-driven. It provides a mechanism for a) removing impulsive noise; b) improved smoothing of nonimpulsive noise and c) preserving edges. Comparisons with conventional techniques are made using real images.>

Image enhancement is useful when the details in an image are lost due to various reasons. It is common to subtract a mask from a given image to enhance the details. The trick is how to obtain a good mask. We describe here how an edge-preserving filter can be used to generate a mask which is smooth over areas with fine details, yet preserving most of the edges. Experiments with real images show that our scheme is very effective.>

Dynamic range compression is useful for displaying and enhancing images. It is common to subtract a mask from a given image to enhance the details. The trick is how to obtain a good mask. The authors describe how an edge-preserving filter can be used to generate a mask that is smooth over areas with fine details, yet preserves most of the edges. Experiments with real images show that the scheme is effective for this task.>

We derive a new clustering algorithm based on information theory and statistical mechanics, which is the only algorithm that incorporates scale. It also introduces a new concept into clustering: cluster independence. The cluster centers correspond to the local minima of a thermodynamic free energy, which are identified as the fixed points of a one-parameter nonlinear map. The algorithm works by melting the system to produce a tree of clusters in the scale space. Melting is also insensitive to variability in cluster densities, cluster sizes, and ellipsoidal shapes and orientations. We tested the algorithm successfully on both simulated data and a Synthetic Aperture Radar image of an agricultural site with 12 attributes for crop identification.

The authors applied a scale-space clustering algorithm to the classification of a multispectral and polarimetric SAR image of an agricultural site. After the initial polarimetric and radiometric calibration and noise cancellation, a 12-dimensional feature vector for each pixel was extracted from the scattering matrix. The clustering algorithm partitioned a set of unlabeled feature vectors from 13 selected sites, each site corresponding to a distinct crop, into 13 clusters without any supervision. The cluster parameters were then used to classify the whole image. The classification map is much less noisy and more accurate than those obtained by hierarchical rules. Starting with every point as a cluster, the algorithm works by melting the system to produce a tree of clusters in the scale space. It can cluster data in any multidimensional space and its insensitive to variability in cluster densities, sizes and ellipsoidal shapes. This algorithm, more powerful than existing ones, may be useful for remote sensing for land use.>

Describes application of scale-space clustering to the classification of a multispectral and polarimetric SAR image of an agricultural site. After polarimetric and radiometric calibration and noise cancellation, the authors extracted a 12-dimensional feature vector for each pixel from the scattering matrix. The algorithm was able to partition without supervision a set of unlabeled vectors from 13 selected sites, each site corresponding to a distinct crop, into 13 clusters. The cluster parameters were then used to classify the whole image. The classification map is much less noisy and more accurate than those obtained by hierarchical rules. The algorithm can handle variabilities in cluster densities, cluster sizes and ellipsoidal shapes.>

A new way to look at the learning algorithm in the cerebellar model articulation controller (CMAC) proposed by J.S. Albus (1975) is presented. A proof that the CMAC learning always converges with arbitrary accuracy on any set of training data is obtained. An alternative way to implement CMAC based on the insights obtained in the process is proposed. The scheme is tested with a computer simulation for learning the inverse dynamics of a two-link robot arm.