The modulated lapped transform (MLT) is used in both audio and video data compression schemes. This paper describes its properties and how it can be used to generate a time-varying filterbank. Examples of its implementation in two audio coding standards are presented.

A technique for generating narrow-band Gaussian audio noise based on the implementation of the multirate filter is described. The technique allows modification of the bandwidth, central frequency, and amplitude of the noise in real time without interrupting the noise generator. The frequency spectrum of the noise has sharp cutoffs which meet the spectral sensitivities of the human auditory system.

The authors introduce an image coding method which unifies two image coding techniques: variable-length transform coding (VLTC) and image-adaptive vector quantization (IAVQ). In both VLTC and IAVQ, the image is first decomposed into a set of blocks. VLTC encodes each block in the transform domain very efficiently: however, it ignores the interblock correlation completely. IAVQ addresses the interblock correlation by using a codebook generated from a subset of the blocks to vector-quantize all blocks. Although the resulting codebook represents the input image better than a universal codebook generated from a large number of training images, it has to be transmitted separately as an overhead, therefore degrading the coding performance at high bit rates.>

An image adaptive vector quantization technique is presented in the transform domain, where the codebook information is interleaved with the label information. The input image is first decomposed into a set of blocks. Then, each block undergoes a two-dimensional block discrete cosine transform (DCT). The DC coefficients are linearly quantized and encoded using a fixed-length code. The remaining AC coefficients are coded using adaptive vector quantization. The AC coefficients within blocks are combined into vectors from which a codebook is generated. In quantization, for each vector, the closest codeword is determined. If it is the first time the codeword is selected, the closest codeword is transmitted. Otherwise, the corresponding label is transmitted. To gain efficiency in transmitting the codewords, variable length coding (VLC) is used in encoding the codewords in the DCT domain. A reduction in bit rate required for transmitting the labels is accomplished by using a technique called codebook self-organization and by exploiting the correlation between the neighboring vectors.>

An optical character reader for processing typeset documents must be able to handle proportional spacing, the presence of touching characters and a wide variety of type fonts. This paper describes the design of a multifont character recognizer which uses a binary decision tree to classify a character on the basis of 197 geometric features. The algorithm for designing the decision tree is based upon an entropy minimization procedure, and makes no assumptions on the distribution or independence of the binary features. The decision tree classifier provides confidence measures which may be used to reduce the substitution error rate at the expense of higher rejection rates. Methods of reducing the overall error rate by combining the decision tree classifier with other classifiers were examined. In particular, the paper evaluates the performance of a classifier using a combination of multiple decision trees, template matching and contextual post-processing. Error rates were highly sensitive to typeface and varied between 10 percent and 0.1 percent. Computer processing times for the various stages of the system are presented.

The singular value decomposition (svd) is usually calculated by the Golub/Kahan QR algorithm. Two simple alternative methods for the partial or complete svd are described which are suited for special applications. Both of these can be implemented easily by the user, even on low-capacity microcomputers. Various properties and limitations of these two methods are discussed.

The paper demonstrates the application of adaptive vector quantization techniques to the coding of monochromatic and color pictures. Codebooks of representative vectors are generated for different portions of the image. By means of extensive simulations, the performance of the coder for monochrome images is estimated as a function of the various coding parameters-vector dimensionality, number of representative vectors, and degree of adaptivity. It is shown that for a vector size of dimension 4, there is a zone of operation, ranging from 1.0 to 1.5 bits/pixel, where adaptive vector quantization is advantageous. Despite the simplicity of the decoder, the performance of the adaptive vector quantizer is found to be comparable with block transform coding schemes for both monochrome and color pictures.

If the linear predictive coding filter coefficients are imagined as specifying a point in ann-dimensional space, then the set of all stable filters fills up a certain region in this space. The shape of this region is determined for low-order filters, and general properties are deduced for higher-order filters.

A method for computing the partial singular value decomposition of a matrix is described. The method is appropriate to problems where the matrix is known to be of low rank and only the principal singular vectors are of interest. The technique is simple, easy to implement in integer arithmetic, and places modest memory requirements. The convergence properties of the algorithm are investigated analytically and by simulation.

A clustering scheme using a multidimensional histogram stored in a table is described and tested on four-dimensional data derived from LANDSAT imagery. By doing all clustering operations on the histogram rather than on the original measurement vectors, it is possible to reduce the computations by a large factor and handle the large sample sizes that are typically encountered in image processing. The clustering algorithm first isolates and delineates the peaks in the multidimensional histogram. These peaks are then used as cluster centers, and all the other measurement vectors in the histogram are assigned to the nearest center. The scheme initially identifies the most separable clusters in the data. It then runs on an interactive basis allowing the user to split specific clusters into subclusters at the expense of less separability. The histogram approach lends itself to statistical analysis using parametric models and the likelihood ratio test. As a starting point, it is assumed that the observed distribution is a mixture of several multivariate Gaussian distributions with unknown mean vectors, covariance matrices, and a priori probabilities. Estimates of the Gaussian parameters are determined ignoring the overlap of the neighboring distributions. The theoretical histogram is then calculated by integrating numerically the probability density function in each of the cells of the histogram, and the likelihood ratio test is applied to measure the departure of the model from the observed data. A statistical measure taking into account the number of degrees of freedom is defined and used to choose between alternative models.