Erlan H. Feria

Linear predictive transform of monochrome images

Digital monochromatic images are encoded using a novel minimum mean square error (MSE) linear predictive transform (LPT) coding formulation. The new formulation is appealing for two important reasons. First, it leads to simple coder implementation with a satisfactory signal-to-noise ratio (SNR). Second, it provides a general theoretical framework from which minimum MSE predictive coding and minimum MSE transform coding arise as special cases. Some specific results of this paper that illustrate the previous ideas are: a simple and generally suboptimum two-dimensional LPT coder operating at 2 bit pixer?1 has approximately one third the complexity of a 4 × 4 Hadamard coder while yielding a better SNR; an optimum 2D LPT coder operating at 2 bit pixer?1 has approximately one sixth the complexity of a 4 × 4 Karhunen-Loeve transform (KLT) coder while yielding a better SNR.

Predictive transform estimation (image processing)

Minimum mean squared error (MSE) linear predictive transform (LPT) source decoding (or modeling) and Kalman estimation are integrated to yield a unified approach to source modeling and estimation called PT estimation. PT estimation enhances classical Kalman estimation in two ways: first, it directly addresses the source modeling problem of scalar or multidimensional Kalman estimation by integrating an exact minimum MSE LPT decoder with a Kalman estimator; second, it provides a transformation mechanism that inherently leads to significant design and implementation simplifications when the state dimensionality is large. In the specific case of image reconstruction, the design and implementation requirements of 2-D LPT smoother structures are lessened with respect to those of classical 2-D Kalman smoother structures with exactly equivalent performance by factors that approach eight and four, respectively. Simple nonadaptive 2-D LPT smoothers perform quite well when compared with previous adaptive linear minimum MSE estimators. >

Matched processors for quantized control: A practical parallel-processing approach

Deterministic ‘quantized control’ problems are investigated from a novel and often computationally feasible ‘matched processors’ (MPs) perspective. It is noted that a bank of independent MPs, each matched to an admissible decision sequence, can evaluate, given the state and remaining duration, the dynamic performance cost-to-go values corresponding to a suitably selected ‘small’ subset of all admissible decision sequences. By choosing the best admissible decision sequence in this subset and using its first decision at each stage a computationally feasible and generally sub-optimum MP control scheme is produced. This scheme is found to be ideally suited for very-large-scale integration. In particular, for processes that are linear in the state and performance criteria that are quadratic in the state, the cost-to-go reduces to a simple quadratic function of the given state which in turn results in an MP control scheme structure similar to offered VLSI-based digital computers. It is demonstrated that the pro...

Predictive-Transform Source Coding With Subbands

Minimum mean squared error (MMSE) predictive-transform (PT) source coding is integrated with subband compression to further improve the performance of low bit rate MMSE PT source coders. A desirable byproduct of the advanced scheme is that the incorporation of joint optimum prediction and transformation from subband to subband is ideally suited to its integration with JPEG2000 to yield even higher compression levels while producing an outstanding objective as well as subjective visual performance.

Decomposed predictive transform estimation

A novel design and implementation decompositions are found to arise for a minimum mean squared error (MMSE) linear predictive transform (LPT) estimator when certain symmetry conditions are satisfied by the first- and second-order statistics used to design the estimator. This results in a decomposed LPT estimator whose design and implementation computational effort is significantly less than that of the original estimator. >

A Composite NTSC Color Video Bandwidth Compressor

Predictors operating on video signals sampled at six times the color subcarrier (6f_{c}) have been investigated and compared with predictors operating at three times the color subcarrier (3f_{c}) . When the best 6f_{c} predictor is used with an adaptive delta modulator (ADM) step size generator, the resulting pictures are similar in quality to 3f_{c} predictors operating with a four-level adaptive DPCM quantizer. An ADM coder was implemented in real time. The coder used between 24 and 34 ICs, depending upon the algorithm, and fits on a single 9 in × 6 in PC board. The transmission rate is 21.4 Mbits/s.

Latency-Information Theory: A Novel Latency Theory Revealed as Time Dual of Information Theory

This paper outlines a space-time duality study started in 2003 and leading to a latency-information theory (LIT) that unifies information theory with a novel latency theory revealed as time-dual. While information theory guides the design of communication systems, latency theory does the same for recognition systems. A unified recognition-communication system is an intelligence system and LIT illuminates its design. LIT naturally arose from the authors desire to systematically address the design of a real-world intelligence system for DARPAs knowledge-aided sensor signal processing expert reasoning (KASSPER) program. This work has led to practical intelligence system solutions that yield outstanding target detections under severely taxing environments, while also exhibiting several orders of magnitude savings in prior-knowledge storage-space, processing-time, and implementation complexity over standard schemes. Moreover, it has led to the discovery of a mathematical-physical duality guiding life system designs.

Accelerated predictive-transform

This paper presents a novel accelerated predictive-transform (APT) modeling methodology for use in compression. The APT scheme is illustrated with a monochrome 2D image compression application yielding very promising results. For example, when the Lena image is compressed to 0.056 bits per pixel and the pixel blocks are of dimension 16×16, it is shown that both the design and implementation computational complexity of the prior predictive-transform (PT) modeling art is improved by a factor of 12 without any loss in the visual quality of the compressed image. The APT methodology can also be used in other application areas such as estimation; detection, identification, channel and source integrated coding, and control and other related areas.

A One-Stage Look-Ahead Algorithm for Delta Modulators

This paper describes a one-stage look-ahead algorithm for adaptive delta modulators. The algorithm does not require the calculation of two encoding paths nor does it require the decision circuitry to choose the optimum path.

Predictive transform: a simple and optimum unified approach to signal modeling, coding, estimation, and control

The decoder of a minimum mean-square-error predictive transform (PT) coder is found to be the key to a simple and optimum unification and transformation enhancement of well-established techniques in linear and nonlinear signal modeling, coding, estimation, and control. The fundamental property of the PT decoder that makes this possible is that the PT decoder is a whitening model of the signal that one wishes to model, encode, estimate, and/or control. In addition, the PT decoder is capable of exactly identifying a large class of linear and nonlinear whitening signal sources. Hence, this whitening signal modeling property of the PT decoder allows its integration with well-known signal-processing techniques such as Kalman estimation and linear quadratic Gaussian (LQG) control. An added bonus of this integration is the appearance of a transformation mechanism that leads to significant simplifications of the Kalman estimator as well as the LQG controller. The transformation-enhanced Kalman estimator is illustrated with monochrome images embedded in additive white noise and is found to produce a design and implementation complexity that is simpler than that of a classical Kalman estimator-without transformation-by a factor that approaches four in the smoothing case.<<ETX>>