Chair: Pao-Ta Yu, National Chung Cheng University, Taiwan
Laurent Lucat, France Telecom - CNET/DSM (France)
Pierre Siohan, France Telecom - CNET/DSM (France)
Vector Median (VM) filters are known to globally perform better than scalar independent median filters around edges. In this work, we examine the filtering of vector edges corrupted by scalar impulsive noise. We show that VM performance is highly dependent on the relative magnitude of the impulses, the noisy component and the noise-free components of the edges. This characteristic gives some indications for adequately choosing the norm (L1/L2) of the VM filter and the standard (RGB/YUV) of data representation.
K.J. Willner, Tampere University of Technology (Finland)
Pauli Kuosmanen, Tampere University of Technology (Finland)
V.V. Lukin, Kharkov Aviation Institute (U.K.)
A.B. Pogrebniak, Kharkov Aviation Institute (U.K.)
Here we analyse some specific properties of several well known nonlinear filters applied to processing of ramp edges or rapidly increasing/decreasing linear parts of signals corrupted by Gaussian or mixed noise. It is shown that depending upon the slope, noise variance, filter type and scanning window size the efficiency of noise suppression and spike removal varies in rather wide range and can differ greatly from that one predicted on basis of standard approach to its analysis carried out for constant signals. Quantitative evaluations based on numerical simulation and partly confirmed by analytical derivations are presented.
Shivaling S. Mahant-Shetti, Texas Instruments DSPS R&D Center (U.S.A.)
David A. Martin, Texas Instruments DSPS R&D Center (U.S.A.)
There is growing interest in CMOS imagers, imagers which can be fabricated in standard CMOS process instead of a CCD process. One problem with these imagers axe the white spots resulting from high junction leakage due to point defects in the photodiode; these high leakage currents show up as spatially invariant "salt" noise. We propose two new non-linear filtering algorithms to compensate for the white spots. Conditional replacement of a pixel (CRP) is an image enhancement technique superior to a median filter with less implementation complexity. The second algorithm is a dark current estimator (DCE), in which the dark current variations in the defective pixels are progressively estimated and compensated. Neither algorithm requires a mechanical shutter or a specific calibration image. These algorithms axe described in detail, and the results of simulations are presented.
Hung-Hsu Tsai, National Chung Cheng University (Taiwan)
Shih-Che Lo, National Chung Cheng University (Taiwan)
Pao-Ta Yu, National Chung Cheng University (Taiwan)
In this paper, we propose a new class of nonlinear filters called rank ordered neural network (RONN) filters based on a detection-estimation strategy and neural network technique. RONN filters not only inherit the adaptive capability from neural networks but also exploit the rank information, from selected observation samples in contrast to that of traditional neural network (NN) filters. As expected, the experimental results reveal that RONN filters are superior than that of other existing filters.
Steven Armstrong, Cambridge University (U.K.)
Anil C. Kokaram, Cambridge University (U.K.)
Peter J.W. Rayner, Cambridge University (U.K.)
This paper proposes the use of min-max functions as a model for the interpolation and reconstruction of missing image data. It is shown how an interpolation equation based on these functions is formed and differentiated. The resulting solution is not closed form, therefore the derived gradient expressions axe employed as part of various numericaI optimisation schemes. Different interpolations can be found by using the squared and absolute errors the latter presents a more complicated solution which is discussed. Results are shown for the interpolation of missing data in an image. These are compared with interpolants derived using a 2D AR model and conclusions are drawn about the suitability of the new technique for reconstructing different image features.
Vijay Sundararajan, University of Minnesota (U.S.A.)
Michael E. Zervakis, University of Crete
Keshab K. Parhi, University of Minnesota (U.S.A.)
Area/power efficient implementation is provided for a new wavelet domain robust image denoising algorithm. A novel approach to system implementation is presented which dramatically simplifies hardware implementation by using a Euclidean-norm approximation technique. Simulation results are presented, which show that the exact and norm approximation based implementations have comparable performance.