A large variety of operations should be supported at the user interface, as indicated in our initial architecture. We propose to incorporate a system such as khoros [69][56] to support the variety of procedures that are required for many images. There are, however, a number of procedures that require additional research before they will be satisfactory for incorporation into the system. In particular we will address the following operations in detail:
Non-linear Algorithms for Image Enhancement: The goal here is to develop appropriate quadratic nonlinear filter for the enhancement and restoration of impulse-noise corrupted satellite images. Linear algorithms are used in most image processing applications due to their simplicity and mathematically tractable properties. However, linear algorithms do not perform satisfactorily in a number of situations, such as in the restoration of images corrupted with signal-dependent or multiplicative noise, as is often the case with satellite imagery. In addition, linear algorithms do not take into account the naturally occurring nonlinear properties of the human visual system or image acquisition systems.
We propose to investigate the application of non-linear techniques for such image processing functions as enhancement, image zooming, edge detection, image restoration, etc. In particular, we plan to investigate the application of non-linear 2-D quadratic filters, a special case of Voltera filters [61], for processing images in Alexandria. We have applied these filters in various image processing tasks resulting in high quality processed images. For example, we have used them for image contrast enhancement; a nonlinear unsharp masking method results in contrast enhanced images that are visually more pleasing than those produced using the conventional unsharp masking approach based on linear highpass filters [47]. We have investigated the use of quadratic filters in the impulse noise removal problem. Preliminary results from this investigation, indicate that our approach performs better than the more widely used median filters [46].
Image Zooming and Printing: The goal here is to design suitable non-linear adaptive interpolation scheme that preserves edges and fine texture detail of images in the digital library for image zooming. Image zooming is usually implemented as an upsampling step followed by a linear lowpass filtering operation. In many cases, however, this very simple and straightforward method does not yield acceptable results. Hence, we propose to investigate a nonlinear adaptive interpolation scheme that preserves edges and fine texture detail. A central part of this system is a nonlinear feature extraction filter. These features are then upsampled and interpolated by a process that adapts to the local image characteristics [66].
An important part of any image processing software package to be associated with a digital library is the ability to provide high-quality half-toned images for printing on a hard-copy device such as the laser printer. Most output devices cannot render the full gray scale resolution of typically 256 levels. This means that images have to be quantized (dithered) to fewer bits. We propose a new dithering algorithm which is an extension to the standard error diffusion algorithm where a nonlinear filter controls the adaptation of the error diffusion lowpass filter to the local image characteristics. Preliminary investigations indicate that this approach produces clearer images and preserve edge sharpness better than the standard approach [67].
Image Registration: We will investigate using the image features detected by non-linear filtering for indexing the image data. In many image processing applications it is necessary to compare multiple images of the same scene acquired by different sensors, or images taken by the same sensor but at different times. At present, there are no automated tools for registering these images available to scientists. Most of the techniques require considerable amount of human interaction with image analysts manually selecting corresponding salient features (or ``control points'' as they are often called) and using them to compute the registration parameters. We recently have developed a registration algorithm based on matching image contours [39] and have tested it on various satellite image data, including optical and radar images [37][39]. We plan to develop this algorithm further and address issues related specifically to the EOS data.
Image Fusion: Intelligent integration or fusion of information from multiple sensors is an important problem with potential applications to automatic target recognition (ATR), target tracking, navigation, and medical image interpretation. Our research will address sensor fusion from an information theory point of view, developing efficient and robust fusion algorithms based on wavelet transforms. We will define a data- and algorithm-independent performance measure that quantifies information preservation and compare the performance of existing algorithms with our proposed scheme. We will discuss ways in which context and prior knowledge about the problem domain can be included in the design and development of the algorithm, and illustrate the techniques for the case of IR and optical data fusion for ATR applications. We are currently exploring a pixel data fusion algorithm based on wavelets [38] and the initial results are very encouraging. This algorithm will also be consistent with the representation and storage of information and hence will not involve any additional computations for wavelet decomposition.