Method and apparatus to compress multi-spectral images to a single color image for display

#The present invention relates generally to image processing techniques. In particular, the invention relates to techniques for gathering image data at N different frequencies or frequency bands and mapping the non-redundant image data into a format suitable for display on a color display screen or printer using a finite number of colors, such as combinations of red, green and blue or on a black and white display or printer as gray-scales.

#Image processing is commonly used to acquire, manipulate, store and display images. It has many useful applications, which illustratively include the areas of photographic processing, photographic restoration and enhancement, medical imaging using x-ray and other frequencies of interest, and satellite imaging for capturing features on the earth's surface, to name a few. The images acquired are typically digitized into imaging data prior to processing, either directly through the use of a digital camera or indirectly by scanning a pre-existing image.

#Imaging data generally consists of an array of pixels, where each pixel represents a position on the image and a spectral profile of the position. The spectral profile is represented by an N-dimensional vector having N spectral intensities. For many applications, N is 3 and each of the three spectral intensities represents a respective intensity in the visible light portion of the spectrum, such as red, green or blue color. Combinations of red, green and blue intensities may be perceived as a large range of colors in the visible spectrum. For other applications, N may be greater or less than 3 and the respective spectral intensities may be measured within and outside of the visible portion of the electromagnetic spectrum. For example, in the case of the TMS and MAMS imaging sensors, N is 11 and includes frequencies in the infra-red portion of the electromagnetic spectrum. In the case of AVIRIS imaging data, N is 224 and includes a wide range of frequencies. AVIRIS and similar sensor data may include sufficient spectral information on the materials represented in the imaging data to allow classification of the materials.

#These and other objects, features and advantages of the invention will be more fully appreciated with reference to the detailed description and appended drawing figures.

#FIG. 1 depicts an image processor coupled to peripheral devices.

#FIG. 2 depicts a method of mapping N spectral bands to M bands with contrast enhancement according to the present invention.

#In the following sections, several embodiments of image contrast enhancement according to the present invention are illustratively described. Three different neural network approaches will be described to accomplish these objectives. The three are Input Training using the General Regression Neural Network (IT-GRNN), Input Training using the Back-Propagation Neural Networks (IT-BPNN), and the Auto Associative Neural Network (AANN). The same concept of mapping image data to a lower dimensional space and/or spreading the data to fill the display space can be expressed in terms of Fuzzy Logic, Genetic Algorithms, applied Statistics, and other neural networks.

#Image Processing and Multi-Dimensional Contrast Enhancement

#N co-registered spectral bands of an object captured by a camera can contain more useful information than the number of bands, M, which can be conveniently displayed on a monitor or a printer. The number of available bands is typically 3 for a color monitor or printer. However, the number may be 1 for a black and white monitor or printer, or possibly just 2 bands if one band is dedicated to overlay information, for example. What is desired is to distill as much as possible of the non-redundant information from the N spectral bands into M color planes for display, where N is greater than or equal to M. There are several methods to achieve this according to the present invention. One preferred method is to use a neural network to compress the information from N dimensions to M dimensions under the control of a fidelity constraint.