Multiscale contrast enhancement with applications to image f

This paper was published in

Multi-scale contrast enhancement with applications to image fusion

Alexander Toet

Copyright 1992 Society of Photo-Optical Instrumentation Engineers.

This paper was published in

Optical Engineering, 31(5), pp. 1026-1031

and is made available as an electronic reprint with permission of SPIE.

One print or electronic copy may be made for personal use only.

Systematic or multiple reproduction, distribution to multiple locationsvia electronic or other means, duplication of any material in this paperfor a fee or for commercial purposes, or modification of the contentof the paper are prohibited.

This paper was published in

Multiscale contrast enhancement withapplications to image fusion

Alexander Toet,MEMBER SPIEInstitute for Perception TNOKampweg 5

Soesterberg NL-3769-DE, The Netherlands

Abstract.A method to merge images from different sensing modalitiesfor visual display was introduced by Toet, van Ruyven, and Valeton in1989, which produces a fused image by nonlinear recombination of theratio of low-pass (RoLP) pyramidal decompositions of the original images.The appearance of merged images that are produced by this scheme ishighly dependent on the contrast and mean gray level of the input images.That nonlinear multiplication of the successive layers of a ratio of low-pass pyramid results in a contrast-enhanced image representation thatis highly invariant for changes in the global gray-level characteristics ofthe original image is shown. Application of this nonlinear multiplicationprocedure in the image fusion process results in composite images thatappear highly independent of changes in lighting and gray-level gradientsin the input images. The method is tested by merging different degradedversions of parallel registered thermal (FLIR) and visual (CCD) images.Subject terms: contrast enhancement, low-pass pyramid; multiscale image rep-resentation; multisensor fusion.

Optical Engineering 31(5), 1026-1031 (May 1992).

1Introduction

The workload of a human operator severely increases withthe number of imaging systems that need simultaneous mon-itoring.Moreover, a human observer cannot reliably inte-grate visual information by viewing multiple images sepa-rately and consecutively. The integration of informationacross multiple human operators is nearly impossible. Animaging system that fuses signals from multiple imagingsensors into a single image is therefore of great practicalvalue.

An image fusion method intended for human observationwas recently introduced.1In this scheme, the input imagesare first decomposed into sets of light and dark blobs ondifferent levels of resolution. This is done by computing aratio of low-pass (RoLP) pyramid2for each of the inputimages. A RoLP pyramid for the fused image is then ob-tained by selecting nodes with maximum absolute gray-levelcontrast from the sets of corresponding nodes in the RoLPpyramids of the inpidual images. The fused image is re-constructed from the set of pyramid nodes or pattern prim-itives thus obtained. As a result, perceptually importantdetails (i.e., details with a relatively high local gray-levelcontrast) of both images are preserved in the compositeimage. A serious shortcoming of this image fusion methodis its inherent sensitivity to variations in mean image inten-sity and global gray-level gradients.

This paper presents a scheme to enhance image contrastby nonlinear multiplication of successive layers of the RoLPimage decomposition. In this multiscale contrast enhance-ment process, image contrast at finer scales is weighted bylocal contrast extrema at coarser scales. As a result, imagecontrast is enhanced at all levels of resolution and fine-scaletexture becomes more visible. The recombination of a contrast-enhanced RoLP pyramid with a constant-valued top layerresults in a reconstructed image that is, to a high degree,independent of global variations in the mean gray level ofthe original image. Hence, the result of the image fusionscheme can also become highly insensitive to global vari-ations in the intensity of the input images when this mul-tiscale contrast enhancement procedure is applied to eachof the input images prior to the actual fusion process.The organization of this paper is as follows: Section 2introduces the nonlinear multiscale contrast enhancementscheme and shows some experimental results. Section 3presents some results of the application of this contrast en-hancement scheme in the image fusion process. Finally,some concluding remarks are given in Sec. 4.

2Toward an Invariant Image Representation2.1

Multiscale Image Decomposition

The RoLP (ratio of low-pass) pyramid was recently intro-duced by Toet.2The construction of the RoLP pyramid isvery similar to that of the popular difference of low-pass3(DoLP) or difference of Gaussians4(DoG) pyramid struc-tures. First a Gaussian or low-pass pyramid is generated for

1026 / OPTICAL ENGINEERING / May 1992 / Vol. 31 No. 5

This paper was published in

MULTISCALE CONTRAST ENHANCEMENT WITH APPLICATIONS TO IMAGE FUSION

the input image. This is a family of low-pass-filtered copiesof the input image, each with a band limit one octave lowerthan its predecessor.

Let array Gcomes the bottom, or zero, level of the pyramid. Each nodeocontain the original image. This array be-of pyramid leveli(1<i<N,whereNis the index of thetop level of the pyramid) is obtained as a (Gaussian) weightedaverage of the nodes at levelxi- 1 that are positioned withina 55 window centered on that node. Because of the re-duction in the spatial frequency content, each image in thesequence can be represented by an array that has half thedimensions of its predecessor.

The process that generates each image in the sequencefrom its predecessor is called a REDUCE operation becauseThus, forboth the …… 此处隐藏：18256字，全部文档内容请下载后查看。喜欢就下载吧 ……