Research On Multi-sensor Image Fusion Technology Based On Multi-scale Analysis

Image fusion is a processing technology which integrates image or image sequenceinformation on the same scene that is acquired by identical or distinct modalities imagingsensor at the same or distinct time into a composite image that is more suitable for thepurposes of human visual perception or computer processing tasks. Multi-scale analysistechnology is derived from the simulation of human perception process for object incomputer vision study. The multi-scale analysis for an image can provide different featuresub-bands from high to low resolution and establish the relation between local and globalprocessing, this structure is consistent with the process of the human visual system perceivingobjects from coarse to fine. So better fusion performance can be obtained when multi-scaleanalysis technology is applied in image fusion domain. In this dissertation, the multi-sensorimage fusion technologies based on multi-scale analysis has been probed and researched onthe following four sections:1. Image fusion technology based on double density dual-tree complex wavelettransform (DD-DTCWT) has been probed and researched. DD-DTCWT is realized byoversampling technique based on two distinct scaling functions and four(two sets) distinctwavelet functions, where the two wavelets in the same set are designed to be offset from oneanother by one half and the wavelets corresponding to each in the distinct set form anapproximate Hilbert transform pair. DD-DTCWT has sixteen main directions and eachdirection is described by two wavelets. DD-DTCWT has anti-aliasing and approximatetranslation invariance. So DD-DTCWT can characterize more detail and improve theaccuracy of image decomposition and reconstruction. In the fusion criterion section,according to the fact that there is a certain correlation among pixels in the local window andthe amount of information carried by the pixels at different spatial position within thewindow is different, so in the low frequency sub-bands, the fusion criterion of weightedwindow coefficient variance as the center point coefficient weight is designed, in the highfrequency sub-bands, the fusion criterion of coefficient absolute value with the neighborhoodaverage energy consistency selection is designed. The defect caused by the traditionalcriterion that the brightness of the fused image is low and he fused image is easy to beaffected by noise is improved effectively. It has been validated by the simulation experiments that the researched fusion technology in this section can eliminate artifact phenomenoncaused by2D DWT with no translation invariance and improve fusion accuracy.2. Image fusion technology of pulse coupled neural network (PCNN) based on regionsingular value decomposition(SVD) in the nonsubsampled contourlet transform (NSCT)domain has been probed and researched. The realization method of NSCT is followed:mapping approach and ladder structure are applied to extend the1D filter to the2D filter,pyramid and fan filters that satisfy the Bezout identity are designed. Nonsubsampled pyramidconstructed by pyramid filters is exploited to decompose the image and band-pass sub-bandsare performed to direction analysis by fan filters, so NSCT with translation invariance、welldirection selection and anisotropy is obtained. In the fusion criterion section, aimed at thedefect that the contrast of the fused image is descended easily or blocking artifacts arisethrough the criterion of weight or mach by local window activity measure, according to thefact that the square of the singular value (SV) matrix’s F-norm represents the energy of theimage and most of the image structure information is contained in the singular value matrix, anovel pulse coupled neural network fusion criterion based on region singular valuedecomposition is designed. A local structure information factor (LSIF) that served as thelinking strength of each neuron in the PCNN is constructed by local region singular value,after the fire processing of the PCNN, new fire mapping images that reflect the feature of asingle pixel and the overall features of its neighborhood pixels are obtained. By thecompare-selection operator with the fire mapping images pixel by pixel, the clear objects ofthe sub-bands are selected and then are merged into a group of clear new sub-bands. It hasbeen validated by the simulation experiments that the researched fusion technology in thissection can eliminate pseudo-Gibbs phenomenon and improve the contrast of the local areaand overall definition3. Image fusion technology based on image quality assessment parameter innonsubsampled shearlet transform (NSST) domain has been probed and researched. TheNSST of an image is a two steps process: Multi-scale analysis and directional localization.First, an image is performed for multi-scale decomposition by nonsubsampled laplacianpyramid to obtain a low-pass sub-band and some band-pass sub-bands. Then Meyer waveletfunction is exploited to produce Meyer window function for realizing small sized shearingfilter. Band-pass sub-bands are convolved with shearing filters to obtain direction sub-bandsin each scale, so the NSST of an image is completed. The NSST with outstandinglocalization、high direction sensitivity and parabolic scaling can capture detail informationwell in the image. In the fusion criterion section, aimed at the defect that the poor structuralinformation and poor definition arise for inappropriate weight through local window activityto weight. A fusion criterion is designed that the structural similarity index with spatial frequency is used as weight of the central coefficient in low frequency sub-band andcoefficient absolute value with its neighborhood average energy is consistent to select orweight the central coefficient in high frequency sub-bands. For evaluating on image qualitybefore fusion，the structure information in the source image is retained effectively and edgedetails are extracted exactly. It has been validated by the simulation experiments that theresearched fusion technology in this section can avoid Gibbs type ringing and blockingartifacts in reconstructed image, can enhance the structure information and improve definition4. Image fusion technology based on finite discrete shearlet transform (FDST) has beenprobed and researched. The FDST of an image is a three steps process: frequencyconversion、multi-scale decomposition and direction analysis. First,2D fast Fourier transform(FFT) is implemented to obtain spectrum image and low frequency sub-band and highfrequency sub-bands are obtained by low-pass and band-pass filtering for spectrum image, soFDST has the multi-scale characteristics; Secondly, the frequency shearlet constructed by theauxiliary function are multiplied with band-pass sub-bands to obtain direction sub-bands;Finally,2D inverse FFT is perform to each sub-band for obtaining time-domain FDST in eachscale. FDST has the parabolic scaling、translation invariance、high direction sensitivity、outstanding direction analysis and high computational efficiency. In the fusion criterionsection, aimed at the defect that the fusion accuracy of edge texture is poor by existing typicalmulti-scale fusion criterion, a fusion criterion is designed that the gradient informationsimilarity index is used as weight of the central coefficient in low frequency sub-band toimprove the fusion accuracy of edge texture and coefficient absolute value with its regionstandard deviation is consistent to select or weight the central coefficient in high frequencysub-bands to improve the definition between the target contour and the non-contour. It hasbeen validated by the simulation experiments that the researched fusion technology in thissection can improve the fusion accuracy of edge texture and the computational efficiency。...