| Multisensor data fusion is a multilevel, multifaceted process dealing with datafrom multiple sources. It is used to eliminate the redundancy and contradiction ofthe multiple sensors, to decrease the uncertainty, and to provide new meaningfulinformation which cannot be acquired by using only a single sensor. As the amountof information rapidly growing and the increasing demand for analyzing theinformation from multiple sensors, multisensor data fusion as a newly subject, hasattracted people's attention greatly.The work of this paper on data fusion is based on the survey of practicalapplications of multisensor systems.The work focuses on two practical systems: thecorona detection system for the power plant and the air target identification system.Defective insulators, broken conductor strands or pollution may cause coronadischarge. Since corona may identify a defective component, it is important to findout the precise location of the corona source so that appropriate action may be taken.Visual observation of corona is difficult because it emits weak radiation, mostly inthe ultra-violet (UV) spectral range, and the corona is virtually invisible except indarkness. A dual-spectrum detection system is designed to detect and locate thecorona in the daylight. The system combines a solar-blind UV ICCD with a visiblecamera to image the corona source, thus generating the dual-spectral images for thesame observed site. The UV image is useful for detecting very weak UV emissionof corona sources and the visible image shows the background to indicate theposition of such a source on the equipment. The information of the images iscomplementary to each other. Then the dual-spectral images are fused by usingproper fusion scheme, the result image illustrates the corona information from theUV image and the background information from the visible image. It is effective forlocate the corona sources precisely. In this paper, we introduced the fundamentalorganization structure and the software framework of the dual-spectral data fusionapplication system.For the multisensor images acquired by the detection system, an efficientfusion method is needed to acquire the fused image that includes the backgrounddetail of the visible image and the position information of the UV image. Amultiwavelet based pixel-level data fusion algorithm is proposed in this paper. Byusing this algorithm, we can fuse the UV image and visible image while eliminatethe interrupt of the useless information. From the fused image, the user can locatethe corona precisely.Air target identification provides the basis for Battlefield Situation/ThreatAssessment. Due to the influence of sensor precisions, components of the integratedsystem and the external circumstance, the identification by a single sensor isuncertain and not suitable to air target identification. Gathering the multi-sourcedata with multisensor to acquire the complete information of the situation andcharacteristics of the observing target, and producing meaningful new fusedinformation can improve the precision of target identification and categorization.We have studied the theory of the data fusion, and proposed several new methods.Based on the application background mentioned above, we described our research ondata fusion from two aspects: pixel-level data fusion and decision-level data fusion.(1) A new image denoising method based on multiwavelet threshold shrinkageand subband enhancement is proposed in this paperMultiwavelet transform has simultaneous orthogonality, symmetry, compactsupport, and vanishing moment, which are not possible with wavelet transform.In this paper, a multiwavelet-based method is proposed to image processing. Animage is often corrupted by noise in its acquisition and transmission. Imagedenoising is used to recover the original image from the noisy data. We want therecovered signal to be as close as possible to the original signal, and to retainmost of its important properties. The information of edges is the most importanthigh frequency information of an image. Therefore we should try to maintainmore information of edges in the process of denoising. Thus we present a newimage denoising method based on multiwavelet threshold shrinkage and subbandenhancement. In this method, multiwavelet transform is the first step, the MWTcoefficients can be divided into two categories: the coefficients associated withnoise and the coefficients associated with edges. The coefficients associated withnoise are reduced by soft threshold multiwavelet shrinkage. But this procedure isa non-linearity transformation. It causes the edge smoothness. So we introducedsubband enhancement function to enhance the edge related coefficients. Theexperimental results showed that our denoising method, compared with thewavelet threshold shrinkage methods, can retain as much as possible theimportant signal features and increase PSNR. This method performs better thanthe threshold shrinkage method commonly in used.(2) A new image fusion method based on the features of the multispectralimages is proposed in this paperA new image fusion method based on the features of the multispectral imagesis proposed for the application of corona detection. In our application, thepurpose of image fusion is to indicate the accurate position of the corona by thehelp of the background information. The visible image represents the backgroundinformation and the UV image shows the corona position, the information of thetwo images is complementary. It is only in the fused image that we can observethe corona while locate its position. There are two popular methods in coefficientcombining: averaging and selection. The combining results acquired by using thetwo methods are not satisfied, the fusion image acquired by the selection methodlacked the background information for the corona, while the fusion imageacquired by the averaging method was too dark to distinguish the backgroundclearly.So we proposed a self-adapted combining method to fuse the usefulinformation of the multispectral images. This image fusion algorithm has beenapplied to process the multispectral UV/visible images. We acquired the fusedimages that include the corona information from the UV image and thebackground information from the visible image. The corona has sharp edge and iseasy to distinguish, and in the area of corona, we can observe details of thebackground, it's helpful to locate the corona precisely. With the subjectiveobservation and object evaluation, fusion results showed that the proposedmethod outperforms the discrete wavelet transform-based approach and the otherfusion methods.(3) An application system is designed for pixel-level multisensor data fusion.A kind of multi-spectrum image fusion application system is designed fordetecting and locating ultraviolet corona. A visible CCD and an ultraviolet CCDare used for input in the system. Two CCDs have the same visual field, so theprocessor can easily fuse two input videos, and then display the real-time fusedimage on LCD for user to observe and locate corona easily.A chip of DSP -TMS320DM642 from Texas Instrument is selected as theimage processor for dual-spectrum data fusion system. RF5 (Referenceframework 5) is provided from TI Company recently as starterware fordeveloping DSP applications. Since common elements such as memorymanagement, device drivers, and channel encapsulation are pre-configured inRF5, developers can focus on their system's unique needs and achieve betteroverall productivity. Basic data elements of RF5 were specially discussed, andthe software in the image processing board based on them was built up for thesystem achieving multi-spectrum image fusion. Until now, the development ofthe product has been accomplished;the sample machines have been put into use.In practice application, the software design method applied in this paper hasthe following characteristics: RF5 enables us to develop applications morequickly, more flexible, and easy-to-maintain than with traditional DSP software;RF5 enables us to call an eXpressDSP-compliant algorithm and organize theprogram flow conveniently;RF5 utilize the Code Composer Studio'ssupporting of simultaneous multi-threading fully. Putting the operation withtolerance of real-time demand in the lower-priority background task ensures thatthe operation execute correctly without affecting the real-time display of thesystem.(4) A new decision-level data fusion method based on D-S theory is proposedin this paperThe D-S Evidence Theory has been widely used in multisensor data fusionbecause of the capability to deal with the uncertain caused by unknown. But it isinvalid when dealing with conflicting probabilities, the highly conflictingevidences tend to produce counterintuitive results. If the evidence offered by onesensor is conflict with other evidences, it indicates that the sensor is in fault. Theconflicts indicate that the evidence offered by different sensors should not betrusted for the same extent. So, the new approach preprocesses the input evidence:if one or minority evidence conflict with the most evidence, the method reducesthe credibility of the minority of evidence, and increases their ignorance degreeso that it could weaken the influence on fusion result. As illustrated by simulateddata, compared with the other combination rules, the proposed method is similarto human experts. Precisions and reliabilities of the system are improved.
|
PDF Full Text Request