The Study Of Multispectral Image Acquisition And Reflectance Reconstruction For Transient Target

In realistic lighting environment,various target scenes are in constant motion instead of being stationary.Based on the current level of optical devices and image sensors,a tri-vectors snapshot multispectral imaging system for specific target scenes was designed.This system utilized the advantage of three channel cameras: short exposure time,high spatial resolution and high energy utilization.To reconstruct the spectral reflectance,a database of specific target scenes was established.Through this database,each vector or pixel output value of specific transient target images obtained by three channel camera was reconstructed to the spectral reflectance.Then the whole spectral or reflectance data cube was established by reconstructed spectral reflectances.The spectral reflectance is considered as the “fingerprint” of an object.The main work of this dissertation was to create a mapping between the low-dimensional channels and the high dimensional spectral reflectance,furthermore reconstructing the spectral reflectance of target scenes accurately.To improve the reconstruction of spectral reflectance,we proposed three reflectance reconstructing methods,which were featured with dynamic selecting and weighting spectral reflectances of specific target training database.The first reconstructing spectral reflectance method was tolerable color difference.This method respectively averaged reflectances of the selected subset from main training samples by prescribing a limitation to the tolerable color difference between the training samples and the camera vector outputs.Different tolerable color differences of training samples were also investigated by this method.The second reconstructing spectral reflectance method was weighting smaller color difference group.To obtain the spectral reflectance of a vector of specific target scene in images,this method assumed that reflectance was a weighted average of reflectances of samples in a selected training group,in which all samples have smaller color difference with that vector.Four weighting models were proposed: the reciprocal square weighting model,the reciprocal weighting model,the reverse squared weighting model,and the reverse weighting model.For the proposed 4 models,different selected numbers of training samples were investigated.Among the four weighting models,the reciprocal square weighting model showed the highest accuracy of spectral reflectance reconstruction and the lowest susceptibility of number of selected training samplesThe third reconstructing spectral reflectance method was improved by adaptively combining colorimetric and reflectance similarities.Specific training samples were respectively sorted based on colorimetric and spectral reflectance similarities.The weighting coefficient matrix was generated through these two sequences.Then the accuracy of the reconstructed spectral reflectance was improved by applying weighting coefficient matrix to the classical method.The influence of different exponential factors of weighting matrix and different sizes of subsets of training samples were investigated by different training samples.Experimental results showed that all the three proposed spectral reconstructing reflectance methods exhibit considerable improvement in terms of the spectral reflectance errors as well as the color differences under different illuminants.The first method is faster in reconstruction of spectral reflectance,the rang of the reconstructed reflectance is between 0-1,which make it have actual physical meanings.The reconstructed speectral reflectance is also between 0-1 by second method,but with high accuracy.The accuracy of third method is high.Besides,there is almost no change for the accuracy of third method when the size of the subset exceeds that of small-sized training samples.We can directly choose a larger subset or the whole training samples for convenience.It can provide a reliable approach to solve the problem of indeterminacy of the training sample subsets number by using a larger number of training sample subsets or all training samples.Both theoretical analysis and experimental results demonstrated that the spectral reflectance and chromatic of a specific transient target scene can be highly reconstructed by using three primary color camera to instantaneously obtain transient target scenes with cooperating the three proposed reconstructing reflectance methods.This research has important theoretical and practical significance for the real-time acquisition and the transient target scene detection,like marine monitoring,agricultural monitoring,environmental assessment,real-time telemedicine and other industrial fields.