Study On Spectral Image Reconsturction And Display Based On Multispectral Imaging Technology

Through the use of spectral image reconstructed based on multispectral imaging technology,the metamerism can be avoided,and the color reproduction among different digital mediums could be achieved accurately.In addition,the color appearance under arbitrary light sources can be represented on basis of the spectral image.Meanwhile,measuring the spectral reflectance with imaging method can overcome the contact measuring of the spectrophotometer and the point by point measuring.Simultaneously,the spectral reflectances of objects in the field of the camera can be obtained through just one shot,so the measuring area is extended vastly,accompanied by the greatly improved efficiency.Hence,this paper focuses on the deep study of theories and key technologies concerning the spectral image reconstruction based on multispectral imaging.The algorithms of spectral reconstruction,filter selection and training samples selection were firstly analyzed,then the filter selection methods based on light source and representative training samples were proposed,respectively.The key technologies of system calibration were studied based on the analyzing of the spectral response of each channel.What's more,the self-training based spectral image reconstruction method was developed to restrain the deterioration of spectral reconstruction caused by inconsistence between training samples and target.Afterwards,the approach for displaying spectral image was explored,and the accuracy of displaying under different light sources was verified.Since the light source and filters adopted in the multispectral imaging system would influence the accuracy of spectral reconstruction,the filter selection method based on light source was proposed combing the adaptive genetic algorithm,which converted the filter selection to finding the filter combinations fulfilling the best spectral reconstruction for target samples.The proposed method did not need to measure the spectral characteristics of light source,filters and sensor.The target samples comprising uniform and representative colors were selected out of the Munsell color book.Then,the optimal filter combinations with different numbers for three light sources were determined from 16 candidate filters of which the wavelengths of peak transmittance are all in the visible bands.As to the spectral reconstruction algorithms based on learning,the training samples are used for building the transforming model from multispectral responses to spectral reflectance,therefore the spectral reconstruction accuracy heavily depends on the training samples.The existing training samples selection methods were summarized,and the adaptive spectral reconstruction method was proposed based on selecting the correlated training samples for the test sample.Moreover,the representative training samples can reflect the characteristic of spectral reflectance for test sample,hence,the filter selection method based on representative training samples was developed to fully consider the spectral characteristics of filters,light source,sensor and test sample for spectral reconstruction.Due to the nonlinearity of electronic amplification and the stray light in the camera,the sensor response is usuallly nonlinear to the flux reflected from the object surface.The nonlinearity between channel responses of multispectral camera and the light intensity was analyzed upon the neutral patches in one commercial target,and this kind of nonlinearity was fitted with cubic polynomial.In the process of capturing multispectral images,the spatial uniformity of light is hard to be achieved,so the non-uniformity was analyzed and corrected for each channel using one uniform gray card.Moreover,the synthetic spectral sensisivity containing the spectral characteristics of light source,optical system,filter and sensor was estimated by the proposed non-negative PE method,simutaneously the single peak characteristic of narrow band was achieved.When the characteristics of spectral reflectances between target sample and training samples are not consistent,the spectral reconstruction accuracy for learning method will decrease.Therefore,the self-training based spectral image reconstruction method was developed to suppress the decrease of spectral reconstruction caused by the reflectance inconstance between the target sample and training samples.The multispectral image of the target was partitioned firstly by the k-means method,and one representative training sample was selected from each cluster.Then the multispectral image of the coordinate paper covered on the target was captured.The spectral reflectances of training samples were measured indirectly with the spectroradiometer.Meanwhile,the measuring area of the spectroradiometer was determined by the circle Hough transform,and mapped into the image of target using the multispectral image of the coordinate paper,then the multichannel responses of the training samples can be obtained.Through simulation,the proper combination of distance measure was determined,which was employed in the processes of clustering multispectral image and selecting the representative training sample in each cluster.Meanwhile,the SPRSQ was chonsen as the metric for determining the proper number of training samples,and the Kernel algorithm was selcted for reconstructing the spectral image of real objects through comparison.Afterwards,the spectral images of one watercolor painting,one oil painting,one ink and wash painting as well as one silk painting were reconstructed using the severally extracted training samples with the proposed method,simutaneously the spectral reconstruction accuracy was compared against that obtained using one commerical target.A professional LCD was employed to study displaying the spectral iamge.Firstly,the key colorimetric characteristics of the display were measured,including the preheating characteristic,chromaticity constancy,channel independence and the TRC(tone reproduction curve).Due to the lack of direct visualization,the spectral image needs to be coverted to CIE XYZ image under one light source or illuminant and further to RGB image,therefore the inverse characterization model of display is necessary.Then four inverse characterization models were compared,among which the best one was selected for the display of spectral image.Afterwards,the spectral power distributions of 6 light sources were measured and the RGB images of one watercolor painting,one oil painting,one ink and wash painting as well as one silk painting under each light source were calculated and represented on the display.In addition,the CIE XYZ tristimulus values of the test samples shown on the display for each painting were measured under the 6 light sources,meanwhile the CIE XYZ tristimulus values of the test samples were calculated on basis of the corresponding meausred spectral reflectances and the spectral power distributions of the 6 light sources.Then the display accuracy was verified by the color differences calculated with the two sets of tristimulus values.Finally,the main research contents and innovative achievements of this dissertation were summarized,and the perspectives of the future study were also forecasted.