6533b82dfe1ef96bd1291ad3

RESEARCH PRODUCT

Optimal extension of multispectral image demosaicking algorithms for setting up a one-shot camera video acquisition system

subject

description

Multispectral images are acquired using multispectral cameras equipped with CCD or CMOS sensors which sample the visible or near infrared spectrum according to specific spectral bands. A mosaic of multispectral MSFA filters is superimposed on the surface of the sensors to acquire a raw image called an MSFA image. In the MSFA image, only one spectral band is available per pixel, the demosaicking process is necessary to estimate the multispectral image at full spatio-spectral resolution. Motivated by the success of single-sensor cameras capturing the image in a single exposure that use CFA filters, we performed a comparative study of a few recent color image demosaicking algorithms and experimentally identified the best performing one, the algorithm based on the wavelet analysis of the luminance component. We have successfully combined biorthogonal wavelets with this approach with very satisfactory results. Color images are limited in the number of spectral bands and to remove this limit equivocation, we have extended our work to multispectral images. Thus, we have proposed a powerful multispectral image demosaicking algorithm which focuses on the green band and the luminance component. We first configured two MSFA models by the binary tree method, one with four bands and the other with five spectral bands with the dominant green band then proposed an algorithm which estimates the missing green components and other missing bands respectively by the method of convolution and weighted bilinear interpolation based on the luminance component. The images from the CAVE multispectral image base used for the simulations are acquired using cameras equipped with a liquid crystal tunable filter LCTF with low-energy sensitivity in blue bands. To balance the energies of the different spectral bands and avoid degradation of one spectral band for the benefit of another, we configured a new model of four-band MSFA with the dominant blue band and proposed a new demosaicking approach that combines the LMMSE method and the adaptive regression kernel. Indeed, we used the LMMSE algorithm to estimate the missing blue bands at the different pixels. For the other missing spectral bands, the directional gradient method which relies on the estimated blue bands is used for their interpolation. The adaptive regression kernel is then applied to each spectral band estimated at the different pixels for their update without persistent artifacts. The results of the various simulations show that our proposed approaches surpass existing algorithms both visually and quantitatively in terms of PSNR, SSIM and RMSE.