Study Of Multispectral Video Acquisition System And Algorithm

The spectrum is a characteristic distribution of the electromagnetic radiation in the full frequency range that a particular object absorbs or emits.Traditional RGB imaging only obtain three channels(i.e.Red,Green,Blue),which are the down-sampled visible spectrum ranging from 400nm to 700nm.Compared with the RGB content,multispectral images have numerous times channels in the spectral dimension,with which the inherent material characteristics of different scenes or objects can be precisely obtained.Therefore,the spectral data is widely used in material identification,precision agriculture,environmental monitoring,biological sciences,criminal investigation and many other fields.Conventional multispectral imaging technologies generally utilize scanning schemes,and they fail to capture spectra of dynamic scenes/objects.Meanwhile,the compact light-weight low-cost multispectral video camera which doesn't need specially-made optical devices and is easy to use is the demand of industry.Focus on how to build a compact multispectral image/video acquisition camera,how to improve the accuracy of multispectral image/video reconstruction and how to effectively remove the Poisson noise in multispectral image/video,the main contribution of the dissertation includes:1.Build a mobile system(21 Omm × 190mm × 40mm,1.3kg)for MS video acquisition at high spatial resolution(2048 × 1536),high spectral resolution(400-1000nm,140 channels)and high temporal resolution(15 fps),which does not require specially-made optical devices and is easy to build with just a few off-the-shelf components;2.Propose a reconstruction algorithm for hybrid-camera coded aperture multispectral acquisition systems.Compared with the conventional coded aperture spectral reconstruction algorithm,we utilize the high spatial resolution image/video to optimize the spectral reconstruction,and the accuracy of multispectral image/video reconstruction is efficiently improved.3.Propose a fast multispectral image denoising method which is based on a successive two step framework,i.e.projection plus filtering.The 3D image cube is projected into the 2D image space to form a candidate reference image with much lower noise than the input MSI,and a 2D spatial denoising is employed to derive the reference image.By applying the denoising process in the 2D reference image space,the computational cost is significantly reduced.After then,the original 3D spectral image is denoised by using the reference guided filtering.