Research About Multispectral Image Reconstruction Algorithnis With Partially Coherent Illumination

Spectral image contains not only abundant two-dimensional spatial information of the objects,but also one-dimensional spectral information,which widely applied in biomedicine,microbiological detection,military reconnaissance and other fields.However,spectral images contain a lot of information with the increase of spectral dimension,and traditional sampling methods are not suitable for spectral image acquisition.Moreover,when the objects have a low optical absorptivity,the contrast imaged is insufficient,which influences the analysis of spectral images.Compressed sensing framework based on partially coherent illumination improves the contrast while sampling spectral data at a frequencies far below Nyquist frequency.In the paper,based on compressive sensing,using dictionary learning method as a framework,combining coherent illumination theory,the research focuses on compressed sampling and accurate reconstruction of Multispectral Images.Firstly,the development and sparse model of compressed sensing are introduced.Sparse representation of signal is applied to spectral imaging,and the mathematical model of spectral imaging based on compressed sensing is analyzed.Secondly,the optimal principle for partially coherent illumination is analyzed,and a contrast enhancement scheme for spectral image based on light source optimization is presented.In the scheme,the effective light source is divided into several independent sub-light sources with Abbe imaging model,and the mathematical model of image intensity is formulated.Then,combining with the idea of optical transmission cross-coefficient of Hope imaging model,the transmission cross-coefficient is separated from the image intensity model,and the cost function of illuminant is established with the transfer function of the image.Then,the optimized illumination is obtained with the Particle Swarm Optimization algorithm.The optimal illumination source for image formation is obtained by optimizing the cost function.The simulation results demonstrate the optimized light source shape show better performance on image contrast than ordinary lighting model.Finally,based on compressed sensing theory,a tensor dictionary learning method for multispectral image reconstruction is proposed,which maps the two-dimensional spatial and one-dimensional spectral characteristics of multispectral images to thirdorder tensors.Different from a two-dimensional matrix dictionary,the third-order tensor dictionary divides the two-dimensional dictionary blocks into a set of subdictionaries which stored in the third dimension with the form of tensor data slices.In the tensor slices,each slice represents a dictionary block and corresponds to a series of coefficients,which makes the sparse expression of atoms have a higher degree of freedom and improves the reconstruction quality.The experimental results based on bio-multispectral data demonstrate that the Peak Signal to Noise Ratio and Structural Similarity Index of reconstructed image under tensor dictionary learning method are higher than traditional dictionary learning method.