Reconstrucción de imágenes a color utilizando muestras incompletas de nueve bandas espectrales

Abstract

In recent years, the use of multispectral cameras has been increasing due to their ability to analyze different materials’ characteristics. Thus, multiple methods to develop them are being studied. A new prototype of a multispectral filter array type multispectral camera, which uses nanowires as filters, has been developed at the Nanophotonics research group at Harvard University. In this type of camera, a filter is superimposed on each of the sensor’s pixels to capture a specific band of light. Since the prototype is under development, it does not have all the functions that a multispectral camera should have. This work is focused on developing one of these functions, which is the reconstruction of color images from its nine captured bands. The development of the color reconstruction function will be carried out in three stages: pre-processing, model acquisition, and post-processing. For the model’s acquisition, the methods of support vector machine (SVM), k-nearest neighbor (k-NN), linear and multivariate nonlinear regression were used. Likewise, in the reconstruction process, high dynamic range (HDR) images were introduced. The results obtained show that the best method to obtain the color model is third-order nonlinear regression. On the other hand, it is essential to note that the k-NN and SVM methods showed better results in a mean squared error (MSE), but not in structural similarity because the color reconstruction in these methods is limited to the 24 colors that are featured in Macbeth’s grid. Finally, the use of HDR in color reconstruction improves the quantitative results in the MSE metric by 29.3 percent.

Date of Award	2021
Original language	Spanish (Peru)
Supervisor	Víctor Manuel Murray Herrera (Asesor)