Point Spread Function Calculation for Unfocused Plenoptic Camera by Random Forest

Abstract

A novel calculation scheme is proposed for the point spread function (PSF) of the unfocused plenoptic camera, which describes the response of image patterns to the point light source locations. It is based on the data-driven machine learning method, random forest (RF). Both synthetic and experimental images are obtained to build the datasets for training and test. By feature construction, transformed variables are selected as the inputs, the intensity values of the illuminated pixels are predicted by the RF. PSF estimation errors of the current scheme are compared with that of the widely used geometrical ray tracing method. The results demonstrate that the method by RF is obviously more accurate and robust, especially for the regions close to the focal plane, mitigating the estimation errors induced by optical aberrations.