Predição de Light Fields Utilizando Redes Neurais Convolucionais

Abstract

Digital images are extremely commons and used in myriad applications. How ever, their bidimensional structure is insufficient for some applications as, for instance, representing immersive environments, post-capture refocus and illumination. Trying to solve this limitation, other representations like point clouds, 3d and omnidirectional images and light fields(LFs)– the last being the focus of this work. Light fields are, essentially, quadridimensional images. Given a set de points in a plane– referred as the sensor plane– LFs describe, for every incidence direction, a possibly different intensity. This can be interpreted in many ways, one of them being a bidimensional matrix where every entry is a bidimensional image, or, in this context, a sub-aperture image(SAI). Given they are quadridimensional structures, the volume of data involved to store and transmit them require a efficient compression. Various proposals were made, in a broader view, they can be classified in two big groups: those who organize the SAIs in a pseudo temporal sequence– which is then compressed using some video encoder, for instance, the High Efficiency Video Coding(HEVC)– and those which directly explore the quadrimentional nature of LFs. Neural networks were already used with success to a myriad of problems, including the processing of visual signals: dentification of written numbers, detection and identification of objects, facial recognition, amongst others. In particular, intra-frame prediction for compression of images and video approaches were developed using neural networks. With that in mind, this work proposes the creation of intra-LF prediction models using convolutional neural networks.