Extração de Mapas de Profundidades de Dense Light Fields usando Deep Learning

Abstract

Several imaging systems use image metrology to measure, identify, inspect and diagnose. The growing demand for visual metrology systems in several areas that require compact and robust sensors has driven the development of capture devices. These sensors use different physical quantities for imaging and calculating distances, each with its limitations and advantages. Among the emerging imaging technologies, which use only visible light information, we highlight the use of light field images captured through dense or sparse light field cameras, which have intrinsic advantages over traditional devices. For example, these cameras are robust in specific occlusion situations and also in scenes with noisy environments (rain, snow, etc.). This makes light field cameras, also called plenoptic cameras, potentially useful as a versatile sensor with multiple applications. This capacity is little used due to the complex optical characteristics related to the capture system and the computational cost of the related processing. To extract the depth map using traditional geometric methods, it is necessary to estimate n-variables, update their values and perform new calculations at each parameter change. When using artificial neural networks, these relations are already implicit in the neural network itself, which allows an immediate response to the dynamic modification of the parameters. This thesis presents two techniques for extracting depth maps from dense light field images based on neural networks with deep learning. The first proposal simplifies the EPINET network, reducing the flow from four inputs to just two inputs. The second proposal explores the multistream input network in an u-shaped convolutional neural network. Each proposal is explored and its advantages and disadvantages are presented. Both proposals calculate depth maps in less time than the original EPINET.