Depth Estimation from a Single Holoscopic 3D Image and Image Up-sampling with Deep-learning

Abstract

3D depth information is widely utilized in industries such as security, autonomous vehicles, robotics, 3D printing, AR/VR entertainment, cinematography and medical science. However, state-of-the-art imaging and 3D depth-sensing technologies are rather complicated or expensive and still lack scalability and interoperability. The research identified, entails the development of an innovative technique for reliable and efficient 3D depth estimation that deliver better accuracy. The proposed (1) multilayer Holoscopic 3D encoding technique reduces the computational cost of extracting viewpoint images from complex structured Holoscopic 3D data by 95%, by using labelled multilayer elemental images. It also addresses misplacement of elemental image pixels due to lens distortion error. The multilayer Holoscopic 3D encoding computing efficiency leads to the implementation of real-time 3D depth-dependent applications. Also, (2) an innovative approach of a deep learning-based single image super-resolution framework is developed and evaluated. It identified that learning-based image up-sampling techniques could be used regardless of inadequate 3D training data, as 2D training data can yield the same results. (3) The research is extended further by implementation of an H3D depth disparity -based framework, where a Holoscopic content adaptation technique for extracting semi-segmented stereo viewpoint image is introduced, and the design of a smart 3D depth mapping technique is proposed. Particularly, it provides a somewhat accurate 3D depth estimation from H3D images in near real-time. Holoscopic 3D image has thousands of perspective elemental images from omnidirectional viewpoint images and (4) a novel 3D depth estimation technique is developed to estimates 3D depth information directly from a single Holoscopic 3D image without the loss of any angular information and the introduction of unwanted artefacts. The proposed 3D depth measurement techniques are computationally efficient and robust with high accuracy; these can be incorporated in real-time applications of autonomous vehicles, security and AR/VR for real-time interaction.