| Summary: | 碩士 === 淡江大學 === 電機工程學系碩士班 === 102 === With the advance of endoscopic technique, the researches of endoscopy are urgently needed for medical devices. However, the endoscopy is usually implemented in monocular image, making it difficult for doctors to judge the visual depth. It therefore may cause unnecessary damages to other tissues. This research work proposes a method using monocular endoscopic video to convert the 2D endoscopy into 3D view. Considering the depth estimation of the entire image, it needs a huge amount of computations, which cannot be attained throughout real time process. We propose the hierarchical similarity analysis to split the images into 16×16, 8×8, and 4×4 to analyze the similarity and depth estimation. However, the depth image may change drastically to make the converting view to have holes with different sizes and directions. It will affect the qualities of the 3D appearance and may take more time for hole-filling. This research hence presents a content-adaptive filtering technique to modify the depth map. A pre-process that applies the respective filter to different sizes and directions of holes is used to optimize the depth map. After the estimation of the depth map, the 2D image can be converted to a 3D image on the 3D display by DIBR (Depth Image Based Rendering) and create 3D modules through OpenGL (Open Graphics Library). Furthermore we implement the proposed approach on GPU (Graphics Processing Unit) to achieve the real time processing.
|
|---|
