利用曲線擬合特性之應用於深度圖邊緣強化與之超解析重建技術

Curve Fitting Based on Edge Enhancement of Depth Map with Super-resolution Reconstruction

林治均(Chih-Chun Lin)

指導教授 ： 李佩君

國立暨南國際大學/科技學院/電機工程學系/碩士(2018年)

摘要

隨著科技日新月異，人們對於影像品質的要求逐漸地提升，隨著DVD演進至現今的HD與Ultra HD，解析度的需求是必須的。高解析的視訊或影像具有高品質的細節資訊以及豐富的色彩內容，因此在各領域之重要性可見，如:數位相機與攝影機、醫學影像、衛星影像和監測系統等皆需要畫質提升之應用。影像經由擷取系統獲得時，通常有可能由於成像條件、成像因素與硬體限制等會使影像遭到破壞而轉變成低解析度影像。而除了彩圖資訊外，深度圖資訊之強化與重建亦是提高畫質與整體立體感品質的關鍵，透過超解析重建理論可使破損與遭到劣化的影像藉由恢復與重建技術來獲取高解析資訊。超解析重建的理論範疇甚廣，本論文討論與著重處在於深度圖邊緣處強化之超解析重建技術。
本論文之研究重點主要探討:如何利用曲線擬合之多項式回歸關係模擬深度圖邊緣處遺失的資訊，對邊緣處採取多項式回歸來重建失去的高頻資訊，對非邊緣處採用傳統的bicubic內插法以減少計算複雜度，利用超解析重建架構來生成高解析度影像序列，以多張影像序列做輸入，結合空域與時域資訊，當下幀部分利用傳統內插法提升解析度，並利用鄰張相關資訊產生之殘差量，帶入相對應之濾波器以提升超解析重建之精度。
實驗結果顯示，利用本論文提出之方法所重建的高解析度影像序列內容之PSNR相較於傳統的bi-cubic內插法平均高約1 ~ 2 dB，邊緣處命中率也比傳統方法高約10~20 %，可有效提升視覺品質。

關鍵字

超解析重建理論 ； 邊緣強化 ； 曲線擬合

並列摘要

With the ever-changing technology, the requirements for image quality are gradually increasing. As the DVD evolves into today’s HD and Ultra HD, the demand of high resolution is necessary. High-resolution video or video with high-quality details and rich color content make it important in all fields, such as digital cameras and video cameras, medical imaging, satellite imaging and video surveillance systems are needed to enhance the picture quality. When an image is acquired via an acquisition system, it is usually possible to transform it into a low-resolution image due to imaging conditions, imaging factors, and hardware limitations that can cause the image to be ruined. In addition to the color map information, the enhancement and reconstruction of the depth map information is also the key to improving the image quality and the overall quality of the three-dimensional image. By using the theory of super-resolution reconstruction, the damaged and corrupted images can obtain lost information through the techniques of restoration and reconstruction. The coverage of super-resolution reconstruction theory is very broad. This dissertation discusses and emphasizes on the super-resolution reconstruction of the enhancement at the edge of depth map.
The main research topic of this paper is: how to use the polynomial regression of
curve fitting to simulate the missing information at the edge of the depth map. At the edge of depth map, polynomial regression is used to reconstruct the lost high frequency information. For the non-edge area, the traditional bicubic interpolation is adopted for reducing compute complexity. Finally, the super-resolution reconstruction architecture is set up to generate high-resolution image sequence. The input is set as multi-frame image sequence. Combining with spatial and temporal information, the traditional interpolation is used at current frame to enhance the resolution and adjacent frame information generated by the residual is utilized to substitute into the corresponding filter to improve the accuracy of super-resolution reconstruction.

The experimental results show that the PSNR of the high-resolution video sequence reconstructed by the proposed method is 1 ~ 2 dB about average higher than the traditional bi-cubic interpolation method, and the hit rate at the edge is also about 10 ~ 20% higher, can effectively enhance the visual quality.

並列關鍵字

Super-resolution reconstruction theory ； Edge enhancement ； Curve fitting

參考文獻

參考文獻

[1] Andrew Burnes, Nvidia-Geforece動態超解析度提升您的遊戲體驗，讓您在 HD 顯示器上也能感受 4K 的畫面品質(Nov. 17 , 2014), [Online] Available: (https://www.geforce.com.tw/whats-new/articles/dynamic-super-resolution-instantly-improves-your-games-with-4k-quality-graphics), (Jan. 2018.)

[2] 雷鋒網，Google 超解析度技術 RAISR：模糊圖片瞬間變清晰，運算速度快十倍(Nov. 16, 2016), [Online] Available:

(https://technews.tw/2016/11/16/raisr-rapid-and-accurate-image-super-resolution/), (Jan. 2018.)

[3] C. Fehn, ”A 3D-TV Based on Video Plus Depth Information,” in Proc. of the Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, Vol. 2, pp. 1529-1533, Nov. 2003.