本計畫的主要目的在研發一個以GPU(Graphics Processing Unit)運算技術為基礎之機器人視覺系 統，以提供一個自主式機器人進行視覺互動行為應用。由於視覺互動行為中的影像處理需要即時且 大量的運算，因此設計一個兼備高效能與高擴充性的運算架構是必須先考慮之重點。除了必要的運 算架構外，影像處理中的彩色影像對比增強、彩色影像雜訊消除及平行化處理也是必須要建立的。 在彩色影像對比增強方面，將基於YCbCr 色差空間，設計出一套能平行化處理的區域型影像增強處 理技術，讓GPU 做有效率的運算與實現。在彩色影像雜訊消除方面，除了傳統灰階影像之雜訊消除 方法以外，由於近年來彩色影像應用的崛起，應用於彩色影像之雙向濾波器(Bilateral Filter)也是很值 得引用的技術。在雙向濾波器的設計上，為了減少色彩空間轉換的過程，將利用與影像增強技術相 同使用的色彩空間來做設計。同時，在提升雜訊濾除效能方面，由於雙向濾波器對於脈衝雜訊無法 提供有效的濾除能力，在雙向濾波器中加入脈衝雜訊偵測器是很值得探討的研究領域，稱為通用型 雙向濾波器。藉由以上技術的結合，預期能發展出一套具有即時運算能力與彩色影像處理功能的機 器人視覺系統。我們預計以一年時間完成此計畫，計畫之重點將集中在GPU 運算平台以及彩色影像 處理演算法的開發。 The purpose of this project is to develop a robotics vision system for autonomous robots to handle the visual interaction tasks based on GPU (Graphics Processing Unit) computing technology. This project involves the development of a GPU computing platform to perform color image enhancement and de-noising in real-time. Since the visual interaction tasks require large computations to achieve real-time applications, it is important to design the vision system architecture with high-performance and high-flexibility. In addition to the design of system architecture, the color image enhancement, color image de-noising, and parallelism processing are also necessary to be constructed. For the development of color image enhancement algorithm, a parallelized local image enhancement method based on YCbCr color space is preferred due to the GPU computing implementation. In the design of color image de-noising algorithm, the bilateral filtering method gains more attention for the applications of color image processing in recent years. In order to reduce the computation on the color space transformation, it is better to use the same color space for both color image enhancement processing and de-noising processing. In the meantime, in order to increase the image de-noising performance, it is important to add an impulse noise detector into the design of bilateral filtering since the bilateral filter lacks of the capability to remove the impulse noise. The design of bilateral filter with impulse noise removing is termed as the unified bilateral filter and is an important issue on the field of image de-noising. A robotics vision system with real-time computing power and color image processing capability can be constructed by the combination of the techniques discussed above. We plan to complete this study in one year. The project will focus on the development of GPU computing platform and color image processing algorithms.
