使用細胞神經網路架構實現影像穩定處理之震動向量估測晶片

Abstract

利用硬體估測攝影裝置於拍攝時，因手晃動或是機體支架震動所產生之不預期背景震動向量，以提供影像穩定系統(Image tabilization)補償為本論文的主要貢獻。影像穩定系統包含背景震動向量估測單元及震動向量之補償單元兩部分，然而計算背景震動向量佔去整個影像穩定系統大部分的處理時間，因此本論文提出一特殊應用導向之晶片(ASIC)，處理拍攝影像背景震動向量的估測。晶片設計採用仿細胞神經網路(Cellular Neural Network, CNN) 架構用以偵測背景震動向量的特徵；仿細胞神經網路之基本架構為一個和週圍細胞(cell)相連並規則排列之二維陣列，具有即時平行運算處理的能力。本論文提出使用CNN可調節偏壓的設計，快速計算影像穩定系統中所需要的背景震動向量，並配合全區域輸出連結鍊(global output connect chains)之設計，將區域背景震動向量位置定址出來。此外，區域類比記憶體(Local Analog Memory)陣列的設計來儲存影像差值資訊。本論文所設計之背景震動向量估測晶片採用TSMC 0.35um混合訊號製程，其大小為8.1mm2 共包含 個像素處理單元。電路模擬及CNNUM分析結果證實採用以CNN實現IS系統中估測背景震動向量具有遠高於DSP處理器的運算速度，讓整體IS系統具有即時運算之能力。

The objective of this thesis is to investigate the hardware design in image stabilization (IS) technique for local motion vectors (LMVs) in the image sequences. The IS technique is used to remove unwanted shaking phenomena in the image sequences captured by hand-held camcorders without affecting moving objects in image sequences and the intentional motion of panning condition, etc. It consists of motion estimation and motion compensation. Most of the complex and time consuming computations occur in motion estimation, an application-specific IC is designed to solve this problem. Cellular Neural Network (CNN) technology is used to implement the local motion estimation chip. CNN is a regular two-dimensional array and connects with its neighborhood locally. Real-time and parallel analog computing elements are contained in the architecture. CNN adaptive threshold template is proposed to extract reliable motion vectors from a given region. The design of global output connected chains can easily decode the LMV address. The local analog memory (LAM) is designed to store image difference information. The size of CNN array is 19 25 pixels. The chip has integrated in the total area of 8.1mm2 by using TSMC 0.35um mixed-signal process. Results with HSPICE simulation and CNNUM analysis prove that the performance of the proposed CNN-based local motion estimation is better than that of a digital signal processor so that the IS system has the capability of real-time operations.