以细胞神經網路為基礎的拍攝影像穩定技術及其類比電路實現

Abstract

近年來，多數的研究已經說明在細胞神經網路(Cellular Neural Networks; CNN) 型態的架構下，能夠提供一個可程式化的方式來處理多數複雜的影像處理工作。CNN的架構中包含了可做即時處理的平行類比計算單元，而這些處理單元是有規則的二維陣列排列，且本身與鄰近的細胞單元為區域性的元件連接。由於此種特性，使得這種架構很容易在超大型積體電路上實現。因此本論文提出以CNN為基礎的影像穩定技術與它的類比電路實現。整篇論文可分成三大主要部份：(1)影像穩定技術；(2)建立於CNN為基礎的影像穩定技術；及(3)以CNN為基礎，實現上述所提到技術的類比電路設計。針對影像穩定技術，其中包含了計算與補償震動所產生的移動向量的兩大主體，本論文中為了加強即時運算方面的能力，將此演算法設計成以CNN為基礎的架構，可以用平行即時處理的類比式電路來實現。 本論文採用應用導向的CNN類比電路來實現提出的影像穩定技術。而CNN的設計電路採多層次 (Multi-layer) 的方式，以3×3為基礎的細胞核心，處理陣列大小是影像的1/120。同時為了降低電路複雜度，採用電流模式 (Current mirror;電流鏡) 的設計架構，且延伸成為可正負雙向電流導通，更容易來實現每個神經細胞的權重比例 (即電流增益)，也使得在節點上的多數訊號易於結合。由於CNN具有陣列式平行處理和區域性的元件連接特性，因此很適合實現於混合訊號標準的CMOS製程上。 最後，展現所提技術在MATLAB和HSPICE電路驗證上的模擬結果。這兩種模擬結果，都顯示我們提出以CNN為基礎的方式，在影像穩定補償上，擁有快速有效且即時的處理能力。

Recently, many researches have been proved that a programmable method can perform most of complicated image process tasks under the architecture of Cellular Neural Network (CNN). Real-time and parallel analog computing elements are contained in the architecture of CNN. There is an ideal characteristic that each computing unit is regular two-dimensional array and connects with its neighborhood locally. Because of this characteristic, the architecture of CNN is easy for VLSI implementation. There are three parts in the thesis: (1) image stabilization technique, (2) CNN-based image stabilization technique and (3) its analog circuit design by (2). Image stabilization technique contains two main blocks. One is the computation of the motion vectors caused by vibration and the other is in compensation for the motion vector. In the thesis, in order to enhance the ability of real-time processing, the algorithm is designed to be the CNN-based architecture and with the implementation of parallel and real-time analog circuit. We aim at the property of the proposed algorithm to design application-driven CNN circuit for the image stabilizer. The CNN circuit is a multi-layer structure and its template is 3□3. The size of CNN array is 1/120 of an image. Current mirror is used to reduce complexity and to extend positive and negative current for weighting of each cell. Current signals are easily combined in the same node as well. Due to parallel processing and local connectivity, CNN is suitable for implementation of standard mixed-signal CNOS process. Finally, the simulation results of MATLAB and HSPICE show that the proposed CNN-based image stabilization technique has the fast and real-time processing ability in image compensation.