以CNN為基礎的紋理邊界偵測技術及其類比電路實現

Abstract

近年來，多數的研究已經說明在仿細胞神經網路(Cellular Neural Networks; CNN) 型態的架構下，提供一個可用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，其中有一個理想的特性是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。因此在本論文中提出以CNN為基礎的紋理邊界偵測之新的影像處理系統與它的類比電路實現 本論文所提出的紋理邊界偵測技術，是模仿人類眼球表面層上的結構行為來偵測影像的紋理邊界。利用多數且平行CNN處理器計算技術的創新，取代以往複雜的數位式影像紋理邊界偵測。對於即時運算方面，它被設計成以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，大大地增加其執行的效率。而CNN的設計電路採多層次 (Multi-layer) 的方式，以5×5為基礎的細胞核心，將處理影像大小擴展成32×32處理陣列。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。

In recent years, many researches have introduced that a programmable method which computes many complex image processing tasks is offered based on the architecture of Cellular Neural Networks (CNN). The architecture of CNN consists of the analog computational units which can do real-time and parallel processing. One ideal property of CNN is that the signal values are placed on a regular geometric 2-D grid, and the direct interactions between signal values are limited within a finite local neighborhood. Based on this property, the architecture is easily implemented on VLSI. Therefore, a new image processing of CNN-based texture boundary detection and its analog circuit implementation are proposed in this thesis. The proposed texture boundary detection technique in this thesis imitates the behavior of the architecture on the surface layer of human eyeballs and then detects the texture boundary of images. The technique use the innovation of many and parallel computational processing units of CNN to replace the complex digital texture boundary detection in the past. For real-time processing, it is designed to be implemented on CNN-based real-time and parallel analog circuits to greatly increase the executive efficiency. The design of CNN circuits, however, adopts the architecture of multi-layer and 5×5 large neighborhood, and extends the size of array on this image processing to 32×32. In order to reduce the circuit complexity at the same time, the current-mode architecture is adopted and the direction of currents is extended to both positive and negative two-direction. Then the weighted ratio (the current amplify) of every neural cell is easily implemented, and the combination of many signals on nodes is easy. Because the CNN has the properties of array-type parallel processing and local connection of devices, it’s suitable for implementation on mix-signal and standard CMOS process.