基於細胞神經網路的仿生型電腦視覺系統

Abstract

自從 Dr. Ramon y Cajal 開始了對視覺神經細胞的研究以來，包含了眼球與大腦的人類視覺系統 （Human Visual System） 就成了一個有趣的研究主題。許多學者們也已提出了他們的研究成果並指出了一些新的研究議題。至目前為止，人類視覺系統與其生理結構已經大概被了解了。一些神經學家也已經開始分析人類視覺系統在神經學上的機制。然而，由於人類視覺系統相當龐大且複雜。至目前為止，我們仍然尚無法完全地了解、分析、甚至是模擬它。

視覺資訊 (Visual Information) 的旅程始於視網膜。過去數十年來人們相信視網膜的作用就如同照相機的感光底片。實際上，視網膜的功能遠超過我們已經知道的。根據 Werblin 與 Roska 在 2005 年的研究，視網膜提供了相當巨量的視覺前處理 (Preliminary Visual Processing)。在視網膜中，視覺資訊首先被光感受體 (Photo-receptor) 所接收。接下來經過一連串的視覺細胞的協同作用後，處理過的視覺資訊被分成超過 12 個不同的頻道並同時透過視神經節即時地傳送到視覺大腦皮質層作更進一步的處理。

延續這些學者的研究成果，本論文提出了一個基於細胞神經網路 (Cellular Neural Networks) 的仿生型電腦視覺系統 (Bionic Computer Visual System)。本系統包含了兩個模型，分別是視網膜中央小窩電腦計算模型 （Computer Fovea Model），與視覺大腦皮質層電腦計算模型 (Computer Cortex Model)。這兩個模型均是基於人類視覺系統而開發。其中大部份並以細胞神經網路加以實現。

為了將所提出的模型以細胞神經網路加以實現，本論文亦提出了一些使用細胞神經網路處理影像的新方法。為了達到與生物體結構一致，本論文提出了一個以六角形狀排列的細胞神經網路 (Hexagonal-type Cellular Neural Network)。同時亦提出將在本研究中被應用的數種基於細胞神經網路的穩態中央線性系統 (Stable Central Linear System of Cellular Neural Networks) 的演算法； 包括了 CNN-based Laplace-type Filtering 、CNN-based Gaussian-type Filtering、與 CNN-based Gabor-type Filtering。這些演算法將成為所提出的視網膜中央小窩電腦計算模型與視覺大腦皮質層電腦計算模型的基本元件。

視網膜中央小窩電腦計算模型可用於模擬某些視網膜的生理結構。透過研究以及模擬這些生理結構，一些人類視覺系統的特性可以被模仿出來。這些特性形成了天然的影像強化機制。因此可以使用所提出的架構來重現這些影像強化機制。這些機制包括了色彩統一性處理 (Color Constancy) 與影像銳化處理 (Image Sharpness)。

視覺大腦皮質層電腦計算模型可用於模擬某些處理視覺的腦皮質區。其提供了簡單的影像紋理辦識及聯想的機制。透過研究腦部視覺皮質區的作用，部份人類視覺系統處理影像紋理的機制也可以被模擬出來。這些機制包含了影像紋理隔離 (Segregration) 、分類 (Classification)、以及辦識 (Identification)。

透過了解這兩個電腦計算模型，部份人類視覺系統中有意義的功能可以被研究、學習、甚至是被模仿。因此，本研究可以提供人機界面領域 (Human-Computer-Interaction) 的研究一個新的方向。

A Human Visual System, including the eyes and the brain, is always an interesting research topic. Since Dr. Ramon y Cajal initialed the study of the visual neurons, many researchers introduced their research results and indicated more and more issues. Currently, the Human Visual System and its biological structure has been roughtly understood. Some neurologists also started to analyze the neurological mechanisms of the Human Visual System. However, the Human Visual System is an extremely huge and complex system; we are not able to fully analyze, to understand, or even to simulate it, yet.

The journey of visual information is started at a retina. For dacades, people believe that the retina is similar to the silver pieces in the film of a camera that react to light. In fact, the functions of the retina are much more than what we already knew. According to Werblin and Roska (2005), the retina provides a huge number of preliminary visual processings. In the retina, first, the visual information has been accepted by the photo-receptors. After the co-operations of some visual neurons have been finished, the visual information is separated into more than 12 channels and are transmitted to the visual cortex of the brain via some neural axises for further processing in real-time.

Following the works of those researchers, in this thesis, a Bionic Computer Visual System based on Cellular Neural Networks (CNNs) has been introduced. This system includes two major parts: a Computer Fovea Model and a Computer Cortex Model. Those two models are based on the biological structures of the Human Visual System and they are implemented based on the Cellular Neural Networks.

For the sake of implementing the proposed models on the Cellular Neural Networks, some advanced approaches for CNN-based image processing are also introducted. In order to consist with the biological structures, a hexagonal-type Cellular Neural Network is introduced. Meanwhile, some operators, which are based on the Stable Central Linear System of the Cellular Neural Networks, are also presented and are used in this thesis. Those operators include CNN-based Gaussian-type filtering, CNN-based Laplace-type filtering and CNN-based Gabor-type filtering. They became the fumdental elements of the proposed Computer Fovea Model and the Computer Cortex Model.

The Computer Fovea Model can be used to simulate the preliminary processing of a retina. Through investigating this model, various properties of the Human Visual System can be simulated. The Human Visual System possesses numerous interesting properties, which provide the natural methods of enhancing visual information. Various visual information enhancing algorithms can be developed using these properties and this model. The proposed algorithms include color constancy, image sharpness.

The Computer Cortex Model provides simple texture recognition and association for the textures on an image. Through investigating the behaviors of this model, some properties of the texture analysis of the Human Visual System also can be simulated, including texture segregation, classification, and identification.

Based on the studies of those two models, some significant functions of the Human Visual System can be investigated, understood, even simulated. Thus, this study can provide a new method for the field of Human-Computer-Interaction.