發展與應用具適應性視覺皮層之類神經網路於物件辨識

CCUR > College of Natural Science and Engineering > College of Engineering > Department of Computer Science and Information Engineering > project > Item 987654321/2170

Please use this identifier to cite or link to this item: https://irlib.pccu.edu.tw/handle/987654321/2170

Title:	發展與應用具適應性視覺皮層之類神經網路於物件辨識
Authors:	李志仁
Keywords:	蓋伯濾波器類神經網路視覺皮層指紋辨識虹膜辨識
Date:	2009
Issue Date:	2009-09-07 15:01:41 (UTC+8)
Abstract:	在物件辨識中，通常包含兩個主要的過程：特徵萃取和分類判斷，在大部分的文獻中這兩個部分通常是採分開設計。事實上我們可以擴展類神經網路的概念，增加特徵萃取層，使特徵萃取相關的參數也能自動學習，如此所使用的特徵萃取處理器的個數和計算得以最精簡。因為蓋伯濾波器(Gabor filters)是模擬人的視覺皮層組織細胞，它們除了能夠偵測輸入訊號的方向和頻率，並且這些資訊可以直接從灰階影像中萃取出來，直接簡化與加速整個萃取過程。因此我們提出主要蓋伯濾波器(principal Gabor filters)的概念，僅留下反應值大的幾個蓋伯濾波器。因為在影像合成的過程中，係數大的濾波器不僅可以代表影像的主要成分，也具有重要特徵的濾波效果。所以在此計畫中我們將利用自動學習的方式，找到較適合代表局部區域影像的主要蓋伯濾波器，同時達到濾波和降低維度的功能，以後就不用再浪費冗長的時間手動調整蓋伯濾波器的參數。也就是本計畫將提出一個結合特徵萃取和分類判斷的類神經網路，並且在特徵萃取層的視覺皮層細胞以可調整「參數」和「個數」的蓋伯濾波器來設計。在本計畫中物件辨識的實際應用，選擇我們最熟悉的生物特徵辨識，其中我們以指紋辨識、虹膜辨識為代表。因為指紋辨識的應用是只挑選一個蓋伯濾波器來代表局部的指紋紋路走向和間距，非常適合所提的類神經網路的初步應用。而進一步的應用是虹膜辨識，我們挑選數個主要蓋伯濾波器來捕捉局部虹膜影像的頻率和方向的變化。最後經過我們的研究後將發展一套跨生物特徵的類神經網路辨識系統 Generally, there are two main processes for object recognition: feature extraction and classifier. Most of researches designed them separately. In fact, the parameters and the number of feature extractors can be adaptively learned by using neural networks. Therefore, not only the number of feature extractors is simplified, but also the computation time is saved. Simulating the visual cortex of human, Gabor filters have the abilities to detect the orientation and spatial frequency of the input signals. And these properties can be extracted from gray-level images directly. In this research, we will propose principal Gabor filters not only to represent the principal components, but also to capture the main characteristics for a local image. So the input dimension is reduced and the key features are also extracted. In order to avoid wasting time to select and adjust adequate Gabor filters manually, we propose a novel neural network incorporating a feature extraction layer to simulate visual cortex and brain. And the proposed neural network has the ability to adjust the number and the parameters of Gabor filters. To verify the proposed neural network is feasible, we select biometric recognitions including fingerprint recognition and iris recognition. For fingerprint recognition, a local image is easily represented by a Gabor filter to capture its orientation and spatial frequency. Another further application is iris recognition. Some principal Gabor filters are necessary for a local iris image. At last, we will apply the proposed neural networks to the other biometric recognition systems.
Appears in Collections:	[Department of Computer Science and Information Engineering] project

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	448	View/Open

Loading...