A Study of Supervised Neural Networks in Semantic Image Content Analysis

• Main Authors: Hung-Jen Wang, 王宏仁
• Other Authors: Chuan-Yu Chang
• Format: Others
• Language: en_US
• Published: 2011
• Online Access: http://ndltd.ncl.edu.tw/handle/80510316682733914796

博士 === 國立雲林科技大學 === 工程科技研究所博士班 === 99 === With rapidly development of digital cameras, digital image processing, database, and internet technologies, image classification has become an important topic in multimedia processing. Therefore, how to efficiently manage a large amount of digital images in content-based image retrieval systems becomes an important issue. Analyzing the contents of an image and retrieving corresponding semantics are important in semantic-based image retrieval system. Recently, neural network-based methods have been proposed to solve the classification problem. Among them, there are some popular supervised neural networks such as support vector machine (SVM), radial basis function (RBF), and so on. In this thesis, we proposed three supervised neural networks including SVM, modular RBF (MRBF), and two-phase fuzzy adaptive resonance theory (Fuzzy-ART) neural networks the solve the gap between the high-level semantic and low-level features in semantic-based content analysis. The radial basis function (RBF) neural network is a most popular architecture because it has good learning and approximate capacity. However, the traditional RBF neural networks were sensitive to center initialization. In order to obtain appropriate centers, it needs to find significant features for further RBF clustering. In addition, the training procedure of the traditional RBF network is time-consuming. Therefore, in this thesis, a combination of self-organizing map (SOM) neural network and learning vector quantization (LVQ) is proposed to select more appropriate centers for RBF network, and a modular RBF (MRBF) neural network is proposed to improve the classification accuracy and speed up the training time. In general, Fuzzy-ART is an unsupervised learning clustering and pattern recognition network. Since, the clustering results of an unsupervised learning are uncertainly, the Fuzzy-ART is modified to supervised two phase Fuzzy-ART to improve the classification accuracy. Traditional region-based retrieval systems attempt to reduce the gap between high-level semantic and low-level features by representing images at the object level. However, in feature extraction, it is difficult to obtain adequate object feature because it is relative to segmentation method. In this thesis, we combine both segmentation method and user interaction to obtain adaptive object features for further semantic-based image analysis and overcome these problems mentioned above using those three proposed neural networks. Moreover, we apply the principal component analysis (PCA) to extract significant image features and then incorporated them with the proposed three neural networks for image content classification. Finally, based on the content analysis by the proposed two-phase Fuzzy-ART, each region in an image is associated with a high-level semantic concept. The proposed system supports both query by keyword and query by specified region(s). Experimental results show that the proposed method has a high accuracy for semantic-based image content analysis, and the result of image content analysis show a reasonable classification results in human visual perception. Furthermore, the semantic-based image retrieval system has high retrieval rate.