The Study of Noisy Face Image Recognition

• Main Authors: Yung-Chang Huang, 黃永昌
• Other Authors: 陳財源
• Format: Others
• Language: zh-TW
• Published: 2009
• Online Access: http://ndltd.ncl.edu.tw/handle/y9ajk5

碩士 === 崑山科技大學 === 數位生活科技研究所 === 97 === The more safety and friend living environment are needed for recent years. A higher speed and more accuracy automatic face recognition are required in many applications. In fact, the recognition rate of automatic face recognition will reduce as face images which are contaminated by different kind of noise. In the thesis, we primary discussed the noisy image recognition process from image enhancement and face feature extraction to noisy face recognition. First in the image enhancement, the median filter and discrete cosine transform are concerned to reduce noisy effect and enhance the image. Then, the principle component analysis (PCA) and the independent component analysis (ICA) for facial feature extraction are consulted. The PCA use second-order statistics as feature and the ICA considers the higher order statistics to get sparser image representation. For more improving recognition rate, the optimal feature subspace is selected by using the genetic algorithm (GA). Finally in the face recognition, the nearest neighbor (NN) and the support vector machine (SVM) are investigated since the NN is one of the simplest and the SVM is one of the most accuracy recognition methods. In the experiment, the noisy image is obtained by adding a series of different levels of Gaussian noise or salt-and-pepper noise into images in the ORL database. The experiments are performed on the noisy image by using a variety of recognition systems that are combinations of different image enhance methods, feature extraction methods, and recognition methods discussed above. The experimental results show that the recognition can be improved by using the optimal feature sub-space selected by genetic algorithm and image enhancement can be achieved by using suitable filter. The facial feature extract from the ICA have no benefits for our noisy face image recognition. The experimental results also demonstrate the fact that the recognition results obtained by the recognition system combined by the 2DPCA and the SVM are not just accuracy but efficiency.