| Summary: | 碩士 === 國立高雄應用科技大學 === 光電與通訊工程研究所 === 102 === Extracting facial feature is a key step in face recognition (FR). Inaccurate feature extraction very often results in erroneous categorizing of persons. Especially in extremely condition, illumination variation is a crucial issue in FR and images which are not properly corrected can look either bleached out or too dark and eventually introduces a false acceptance. In this thesis, we present a novel framework, local color correction in extremely illumination condition (LCC-EIC), working toward to reproduce facial colors accurately. Adaptively varying the amount of gamma correction based on the histogram of the image changes on the brightness, and thus solving the problem in which the contour disappeared from the face under the shadowed side. To remove unrealistic color casts, so that faces which appear white in person are rendered white in our images. Smoothing white balance (SWB) is followed up. The original mean of RGB three color channels to the mean of any two channels which nearest the comfort intensity zone is adaptively tuned up making the white balance would not have overexposed and loss of color saturation. Finally, we propose three objective indexes for image coherence to measure the performance of the proposed algorithm. The indexes are the average intensity of the image, the mean of the standard deviation on individual intensity, and the mean of the standard deviation of the ratio of RGB three color channels. Using these three indexes to evaluate the invariance of intensity and skin color under different lighting conditions can be regarded as almost equal to make a questionnaire for people. We use CMU-PIE face database and compare with our previous works. The result shows that the proposed algorithm makes the great improvement on image quality based on human eyes and our three image coherence indexes. In final, we combine our result with the related work of Gradientfaces for facial feature extraction, then using Eigenface and the nearest neighbor for FR. Experimental results show that this work leads 15% recognition rate compared to the related works with first four Eigenfaces. It means that our method not only satisfies the human vision, but also makes the better job on machine vision.
|
|---|
