| Summary: | 碩士 === 長庚大學 === 電機工程研究所 === 90 === Feature extraction has been extensively applied to recognition of human face and fingerprint. In this work, it is different from the conventional analysis of single profile and we propose a few outlines of human face to extract features. To characterize the human face, we apply wavelet analysis to select some useful points on these curves, and construct the facial features with these characteristic curves.
In the case of recognition mechanism, we compare the results of face recognition on many kinds of neural network and obtain the feasible framework of the 3-D human-face recognition system. Original 3-D data are rotated in different direction within certain range of degrees, so that we can simulate the perturbation error of the same tester on the different attitude during measuring time.
Based on the combination of both original data and simulated data as described previously, we can set up the mixed training group for the basis of discipline and obtain the high accuracy of neural network. Under the sufficient training group and well-disciplined neural network operation, we can obtain 100% accurate recognition of front face through the searching conditions on near 30 groups. Even though the worst case on the facial data at 10 degrees of left and right rotations, we still provide the accuracy of 82 percents.
In the future, efforts on the reduction of groups within 10 groups are investigated. Our goal purposes to improve the recognition accuracy after data rotation and to reduce the number of characteristic data.
|
|---|
