| Summary: | 碩士 === 國立高雄第一科技大學 === 電腦與通訊工程所 === 91 === This thesis is based on general family personal identification system, designed for face recognition system by simple and low cost equipments.
The experiment of this study uses the gray level image with low resolution to extract the facial image by pixels standard fuzzy facial pattern detecting after noise filter. Because we take the LMSE, Least Mean Square Error, method to detect and extract the facial image, it will make some facial detecting wrong when the average gray of image background is equal to the standard fuzzy facial pattern, or the clothes of figure is more complexity. Therefore, we use the outline of human’s head to erase the figure background and clothes to avoid detecting wrong after noise filter.
After getting the facial image, we extract the outline of facial figures by Laplacian operation. As a result, that fringe of hair always cover the eyebrows, women figure especially, we cut the outline of eyebrows again. So the facial image remains the region surrounded by eyes, nose and mouth. Then we horizontally adjust and normalize the facial image to eighty square pixels.
Finishing the normalization, we extract the sixty-four features by the distance between facial figures and pixels density of grid model. Finally, we input the sixth-four features to the plastic perceptron back propagation neural network for training. Then the trained neural networks of facial recognition are obtained.
There are thirty persons to participate in this experiment. Every person provides at least three images with different facial expression for neural network training. The successful recognition rate achieves 100% for inside testing, and 90.77 % for outside testing. Therefore, we have a satisfactory achievement.
|
|---|
