智慧型多特徵人臉偵測系統設計

Abstract

本篇論文針對人臉偵測提出一個智慧型多特徵偵測系統，以達到提高人臉偵測率的目的。此系統可主動在串列影像中找出人類視覺上判斷人臉的兩個重要特徵，即膚色及輪廓，其中膚色判斷的結果可先視為可能的人臉區域。在本篇論文中，與一般人臉偵測系統不同之處，主要在於膚色擷取、輪廓擷取與人臉辨別等三方面，皆是經由監督式學習的類神經網路設計完成，以達到人臉偵測系統智慧化的目的。此外針對人臉辨別，本論文將可能的人臉區域結合膚色及輪廓特徵，先取得區域膚色影像及區域輪廓影像再做不同的影像處理，包括影像聯集、水平迴旋積、及影像相乘三種方式，最後經由類神經網路辦別是否為人臉。由實驗結果可知多特徵的人臉辨別率可優於單一特徵的類神經網路架構，其中又以影像聯集為最優，可提升辨別率約3%。

This thesis proposes an intelligent face detection system to find out faces from a sequence of images based on the skin color and edge, two important features in human vision for face detection. The system first searches the potential face regions from an image via the skin color. Different to the conventional face detection, the skin color extraction, edge extraction, and face classification are all implemented by neural networks under supervised learning. Moreover, with the use of skin feature image and edge feature image there are three combined methods developed for face classification, named as Union-NN, Convolution-NN and Multiply-NN. From the experiment results, it is clear that neural networks of multi-feature could improve the recognition rate when compared with single-feature. Besides, the Union-NN has the best recognition rate in face detection.