多角度彩色人臉偵測

Abstract

人臉偵測在很多應用中扮演著重要的角色，越來越多的研究及方法再解決人 臉偵測的問題。人臉偵測的最終結果即是自動分析一張輸入影像是否含有人臉的 存在，如果有的話即輸出人臉所在的座標，標示出人臉的位置。 近年來，許多研究嘗試改進經典的人臉偵測方法”Viola&Jones (Adaboost)”使這 個原始只能偵測正臉的演算法可以偵測多角度人臉，到目前為止，多角度的偵測 是一個還未被完美解決的問題，在這篇文章當中，我們提出了一個基於 Adaboost 正臉偵測器但卻可以偵測多角度人臉的方法，在不改變經典架構的前提下使其能 解決多角度的問題。在此方法中包含了以下的技術，膚色切割、超像素使我們得 到可能為人臉的候選區域，藉由角度補償及精確修正可以將被旋轉過的人臉轉回 正面，至於側臉的部分使用對稱延伸的方法，將一個側臉左右顛倒後和原始側臉 結合，創造、延伸出一個相似於真正人臉的正臉，經由這些處理後，多角度的人 臉都可以使用單一一個 Adaboost 的正臉偵測器來做偵測，不需要訓練別的角度偵 測器或是改變 Adaboost 的演算法就可以有相等甚至更好的效果，使 Adaboost 有更 完好的效能。

Face detection plays an important role in many computer vision applications and has drawn significant research attention nowadays. The objective of face detection is to analyze whether the image contains face or not, and if it does, output the location of the bounding box for each face. In recent years, many researches attempt to extend the well-established Viola & Jones (Adaboost) face detection algorithm to suitable for multi-view face detection. Until now, it is a challenge to detect in-plane, rotated, and out-of-plane face simultaneously. In this thesis, a very robust multi-view face detection algorithm is proposed. Although it is essentially a frontal face detector, it can well detect rotated, in-plane, and out-of-plane face without rotated training faces. First, several techniques, including the skin filter and entropy rate superpixel (ERS) are applied to obtain face candidate regions. Then, angle compensation and refinement are applied to improve the accuracy of face detection in in-plane case. Moreover, to find the out-of-plane face, one can apply the symmetry extension technique, i.e., extending the face candidate with its flipping version to create a face that is similar to the frontal one. With it, even if there are no training data for out-of-plane face, one can successfully detect the face in the out-of-plane case. Simulations on the FEI, Pointing'04, Bao, Group, Utrecht, and our dataset show that the proposed algorithm is effective and outperforms state-of-the-art face detection approaches.