於複雜背景及不同光影環境下之即時人臉偵測系統

Abstract

人臉偵測於近年來受到重視,並廣泛運用於各種領域,如:人臉身份辨識、人臉追蹤及以內容為主之影像檢索系統。此方面的研究,皆須偵測人臉並定位以進行後續的處理,因此如何精確並快速偵測人臉為相當重要之議題。本研究提出以梯度為主之即時人臉偵測系統,將偵測分為兩階段,第一階段以人臉及非人臉之梯度分佈高斯混合模型,並使用動態間隔偵測法,大幅降低需掃瞄之視窗數目,第二階段串連七個梯度空間相關性模型,進行人臉精確定位並有效移除誤判視窗,且保留人臉視窗。實驗證實,本研究所提出之梯度分佈特徵對臉部姿勢、表情、轉頭及傾斜有良好的強健性,並於複雜背景及光源變化等情況,仍可精確定位人臉,在實驗影像資料庫BioID及Viplab各達到91%及95%之偵測率,並維持極低的誤判視窗數目,且於Pentium M 1.5GHz之筆記型電腦上,每秒可處理10張320×240影像,亦滿足即時偵測之需求。
Human face detection is an important capabilities in a wide range of applications, such as face recognition, face tracking, and content-based image retrieve. Detecting and locating face in image is a necessary procedure before any future processing. We proposed a real-time face detection system including two gradient-based models. In first stage, two Gaussian mixtures of facial and non-facial weighted gradient distribution are used to roughly locate face in image. For accelerating detecting speed, dynamic interval detection algorithm is proposed to avoid redundant computations. In second stage, spatial gradient relation model is proposed to remove false detection and locate the facial positions precisely. In experimental results, weighted gradient distribution and spatial gradient relation model are proven to robust to different facial pose, expression, and rotation. Proposed methods can achieved detection rate of 91% and 95% respectively in database of BioID and Viplab under complex background and varying light condition. Proposed system can detect faces in 10 frames per second with size of 320×240 on a Intel Pentium M 1.5GHz notebook.

Description

Keywords

人臉偵測, 圖形識別, 高斯混合模型, 貝氏定理, 粗糙分類器, 梯度, 即時系統, 串連式架構, Face Detection, Pattern Recognition, Gaussian mixture, Bay's theorm, weak classifier, AdaBoost, Gradient, Real-Time System, Cascade, BioID

Citation

Collections

Endorsement

Review

Supplemented By

Referenced By