[研究成果] 數位相片之人臉註釋與辨識系統

數位相片之人臉註釋與辨識系統

研究生
蔡承穎

指導教授
陳定宏

摘要

摘 要 輕巧的數位影像擷取設備如數位相機與具備照相功能的手機大量普及，使 得數位影像的產生十分便利。而相對而來所造成的龐大影像資料之管理、註釋 與檢索也變成研究的重要課題。 本篇論文中，我們針對相片中人物的註釋與辨識問題，以一種半自動的方 式提出一個初步可行的解決方法。在人臉偵測及人臉辨識的相關研究中，所需 加強的是準確率的提升；而輸入影像的前處理步驟及輸入影像的品質也都可能 明顯影響後續處理人臉偵測及辨識的準確性。在本論文中，我們首先尋找相片 中人臉部位，利用非線性的YCb'Cr'色彩空間轉換來找出膚色位置，再利用人 臉上的特徵來分析判斷此區塊是否為人臉部位。在此我們假設人們在拍攝一系 列的相片時，他們所穿的衣服都是不變的，因此衣服的顏色可以用來當作人臉 的一種特徵。確定人臉位置後，我們擷取臉部下方的衣服區塊的色彩值來當作 人物註釋與辨識的依據。為了不使衣服區塊因顏色眾多而影響辨識，我們利用 K-means 分群法分析此區塊，選出區塊中顏色比重最多者作為樣本特徵。之後 再以LVQ 演算法重複訓練，使衣服區塊之顏色樣群分佈更加明確。此時之衣服 顏色便可用來當作確認辨識人物的特徵。 我們實作出一個具有註釋能力的數位相片管理系統以驗證本論文所提方 法之可行性。系統的操作流程如下所述。首先系統自動定位人臉位置，使用者 可將人物名字或描述文字註釋到相片中的人臉位置。在連續拍攝的相片集中， 系統會依據之前相片所儲存的人臉衣服對應關係，偵測並辨識相片中的每一個 的人臉位置；並根據資料庫中的特徵，對於人臉加以自動註釋。若使用者發現 註釋錯誤，則可針對錯誤加以更新。更新後的資訊將會加入資料庫內，用以修 正人臉與衣服的對應關係。重複以上述方法，我們將可建構出人物影像特徵與 ii 人物說明資料庫，以提供大量相片所需的註釋與管理功能。實驗結果顯示，本 論文所採用之影像特徵與其他的註釋項目相互搭配後，可提供影像檢索與相片 管理作業中良好的註釋依據。 i

Abstract

The handy equipments such as digital cameras and cellular phones with camera function provide a convienient way to capture images. Therefore, the additional demands such as image management, image annotation, and image retrieval have become the major issues. In this thesis, we will focus ourself on the face annotation problem and propose a feasible solution. The major concern in face detection and recognition problems is the accuracy in recognizing the people in the photo with the exact identity. Many factors could influence the accuracy of face detection and face recognition. In this thesis, we propose a semi-automatic approach to solve these problems. First, we use a non-linear YCb'Cr' color space transformation to detect the location of skin color, and then we search the face positions by using the geometric properties of faces. Here we make an assumption that people generally wear the same clothes in taking a series of photos. While the faces are located, the RGB values of the clothes beneath the faces are used as the features to represent the faces. In order to simplify the color feature of clothes, we use K-means clustering to analyze this cloth block and select the major color as the cloth feature. The LVQ algorithm training process is repeated to make the color distributing between different cloth blocks more clear. We have developed a sytem to verify the proposed method. The sytem is performed as follows. First, the system automatically locates the face position and computes the corresponding cloth feature. User gives a name or text description to this face, the correspondence between the face and the annotation text has been established. The system will detect the face positions in a series of photos and give iv the detected faces the existent annotation text. If the annotation does not match to the detected faces, user can add a new annotation or correct the wrong matchings. The experiment shows the proposed method can achieve a good performance in face annotation and recognition. It also provides a convient tool to manage and retrieve the huge amount of digital photos.