多層次臉部細紋之分析與合成

Abstract

本論文提出一個可應用於產生人臉主要運動表情並保有細緻人臉細紋之影像多層次頻率分析與合成技術。我們使用對人臉的不同區域彼此間之相關係數作為人臉分群之準則，並利用分群出的人臉區域獨立進行表情的合成，藉以利用少量資料產生豐富變化之人臉表情。我們使用高頻資訊增強之控向金字塔分層法(steerable pyramid)對範例影像進行不同頻率成分之分解。在人臉高頻影像合成的過程，我們使用形態融合(blend shape)處理最低頻之資訊，並保留合適的高頻資訊，以作為後續標準差比對合成。藉由高低頻成分的獨立處理，我們的技術可以合成栩栩如生之人臉影像。將我們的系統與時間空間之相關性以及紋理對齊之技術結合，可呈現出具細緻紋理、豐富表情變化之人臉動畫。

This thesis presents a multi-layer analysis and synthesis approach to synthesize novel expressions with fine facial details and global features. We analyze correlation between different face regions as the criterion for face segmentation, and process each sub-region independently for generating various appearances. In synthesis procedure, we use the high-band enhancement steerable pyramid for decomposing various frequency components. The lowest sub-band is used for primary deformation by blend shape. For high-band, the statistic-based feature matching and high-band enhancement synthesis is applied. Due to our high-band\low-band separate procedures, we can synthesize photorealistic facial expressions. Integrate our approach with spatial-temporal constraints and texture alignment, we can generate detail-preserved facial animation.