人 臉 影 像 壓 縮 之 研 究

Abstract

本論文提出一套人臉影像壓縮方法，基於人類對於影像所付予的注意力會根據一些特徵(或性質)而有所不同，我們將人臉影像中的特徵區塊與臉部其它部份分別給于不同程度的壓縮條件進行壓縮，期望能達到高壓縮倍率並保持高品質的可辨識率。 為了達到這個目標方法主要分成下列兩個部份 : 臉部特徵的擷取和壓縮。人臉特徵擷取的部份主要是利用人類臉部特徵的左右對稱特性，取出中心對稱線。再利用頭髮、眼睛、嘴巴影像灰階值變化的不同特性定出特徵點初略的位置，得知眼睛、嘴巴的位置後接著就框出眼睛、嘴巴等臉部特徵所在的範圍。利用眼睛、鼻子和嘴巴的相對關係，我們大略地框出鼻子所在的範圍。經過第一階段後﹐我們將得到眼睛、鼻子、嘴巴三個人臉特徵區塊。 在壓縮的部份﹐我們將人臉影像轉換到wavelet domain，利用區域性與多重解析度的特性，對於不同區塊給予不同的quantization scale 和 decomposition level 分別進行壓縮。 最後的實驗證明，我們所提出的人臉壓縮方法確實能達到高壓縮倍率並保持高品質的可辨識率。

A face image compression system is proposed in the thesis. The system mainly consists of two phases: facial feature extraction and compression process. In facial feature extraction, based on the symmetric property of a face, the vertical central line of a face is found and the locations of the eyes and mouth are extracted according to the characteristics of the gray values of facial features. Furthermore, we use the relative positions between the eyes and nose regions to bound the nose region roughly. In the second phase, we use a fact, that human usually pays more attention in facial features for face recognition, to do compression in and expect to improve the compression rate. Wavelet transformation is used. By the local property of wavelet transformation, different compression levels are assigned to the facial feature regions, e.g., eyes, nose and mouth, according to the attention degree. Experimental results show that the proposed method has high compression ratio and still keep the face image in a recognizable quality.