以十字旋轉搜尋法為基礎之人臉自動標點之研究

The Cross-rotation Search Approach Based on Face Auto-labeling

一般人臉辨識是採用主動式外形模型(Active Shape Model, ASM)的方法。由於目前傳統的主動式外形模型需要手動標示特徵點，過程中十分耗時費力並且非符合即時性，因此，本文將其改良為全自動化人臉辨識糸統。在自動偵測方面，採用哈爾特徵(Haar-like Features)的人臉偵測方法，主要是它在物件偵測技術的速度上具有優越性，可快速比對出其特徵，而提升偵測率。在自動標點方面，採用索貝爾邊緣檢測(Sobel Edge Detection)技術與自創新的十字旋轉搜尋法(Cross-rotation Search Approach)，並且運用兩眼中心定位法尋找出人臉的大概位置，接著計算出偏移量加以調整，持續進行迭代運算直到收斂為止。經實驗證明，其自動標點所花時間約0.016秒，且整體辨識率可達75%至87.5%。顯然地，可以節省人工標點的時間成本。

Traditional active shape model (ASM) is manually labeled feature points. This process is very time-consuming and non-real-time; therefore, this paper improves active shape model for automatic punctuation of face recognition systems. In automatic detection, we adopted Hal features (Haar-like Features) face detection method. The main advantage is that it has the technology to detect the speed of an object, which can quickly compare the features and enhance the detection rate. In automatic labeling points, an efficient technique is developed to incorporate the Sobel edge detection method and the new cross-rotation search approach. This technique is the use of eyes center position to find out the approximate location of the human face, then to calculate and adjust the offset, and continuous iteration until convergence. The experiment proved that the automatic punctuation time spent approximately 0.016 seconds, and the recognition rate was 75 to 87.5%. Obviously, we can save time cost f artificial labeling points.