本文提出基於人臉偵測之Haar-like型特徵樹狀結構分類器，訓練加速技巧與分類器放大修正法。在Adaboost的訓練法中，正樣本資料為非理想形狀構成時，將使得演算法使用更多的分類器，且效率不彰，樹狀結構為改善此點所設計。訓練資料的差異會造成不同的分佈值，對被挑選出來的分類器而言，分類器值的高低可以反映分佈值。利用具有分支的節點使得不同分佈，不同分類器值的訓練資料分離，使具有較大差異的資料能夠分別訓練，可減少無效率的分類器產生。訓練出的樹狀分類器可將偵測結果作無預先知識的分類，可供後續處理使用。訓練加速技巧使得訓練的時間複雜度有效地降低。於實作過程中發現，分類器在偵測大圖時的放大過程中，因非整數放大倍率造成分類器具有實數值的形狀，無法對應取得圖片上只有整數位置的像素值，而造成誤差，導致判斷結果不良，故提出分類器放大修正法,可大幅改善Haar-like型特徵放大的盲點。偵測法的改善使得訓練所大量需求的負樣本取得變得容易，有益於訓練出具有普遍性的分類器。 This thesis present the tree structure detector of Haar-like feature for face detector, a training speedup tick, and the feature scale correct method. In the feature training of Adaboost algorithm, the process adapt much more features for complex positive data, and the additional features are not efficient to the final result. We propose the tree structure to improve the data path. The distributions of trained distinct data are separated as high and low, and the difference can be also observed on the value of stage detector. Tree nodes are used to separate data by their detector values. In next stage, the data with difference will be trained independently. It decreases the production of ineffective features, and its data path can be a support of intra data classify. In my experiments, I found a scaled haar-like feature can’t fetch a pixel with float point position, and the results of detection are worse to the non-scaled feature. The scaled feature need a proper approximation. The Scaled Feature Correction method we proposed improve the detect rate of scaled feature. It also let the gather of negative samples become more easily when we want to train a classifier with generalization.
