| 摘要: | Adaboost最廣泛是用於人臉偵測系統。長久以來的研究，是以人臉影像特徵的固定整數權重為特徵訓練人臉來進行偵測，而Adaboost挑選出來的人臉影像特徵是否能保證為最佳的特徵，可否能再提高偵測的準確率，而false positive rate是否能再快速的下降，因此需要開始思考，當人臉影像偵測已經發展到一個階段時，如何突破現有的瓶頸，朝向更好的方法邁進。 本篇論文提出一種利用新的訓練方法來改進Adaboost分類器的偵測效果。目的是希望降低原始的Adaboost偵測所產生的false alarm，由於Adaboost的特徵權重為固定的整數權重，但實驗結果顯示固定的權重並非是最佳挑選出欲偵測的特徵，因此我們推翻以往的訓練方法，利用LDA的方式和PCA來調整訓練每一種特徵最佳的權重。除了利用LDA和PCA特徵之外，並加入整合NDA，來達到降低更多false alarm偵測效果，並且在訓練階段可以快速的下降false positive rate，以達到我們論文的偵測目的。 在速度方面我們也把原始浮點數的系統訓練架構[9]做一個更快速的改進，也就是由原始Adaboost訓練好的特徵來進行這些演算法的訓練調整權重組合，透過平行的處理每一個挑選出來的特徵，並利用投票法來挑選最後的特徵，來達到更好的人臉偵測率。 關鍵字: Face detection, AdaBoost, FLDA, PCA, NDA.Adaboost is the most widely used in face detection system. As reported in several literatures, face detection is usually carried out based on face image feature with a fixed integer weight through training. Whether the selected features from human face samples guarantee to be the best features or they can improve the potential correct rate while decreasing the false positive rate quickly is a topic worth investigating. Moreover, once the face detection techniques have been developed to a certain degree, the breaking of current bottlenecks toward a profound standard is still a long way to go. In this thesis, we present a novel training method to improve Adaboost classifier for enhancing the detection performance. In our proposed method, it will reduce more false alarm than traditional Adaboost because the original features in Viola’s Adaboost used fixed integer weight features. The results show that the using of floating point weight will indeed select better features for detection. We abandon traditional training methods by using LDA and PCA to adjust each weight for obtaining the best feature. In addition to using LDA and PCA features, NDA feature is also integrated to decrease false alarm. In the training stage, we can also quickly decrease the false positive rate while achieving better detection results. We then train the system architecture to make profound improvement for original floating point feature weight [9] in respect to speed. That is features are trained from the original Adaboost, then be fed to the second training process using LDA,PCA and NDA algorithms to adjust the feature weight combination as parallel processing for each selected feature. Finally, voting mechanism is adopted to select the last feature to achieve the best face detection rate. Keyword : Face detection, AdaBoost, FLDA, PCA, NDA. |
