| 摘要: | 在近幾年來，物件偵測與辨識在電腦視覺上的研究與應用相當的多。而其中人臉與車牌的偵測和辨識是相當常見的領域。首先在偵測方面，通常是先訓練一個分類器去偵測物體的所在位置。而常見的分類器有類神經網路、支持向量機…，但運算量通常很大。後來有學者提出以Adaboost訓練的串聯式分類器並利用Haar特徵結合積分影像去快速地過濾掉背景影像以達到即時的偵測，但卻需要相當長的訓練時間。
在本研究中，一個基於迴積類神經網路的特徵映像與兩階式串聯式分類器被提出來做物件的偵測。首先迴積與取樣運算可用來減緩物體受光線、旋轉及扭曲的干擾。然後接在後面的兩個分類器以粗至精細的機制去過濾掉大量的背景影像。由於輸入至粗階分類器的是取樣的特徵映像，其要檢查的視窗數縮小至原影像大小的四分之一。而剩下來的視窗再進一步由細階分類器來檢查。除了改善偵測流程，提出來的架構也大大地提升了訓練的速度。由較小的視窗所產生出來的少量特徵數被用來訓練一個粗階分類器。此外在訓練細階分類器時，特徵分級演算法從大量的特徵數只保留住少數有鑑別力的特徵並在不減少偵測效果的情況下以加快訓練的速度。最後的實驗，我們提出的偵測演算法和其它著名演算法比較後皆有較佳的結果。
在人臉辨識方面，我們在本研究中提出了正交化最近鄰居特徵線嵌入(Orthogonal Nearest Neighbour Feature Line Embedding，ONNFLE)。由於最近特徵線嵌入(Nearest Feature Line Embedding，NFLE)會具有內外插誤差，而且當樣本增加時會大幅增加計算量。因此為改良這些缺點，我們在產生最近特徵線時先選擇鄰近樣本點，再由這樣鄰近樣本點產生最近特徵線。如此則可以降低上述所提及的內外插誤差以及計算量增加的問題。在最後的實驗結果中，都能顯示出我們改良後的辨識演算法皆有顯著的效果。
Recently, the researches and applications about object detection and recognition grow rapidly in the area of computer vision. Among these, the detection and recognition of human face and license plate are typical applications. To achieve the detection goal, an object is firstly detected by a trained classifier. The commonly used classifiers are neural networks, support vector machines, etc. However, the computation load is very heavy. To remedy the drawback, researcher proposed a cascade classifier trained by Adaboost and combined Haar-like features with integral images to quickly filter out background regions to achieve real-time detection task. However, it is still time consuming in training the classifiers.
In this dissertation, an object detector is proposed based on a convolution/sub-sampling feature map and a two-level cascade classifier. First, a convolution/subsampling operation is proposed to alleviate the suffering of the illumination, rotation, and distortion variances. Then, two classifiers are concatenated to check a large number of windows using a coarse-to-fine strategy. Since the sub-sampled feature map with enhanced pixels is fed into the coarse-level classifier, the size of the feature map is drastically reduced to a quarter of the original image. A few surviving windows with detailed data are further checked with the fine-level classifier. In addition to improving the detection process, the proposed mechanism also speeds up the training process. A few features generated from the prototypes within the small window are selected and trained to obtain the coarse-level classifier. Moreover, a feature ranking algorithm is proposed to reduce the huge feature pool to a small set for speeding up the training process without losing the generality of the feature pool. Finally, some experiments were conducted to show the feasibility of the proposed method.
As to the recognition, a novel manifold learning algorithm, called orthogonal nearest neighbour feature line embedding (ONNFLE), for face recognition is also proposed. In the proposed ONNFLE, two drawbacks of our earlier proposed nearest feature line embedding (NFLE) method are resolved. They are the extrapolation/interpolation error, and high computational load. The extrapolation error occurs if the distance from a specified point to one line is small when that line passes through two farther points. The scatter matrix generated by the invalid discriminant vectors does not efficiently preserve the locally topological structure which results in incorrect selection while reducing recognition performance. To remedy this problem, the nearest neighbour (NN) selection strategy is used in the proposed method. In addition, the high computational load is also reduced using a selection strategy. Finally, some experiments were conducted to demonstrate the effectiveness of the proposed algorithm. |
