Relevant Feature Analysis In Face Detection

Human face detection and recognition play an important role in plenty of applica-tions such as authentication, dossier management, visual communication, etc. The studyis roughly comprised of face detection, face tracking and face recognition. Face defection,as the first step in the analysis process, is to find out human faces in images or video se-quences rapidly and accurately. Currently, it is still a complicated problem to extract goodfeatures for face detection from the given images so as to improve the detection accuracy.In this thesis, 4 feature extraction methods including fourier transform, wavelets trans-form, independent component analysis(ICA) and sparse coding are employed to extract ef-fective facial features from gray-level still images. The main idea is to find out a group ofeffective bases functions which can capture facial structures firstly, and then face imagesand non-face images can be projected on these bases functions. Projection coefficients rep-resenting original images are used for classification. Since most frontal face images havesimilar structures, the coefficients as features on the same set of bases functions are veryclose. However, nature images not including faces have different coefficients on those basesfunctions. Comparing facial coefficients with nature coefficients helps to distinguish them.Among these four bases functions, fourier bases and wavelets bases are fixed ones, whereasICA bases and sparse bases are adaptively learned from the given face images as trainingsamples.A feature selection algorithm based on mutual information is also employed to select asubset of features relevant to objective from the original set. This process removes redundantand irrelevant features in order to improve the detection accuracy. Support vector machine istrained as a classifier with training data for classification. The final face detector is cascadedby 4 SVMs trained with different kinds of extracted features. Computer simulations showthat all these four extraction methods provide good features for classification. ICA and over-complete bases have better classification performances than the fixed ones because of theiradaption learning from training samples. The cascade classifier yields a good performanceunder some restrictions as well.