Face Detection Based On Spectral Histograms And SVMs

As a critical technology of face information processing, face detection has been paid much attention and become a very active research branch in pattern recognition and computer vision application areas. With the development of intelligent information processing, face detection will be broadly applied in identity recognition, content-based retrieval, surveillance and human computer interaction.In this thesis, the definition and properties of spectral histogram representation is given, and a face detection algorithm based on spectral histograms and support vector machines (SVM) has been studied. More importantly, the performance of algorithm is improved.(1) In this thesis, spectral histograms and SVMs are combined effectively, and fairly good results to detect frontal faces in complex images are achieved. The performance of the algorithm can be attributed to the desirable properties of spectral histogram representations and the generalization property of SVMs. The proposed algorithm has strong adaptation and robustness.(2) The performance of algorithm is improved, by computing spectral histograms which LBP operator used to replace Gabor filter to construct filters. Local binary patterns (LBP) describes the structure of the texture by charactering changes of gray of the neighbourhood of each pixel; it is efficient not only for making SVMs classification more accurately but also for reducing the computational cost.(3) To detect faces with different sizes, a strategy of multi-resolution sliding window is used to accomplish invariance of location and scale in face detection.Experiment results show that the algorithm can detect frontal faces in complex images; then it is a robust and effective algorithm.