LBP based on multi wavelet sub-bands feature extraction used for face recognition

The strategy of extracting discriminant features from a face image is immensely important to accurate face recognition. This paper proposes a feature extraction algorithm based on wavelets and local binary patterns (LBPs). The proposed method decomposes a face image into multiple sub-bands of frequencies using wavelet transform. Each sub-band in the wavelet domain is divided into non-overlapping sub-regions. Then LBP histograms based on the traditional 8-neighbour sampling points are extracted from the approximation sub-band, whilst 4-neighbour sampling points are used to extract LBPHs from detail sub-bands. Finally, all LBPHs are concatenated into a single feature histogram to effectively represent the face image. Euclidean distance is used to measure the similarity of different feature histograms and the final recognition is performed by the nearest-neighbour classifier. The above strategy was tested on two publicly available face databases (Yale and ORL) using different scenarios and different combination of sub-bands. Results show that the proposed method outperforms the traditional LBP based features.