Spatial Scalability in Face Recognition using Laplacian Pyramid

The major challenge in face recognition is to deal with some problems such as pose, expression and lighting variations, occlusion, aging, and low image resolutions. On the other hand, image compression methods are generally used to reduce memory demand in large face databases. The present study undertakes to find a proper method that not only provides image compression, but also solves lighting and occlusion problems. It can also make a trade-off between speed and accuracy. In the proposed algorithm, input images pass through a Laplacian pyramid, in which they are divided into layers and then compressed. In the next step, DCT is performed on each layer to extract the important features from face image. Since DC is the first element of these transformations, its elimination may decrease the dependency of recognition on changes in lighting conditions. In order to reduce size, insignificant coefficients can be removed and PCA is utilized in the following stage. Finally, face recognition is performed by calculating Euclidean distance and finding the closest similarity. Therefore, the coarsest layer of the pyramid is used to gain low-accuracy and high-speed recognition. The upper layers are used to increase accuracy, at the price of losing speed.