چكيده به لاتين
Face recognition is an important topic in machine vision and pattern recognition, and is one of the most successful in biometrics that has recently been used by many researchers. Also, among feature extraction algorithms, texture descriptors have a good performance in image recognition, so in this study two texture descriptors have been proposed for face recognition.
Descriptors based on Local Binary Pattern (LBP) have good performance in image recognition. Improved versions of LBP, such as Center-Symmetric LBP (CS-LBP) and Local Ternary Pattern (LTP), are also successfully applied to image recognition but, it's hard to manually set a suitable threshold for them to address the noise. To overcome this problem, inspired by Weber's law, adaptive local feature descriptors have been proposed based on an automatic strategy selecting the threshold and are more robust against noise. The first proposed method is also derived from the automated threshold according to Weber's law and the Orthogonal-Symmetric LTP (OS-LTP) algorithm that is called Orthogonal-Symmetric Adaptive LTP (OS-ALTP). There are many methods in the evolution of texture descriptors. Patterns of Oriented Edge Magnitudes (POEM) is a strong descriptor that has good performance in face recognition. The improved version of the POEM, called Patterns of Dominant Orientations (PDO) which consider the relationships between dominant orientations of local image regions. Since the POEM and PDO are complementary strength, the second proposed method has been introduced by combination of the improved versions of these operators. The proposed method is called Patterns of Orientations and Magnitudes (POM).
For the first proposed method, the experiments on noisy ORL database show the high accuracy of the proposed method with an average recognition rate of 5.64% compared to the best adaptive algorithm with different number of gallery images. Also the results of the FERET database which includes high quality images, show the ability of the proposed algorithm to compete with the other mentioned algorithms. For the second proposed method, the experiments on the ORL database with 4.45% for single gallery image, the FERET database with 2.14% for single gallery image and 0.87% for the maximum possible gallery images, and the LFW database with 2.69%, show the better recognition rates of the algorithm compared to the best texture descriptor.
