Research Of Correlation Filters And Its Application For Face Recognition

The modern city is becoming more intelligent, due to the fast development of society. A fast and accurate identification technology is of great importance in the modern city. The face recognition technology is becoming more popular than the traditional identification methods, due to its convenience and efficiency. Compared with other biometric technologies, including fingerprint recognition, iris recognition, the face recognition technology has several unique advantages, such as non-contact and intuitive. Although the face recognition technology has been widely used in the field of public security and digital entertainment, there are still several challenging problems to be solved, such as occlusions, pose variations, illumination variations, expression variations, etc. To deal with the key issues, including head pose estimation and feature fusion, in face recognition, we have intensively investigated the related work and proposed several solutions. The innovation points of our work are summarized as follows:●We propose a new correlation filter, called Principal Optimal Tradeoff Filter (POTF). POTF is a correlation filter which is designed in the1D frequency domain of the subspace obtained by PCA. Compared with the traditional correlation filters, POTF requires less computational complexity and is more robust to deal with illumination and expression variations.●We propose a robust head pose estimation method, called Directional Correlation Filter Set (DCFS). DCFS is composed of different directional correlation filters (DCF) corresponding to different poses. Each DCF consists of several POTFs designed by the samples corresponding to a specific pose. Experimental results on the PIE, HPI and UMIST face databases have shown that the proposed method can achieve the detection rate of more than99%, and can effectively improve the performance of face recognition. Furthermore, we extend the proposed method to occlusion detection. Experimental results on the AR database have demonstrated that the detection rate can achieve100%.●We proposed a nonlinear feature fusion method, called Non-linear Feature Fusion Based on Polynomial Correlation Filter (NF-PCF). We show that combining the global features and local features can effectively improve the robustness of the face features. High-order statistics information is beneficial for face recognition due to the fact that the facial structure is highly nonlinear. Experimental results on Yale, PIE, GTFD and LFW face databases have shown that NF-PCF achieved a better performance than the state-of-art methods, such as LPP, etc. Moreover, NF-PCF has a similar performance to the SRC method, but NF-PCF is much faster than the SRC method.