An efficient fusion algorithm for large scale face verification based on KISSME and cosine similarity

Although the KISSME approach can effectively reduce the correlation between feature vectors of samples, it can´t restrain the multimodal distribution of vectors in the global level. This drawback brings negative impact on classification performance. Inspired by KISSME, we propose our fusion algorithm of Likelihood Ratio Test and cosine similarity for large scale face verification (CS-KISSME). In our algorithm, we obtain an approximate optimum Mahalanobis distance matrix and use it as a projection matrix. After that, we measure the dissimilarities by cosine similarity in the linear transformed subspace. In this paper, we expound merits of CS-KISSME theoretically, test it on two challenging face verification datasets and achieve higher accuracy and better robustness with little time consumption.