Safe binary particle swam algorithm for an enhanced unsupervised label refinement in automatic face annotation

Abstract

Mining web facial images on the internet has become as a profitable and important paradigm towards auto face annotation technique. The unsupervised label refinement (ULR) is an effective method that can fix weakly labeled facial images data which are collected from the internet and included some images with wrong label. In order to improve the correction accuracy of ULR, particle swarm optimization (PSO) and binary particle swarm optimization (BPSO) are used for solving binary constraint optimization task in this study. A novel method named safe binary particle swam optimization (SBPSO) is also proposed to improve BPSO which has the probability over range problem for using the ULR. In addition, SBPSO is also employed for an enhanced ULR (EULR) objective function which is created by modifying the original formula of ULR to improve the accuracy of labeled facial image. An experimental database is queried from IMDb website which collected the actors who were bored in 1950 to 1990. Some error flags are randomly added in the database for the correction tests by different methods. The results showed that the SBPSO Algorithm for the EULR in automatic face annotation have the better label correction rate and convergence effect.