The identification of the most discriminative features in an explainable AI decision-making process is a challenging problem. This research tackles such challenges by proposing Particle Swarm Optimization (PSO) variants embedded with novel mutation and sampling iteration operations for feature selection in object recognition. Specifically, five PSO variants integrating different mutation and sampling strategies have been proposed to select the most discriminative feature subsets for the classification of different objects. A mutation strategy is firstly proposed by randomly flipping the particle positions in some dimensions to generate new feature interactions. Moreover, instead of embarking the position updating evolution in PSO, the proposed PSO variants generate offspring solutions through a sampling mechanism during the initial search process. Two offspring generation sampling schemes are investigated, i.e. the employment of the personal and global best solutions obtained using the mutation mechanism, respectively, as the starting positions for the subsequent search process. Subsequently, several machine learning algorithms are used in conjunction with the proposed PSO variants to perform object classification. As evidenced by the empirical results, the proposed PSO variants outperform the original PSO algorithm, significantly, for feature optimization.