Particle swarm optimization (PSO) has been widely applied to solving complex optimization problems due to its simple parameter settings and fast convergence speed. However, the classical PSO algorithm suffers from inherent limitations, such as premature convergence and slow convergence in the later stages. Therefore, a cooperative self-evolutionary particle swarm optimization algorithm is proposed. The proposed algorithm employs a novel dual-population cooperative co-evolution strategy to accelerate the convergence process. Meanwhile, to balance global exploration and local exploitation capabilities, a self-evolutionary framework is introduced, in which a probabilistic biased directional learning strategy is combined with a decaying hybrid perturbation mechanism, effectively enhancing the overall optimization performance. In addition, the boundary-handling mechanism of the algorithm is further improved, thereby increasing its adaptability across a wide range of optimization problems. The improved algorithm is evaluated on the CEC-2017 benchmark function sets, demonstrating fast convergence and superior optimization performance on low-, medium-, and high-dimensional complex problems. Finally, the proposed algorithm is applied to the threshold optimization problem in multilevel image segmentation. Experimental results show that the improved algorithm significantly enhances segmentation accuracy and efficiency, thus validating its effectiveness in solving real-world optimization problems.