Online dictionary learning (ODL) updates all dictionary atoms and it is difficult to estimate the optimization direction, so the accuracy is decreased. Aiming at this drawback, we propose a method of online dictionary learning based on modified particle swarm optimization (MPSO). On the basis of ODL, the algorithm optimizes the gradient descent function in the iterative process of dictionary learning.  A special atom is selected by a rule in the dictionary-updating stage, which is linearly represented by the other atoms in the dictionary. The coefficient of the linear representation is the position of the particles in the MPSO, which is introduced to eliminate the particles by their fitness while leaving the more suitable particles in the next iteration. Furthermore, intermediate variables, which carry the prior reference data, are introduced into the MPSO to guide the optimization direction, so that the  direction is constrained to improve the accuracy and effectiveness of the dictionary. Experimental results confirm that compared with the other three algorithms, this algorithm has a better performance in noise suppression and improves the performance of large-scale data computation.