Feature Selection (FS) is a crucial data preparation method used in the classification process. FS challenges have been successfully solved with a variety of evolutionary computational approaches. Aiming at the shortcomings of the existing wrapping feature selection algorithm with low search efficiency, and considering the complementary advantages of the filtering method and the wrapping method in feature selection, this paper proposes a multi-objective feature selection algorithm that combines mutual information and NSGA-II (MOFS-MIGA). Firstly, in order to form a good frontier in the initial generation and improve the convergence speed of the algorithm, the symmetric uncertainty (SU) between each feature and class label is calculated based on the mutual information, and the population is initialized based on this uncertainty; secondly, a new crossover and mutation method is suggested to enhance the directionality and convergence of NSGA-II. Ultimately, the suggested algorithm achieves a set of converged and well-distributed solutions on the majority of datasets and has good feature selection performance. This is demonstrated by the experimental results, which compare the algorithm's performance to five traditional multi-objective feature selection algorithms on nine datasets of varying dimensions.