We propose a new bound for generalization of neural networks using Koopman operators. Unlike most of the existing works, we focus on the role of the final nonlinear transformation of the networks. Our bound is described by the reciprocal of the determinant of the weight matrices and is tighter than existing norm-based bounds when the weight matrices do not have small singular values. According to existing theories about the low-rankness of the weight matrices, it may be counter-intuitive that we focus on the case where singular values of weight matrices are not small. However, motivated by the final nonlinear transformation, we can see that our result sheds light on a new perspective regarding a noise filtering property of neural networks. Since our bound comes from Koopman operators, this work also provides a connection between operator-theoretic analysis and generalization of neural networks. Numerical results support the validity of our theoretical results.