This paper is concerned with a partially observed optimal control problem described by mean-field forward and backward stochastic differential equations. Moreover, the control variable enters the diffusion coefficient and the control domain is non-convex. Utilising Girsanov's theorem as well as extended Ekeland's variational principle, a maximum principle is established in the form of Pontryagin's type. As an application, a linear-quadratic control problem is studied in terms of the stochastic filtering.