This article develops a novel extremum seeking control (ESC) scheme to address an input dead-zone. The new ESC scheme includes two parts: one is the learning mechanism that can rapidly compensate the effect of the dead-zone; the other is the classic ESC scheme, which can regulate the output of the system to an arbitrarily small neighborhood of its extremum. The proposed scheme has a two-time-scale structure, and the boundary layer system is uniformly, practically, and asymptotically stable. By applying results that exist in the literature, semi-global practical asymptotic stability of the new ESC scheme is established. Finally, numerical examples are presented to illustrate the effectiveness of the proposed scheme.