The vibration signals of faulty rolling bearings usually contain different components. The separation of the fault feature component from the bearing vibration signal is significant for diagnosing bearing faults. Empirical Fourier decomposition (EFD) is an adaptive signal decomposition method based on spectrum division and Fourier transform theory. It provides an efficient tool for extracting rolling bearing fault features. Nonetheless, the existing spectrum division methods have the shortcoming of over-division, which could cause the over-decomposition of EFD and weaken the fault feature extraction capability. This paper provides an improved EFD (IEFD) approach that aims to inhibit the over-decomposition from two aspects. First, an improved scale-space division based on fine to coarse (FTC) segmentation algorithm is developed for spectrum over-division inhibition. Second, an adjacent modes merging framework, based on envelope spectrum kurtosis (ESK) and envelop spectrum overlap coefficient (ESOC) indexes, is exploited to inhibit the modes over-decomposition. The effectiveness and practicability of IEFD are investigated by simulated analysis and experimental application. In comparison to variational mode decomposition (VMD) and EFD, IEFD has superiority in inhibiting over-decomposition and extracting more abundant bearing fault information, which indicates that IEFD can be judged as a reliable bearing fault diagnosis method.