The existing bidimensional empirical mode decomposition (BEMD) algorithms are inefficient in extrema searching, intrinsic mode functions sifting and iteration processing, moreover the adaptability needs to be further improved. Therefore, this paper proposes an improved BEMD method based on multi-scale extrema. Firstly, the concept and establishment method of bidimensional multi-scale local extrema binary tree are given, and then a new approach based on multi-scale extrema is presented to determine window sizes for order-statistics and smoothing filters. This method significantly improves the adaptability of multi-scale decomposition of two-dimensional signals, and also significantly improves the decomposition efficiency. Experimental results of natural image and synthetic texture image decomposition show that the proposed method has obvious advantages in adaptability and efficiency compared with the existing fast adaptive EBMD method.