Structural analysis of heterogeneous and cracked walls and bridges has been a challenge. The major difficulty is the introduction of the geometry and required pattern and damage information. Several works report on automatic creation of geometry data and subsequent structural analysis using linear and nonlinear finite element or discrete element models. Image analysis supported by deep learning is able to capture this piece of information and provide detailed geometrical models for all visible parts of the structure. The complexity of the resulting mechanical model still remains high. The solution to this problem is provided by numerical homogenization, which can be used, instead of currently used discrete and interface models, in order to describe the overall behavior by taking into account the real microstructure. Numerical homogenization that introduces homogenized material properties at each point by using detailed solution of representative volume elements at different loading levels and subsequent neural network based approximation of the constitutive parameter have been introduced in and extended here in order to use different microstructures at different parts of the structure. The proposed methodology will be outlined for static analysis tasks, constituting a contribution towards a combined geometric-mechanical digital twin of masonry structures.