To this day, various methods are used in order to reveal the microme-chanical mechanisms of damage in materials. Post-mortem analysis at different stages of stress reveals only snapshots of the material, while in-situ methods are spatially limited to observing the evolution of only a few damage events. A limiting factor in all those methods is the amount of work involved in controlling the microscope and the image analysis. In this work, we implement different structures and architectures of neural networks for the localization and classification of damages. As a sample material dual-phase steels are chosen, due to the different responses of the ductile ferrite matrix and the brit-tle martensite islands to stress, resulting in the formation of damage sites belonging to distinct classes at early stages of deformation. The developed algorithms can on the one hand be used in order to auto-mate the statistical evaluation of post-mortem micrographs, while on the other hand enabling in-situ experiments to generate statistically relevant data. Due to the computational nature of this method, a high throughput of data is possible, enabling a more complete understand-ing of failure mechanisms in many materials.