In situ TEM is by far the most commonly used microscopy method for imaging dislocations, i.e., line-like defects in crystalline materials. However, quantitative image analysis so far was not possible, implying that also statistical analyses were strongly limited. In this work, we created a deep learning-based digital twin of an in situ TEM straining experiment, additionally allowing to perform matching simulations. As application we extract spatio-temporal information of moving dislocations from experiments carried out on a Cantor high entropy alloy and investigate the universality class of plastic strain avalanches. We can directly observe “stick–slip motion” of single dislocations and compute the corresponding avalanche statistics. The distributions turn out to be scale-free, and the exponent of the power law distribution exhibits independence on the driving stress. The introduced methodology is entirely generic and has the potential to turn meso-scale TEM microscopy into a truly quantitative and reproducible approach.