The work investigates the automated surface crack detection in fracture experiments. For the analysis, widely employed uniaxial tension, shear and punch loading experimental image analysis data are employed. Thereupon, deep neural network modelling architectures that are based on texture features are elaborated, characterizing their applicability and accuracy limits [1]. What is more, convolutional neural network architectures are considered as generalized image data-based fracture classification and identification algorithms. To that scope, transfer learning, as well as greedy and Bayesian search approaches are employed. Low computational cost convolutional neural network architectures that are capable of automatically identifying and localizing cracks in material testing experiments are elaborated, making use of the raw and unprocessed experimental data. The required dataset size, as well as the optimized total number of parameters and their inner convolution to dense layer distribution are analysed for each material testing case.