Offshore Renewable Energy (ORE) is developing worldwide, for which biofouling is a crucial parameter to consider, both for engineering and environmental monitoring purposes. In this study, machine learning tools are used to classify macro-biocolonisation images into four categories: 'mussels', 'barnacles', ‘calcareous worms' and ‘no macro-biocolonisation’ as part of the suspected “most impacting species” of the fluid/Structure behavior of colonized components. A transfer learning approach is investigated using a state-of-the-art convolution neural network (CNN) architecture and an open-source training algorithm is specifically modified to ensure rapid reproducibility of the methodology: image selection, CNN adaptation, training/validation process and quality assessment. A database of 1261 images is set up to train the models; their performances are tested using images of offshore components with and without biocolonisation. Two stochasticity sources were evaluated: coming from the modest size of the data base trough cross validation on 5 folds, and sensitivity to the numerical model through the building of 30 models per fold. The performance was discussed with particular care, given the small number of images available and the statistical uncertainties of the metrics evaluated from the validation process. The average performance of the models on the testing images is 69% of good detection for all classes combined. In engineering point of view the results are satisfactory since the two classes with maximum hydrodynamic impact (mussel and ‘no macro-biocolonisation’) provides average detection of 81% and 79%. Recommendations are proposed to enrich the training image database and photographic surveys on offshore structures, in providing metrics enabling algorithm optimization in engineering purposes. The main innovation of this study is to adapt existing machine learning tools to a new and complex application area: the biofouling.