We propose a method for creating a semantic profile of user's interests emerging from pictures by application of neural networks trained for object recognition. We use BabelNet, an online encyclopaedic dictionary, to generalise object names into categories of interests. Our method is evaluated with ground-truth data based on social tagging mechanism. Experiments are conducted entirely on original data containing 60,000 images crawled from Flickr, evenly distributed among 300 users. Results show that object recognition methods combined with object category generalisation can be effectively used to predict user's interests. The accuracy of the presented method seems to change with the neural network used for object recognition (5 NN tested in total), therefore it has a strong potential for further development. ResNet-50 turned out to be the most accurate network in our experiment.