Deep learning is promising approach to extract useful nonlinear representations of data. However, it is usually applied with large training sets, which are not always available in practical tasks. In this paper, we consider stacked autoencoders with logistic regression as the classification layer and study their usefulness for the task of image categorization depending on the size of training sets. Hand-crafted image descriptors are proposed and used for training autoencoders in addition to pixel-level features. New multi-column architecture for autoencoders is also proposed. Conducted experiments showed that useful nonlinear features can be learnt by (stacked) autoencoders only using large training sets, but they can yield positive results due to redundancy reduction also on small training sets. Practically useful results (9.1% error rate for 6 classes) were achieved only using hand-crafted features on the training set containing 4800 images.