A strengthened autoencoder formed by placing an object detector upstream of a decoder is here developed in the context of the model-helped human analysis of composed ornaments from a dictionary of vignettes. The detection part is in charge to detect regions of interest containing some vignette features, and the decoding part to ensure vignette reconstruction with a relative quality depending on feature match. Images of ornaments without typographical composition are generated in order to properly assess the performance of each of the two parts.