The idea suggested in this paper consists in forensic sketches that have to be drawn and furthermore analyzed from an emotional face recognition point of view. The detection itself is more effective and useful to further identify the future behavior and possible reactions of the suspect based on the emotional analysis of the areas on the face. The study embraces this idea and experiments with forensic like sketch images within a system that is trained on a FER+ database, and executed using two different schemes, ML (Multi-Label learning) and CEL (Cross-Entropy Loss). The conducting experiment resulted that for forensic sketch images, CEL system is more efficient than ML training method 69.77% vs 69.05% for the best test accuracy, even if the experiment on the standard image equivalents proved otherwise.