Performance assessment of morphological dynamic link architecture under optimal and real operating conditions

In this paper, the performance of morphological dynamic link architecture (MDLA) is assessed under optimal and real operating conditions. It is shown that MDLA achieves a very low equal error rate on the extended M2VTS database which contains 295 persons’ video data in 8 shots recorded under optimal conditions. However, its performance severely deteriorates, when it is applied to the IBERMATICA database that has been recorded under conditions simulating a real access-control to an automatic teller machine. The compensation for the image variations attributed to the variable recording conditions, i.e., changes in illumination, face size differences and varying face position, is addressed next. The use of simple and powerful pre-processing techniques aiming at compensating for the aforementioned variations prior to the application of MDLA is proposed. The first results obtained indicate that the proposed approach overcomes the image variations and stabilizes the performance of MDLA.