Recognition of human emotions on digital images using convolutional neural networks

V magistrskem delu smo se ukvarjali z razvrščanjem šestih osnovnih človeških emocij in nevtralnega izraza s pomočjo digitalnih posnetkov in konvolucijskih nevronskih mrež. Pregledali smo področje razpoznavanja človeških emocij in natančno preučili konvolucijske nevronske mreže. Implementirali smo več modelov sodobnih konvolucijskih nevronskih mrež, ob tem pa razvili tudi lastne modele. Uporabili smo knjižnico Tensorflow in programski jezik Python. Naše predlagane rešitve smo preizkusili na prosto dostopnih podatkovnih zbirkah CK+, MMI in JAFFE. Slike iz podatkovnih zbirk smo obogatili z zrcaljenjem in rotiranjem, tako da smo dobili večjo količino podatkov. Za validiranje smo uporabili pristop, neodvisen od subjekta, in petkratno navzkrižno validacijo. Najboljši rezultati razvrščanja z našimi predlaganimi metodami so bili 91,65 % na zbirki CK+, 59,08 % na zbirki MMI in 67,86 % na zbirki JAFFE. Rezultati na zbirki CK+ so primerljivi z rezultati sorodnih del, na preostalih dveh zbirkah pa je uspešnost razvrščanja z našimi pristopi bistveno slabša od rezultatov sorodnih del.In the master\u27s thesis, we dealt with the classification of six basic human emotions and the neutral expression on digital images using convolutional neural networks. We reviewed the field of human emotion recognition and examined convolutional neural networks. We implemented several existing models of convolutional neural networks and also developed our own models. We used the Tensorflow library and the Python programming language. We tested our solutions on freely accessible CK +, MMI and JAFFE databases. The images from the databases were augmented by mirroring and rotating the images in order to obtain a larger amount of data. We used a subject-independet aproach for validation and 5-fold cross validation. The best classification results with our proposed methods were 91,65 % on the CK+ database, 59,08 % on the MMI database and 67,68% on the JAFFE database. The results on the CK+ database are comparable to the results of related works, while on the other two databases the success of the classification with our best approaches is significantly worse than the results of related works