2026 m. kovo 2 d., 13 val.
Vilnius, Akademijos g. 4, 203 kab.
Nuotoliniu būdu „MS Teams“ aplinkoje (https://bit.ly/DMSTI_2026-03-02)
Modestas Motiejauskas
(vadovas prof. habil. dr. Gintautas Dzemyda)
„Konvoliucinių neuroninių tinklų architektūrų ir mokymo gerinimas: emocijų atpažinimo nuotraukose atvejis“
Anotacija: This dissertation presents a deep learning framework for visual emotion recognition in general-purpose images. It improves classification by reducing confusion between visually similar emotion classes under subjective and potentially culture-dependent labeling. Emotions are represented using an eight-category label set drawn from open-access datasets. The approach extends an EfficientNetV2S backbone with Gram matrix modules to capture stylistic representations alongside semantic CNN features, and integrates contrastive-center loss to enforce tighter intra-class clustering and greater inter-class separation in the embedding space. Experiments on WEBEmo sadness, FI-8, EmoSet-118K, and WikiArt show consistent improvements over baselines and prior Gram matrix–based approaches. Adding four Gram matrix modules increases accuracy on WEBEmo sadness from 81.31% to 82.52% (+1.21%). Incorporating contrastive-center loss further increases accuracy on WEBEmo sadness from 82.52% to 83.74% (+1.22%), on FI-8 from 69.68% to 71.88% (+2.20%), and on EmoSet-118K from 77.94% to 80.46% (+2.52%). Cluster analysis confirms a better-organized feature space with fewer ambiguous samples. A complementary top-2 cross-sentiment consistency metric shows substantial reductions in opposite-sentiment top-2 predictions on WikiArt, from 41.4% to 19.4% (−22.0%) without confidence thresholding and from 43.5% to 19.9% (−23.6%) with confidence filtering. Overall, the architectural enhancements, optimization strategy, and evaluation methodology support more discriminative and reliable visual emotion classification across diverse real-world datasets.