Face Expression Recognition via transformer-based classification models

Balkan Journal of Electrical and Computer Engineering, Journal Year: 2024, Volume and Issue: unknown

Published: Aug. 20, 2024

Facial Expression Recognition (FER) tasks have widely studied in the literature since it has many applications. Fast development of technology deep learning computer vision algorithms, especially, transformer-based classification models, makes hard to select most appropriate models. Using complex model may increase accuracy performance but decreasing inference time which is a crucial near real-time On other hand, small models not give desired results. In this study, we aimed examine 5 different relatively image algorithms for FER tasks. We used vanilla ViT, PiT, Swin, DeiT, and CrossViT with considering their trainable parameter size architectures. Each 20-30M parameters means small. Moreover, each As an illustration, focuses on using multi-scale patches PiT introduces convolution layers pooling techniques ViT model. obtained all results datasets: CK+ KDEF. observed that, achieves best scores 0.9513 0.9090 KDEF datasets, respectively

Language: Английский

Improved optimizer with deep learning model for emotion detection and classification

Mathematical Biosciences & Engineering, Journal Year: 2024, Volume and Issue: 21(7), P. 6631 - 6657

Published: Jan. 1, 2024

Facial emotion recognition (FER) is largely utilized to analyze human in order address the needs of many real-time applications such as computer-human interfaces, detection, forensics, biometrics, and human-robot collaboration. Nonetheless, existing methods are mostly unable offer correct predictions with a minimum error rate. In this paper, an innovative facial framework, termed extended walrus-based deep learning Botox feature selection network (EWDL-BFSN), was designed accurately detect emotions. The main goals EWDL-BFSN identify emotions automatically effectively by choosing optimal features adjusting hyperparameters classifier. gradient wavelet anisotropic filter (GWAF) can be used for image pre-processing model. Additionally, SqueezeNet extract significant features. improved optimization algorithm (IBoA) then choose best Lastly, FER classification accomplished through use enhanced optimization-based kernel residual 50 (EK-ResNet50) network. Meanwhile, nature-inspired metaheuristic, walrus (WOA) pick EK-ResNet50 model trained tested publicly available CK+ FER-2013 datasets. Python platform applied implementation, various performance metrics accuracy, sensitivity, specificity, F1-score were analyzed state-of-the-art methods. proposed acquired overall accuracy 99.37 99.25% both datasets proved its superiority predicting over

Language: Английский