Face Recognition on Low Resolution Face Image With TBE-CNN Architecture

Abstract

Face recognition in low resolution images has challenges in active research because face recognition is usually implemented in high resolution images (HR). In general, research leads to a combination of pre-processing and training models. Therefore, this study aims to classify low-resolution face images using a combination of pre-processing and deep learning. In addition, this study also aims to compare evaluation results based on differences in epoch values. In this research will use the Trunk Branch Ensemble-Convolutional Neural Network (TBE-CNN) as one of Convolutional Neural Network (CNN) architecture combined with the pre-processing method which is super resolution to get better accuracy compared to the state-of-art. Then the model is trained with different epoch values, that is 40 and 100, to comparing the best classification. This study was evaluated by YTD Dataset. Based on the test results, this proposed method has achieved better results compared to the previous method, which gives an increase of 1%. After that, training results from different epoch values used produce different accuracy at the training model but have no effect on validation and testing model. The training using epoch = 100 has an accuracy increase of 3% compared to training using epoch = 40 and the loss ratio obtained at the training is decreased by 20% compared to training at epoch = 40. Furthermore, optimization parameters used and reducing computing time when training models and validation will be future research.

Convolutional Neural Networks for Face Recognition: A Systematic Literature Review

“…(Sepas-Moghaddam et al, 2020) proposes a double-deep spatio-angular framework for light field-based face recognition, which is able to model both the intra-view/spatial and inter-view/angular information using two DCNN in sequence. (Suharjito and Puspita, 2020) classify low-resolution face images using a combination of pre-processing and deep learning. They use the Trunk Branch Ensemble CNN.…”

“…One very important finding was that there were 17 papers (Zafar et al, 2019;Khalajzadeh et al, 2013;Syafeeza et al, 2015;Peng et al, 2016;Bussey et al, 2017;Zeng et al, 2017;Almabdy and Elrefaei, 2019;Zeng et al, 2018a;Khan et al, 2020;Okokpujie and Apeh, 2020;Choi and Lee, 2020;Sepas-Moghaddam et al, 2020;Kim et al, 2020;Suharjito and Puspita, 2020;Darma and Mohamad, 2021;Alhanaee et al, 2021;Li et al, 2022) that reported an accuracy od 100%. With 4 papers (Bussey et al, 2017;Almabdy and Elrefaei, 2019;Sepas-Moghaddam et al, 2020;Li et al, 2022) reporting an accuracy of 100% in more than one category.…”

“…They achieve promising results using a fine-tuned DCNN model, indicating the potential of deep learning for such scenarios. (Suharjito and Puspita, 2020) aim to classify low-resolution face images using a combination of pre-processing and deep learning, employing super-resolution for better accuracy. (Zeng et al, 2018a) proposes a scheme that combines traditional and deep learning (TDL) methods for face recognition.…”