Facial Image Quality Assessment using an Ensemble of Pre-Trained Deep Learning Models (EFQnet)
Adebayo Adewumi Abayomi-Alli, Olabode Atinuke, Adebukola Onashoga, Sanjay Misra, Oluwasefunmi 'Tale Arogundade, Olusola Abayomi‐Alli · 2020
Facial recognition is a type of biometric that deals with the facial region of a human image. The low recognition accuracy of existing Facial Recognition Systems (FRS) is due to the low quality of captured images. The assessment of image quality, therefore, becomes a requirement to be taken before passing such an image through the FRS. This study presents a proposed Facial Image Quality Assessment (FIQA) model using an ensemble of pre-trained deep learning models (EFQnet). The system known as EFQnet is an ensemble of ResNet-50, DenseNet, and Inception-Net CNN pre-trained models. It utilizes a performance-based ground truth that forecasts a quality score for the input image between 0 to 1. The three models are ensemble using full fully connected Feedforward Neural Network and AMSGrad stochastic gradient Descent algorithm. When trained and fully implemented EFQnet will be evaluated on standard IQA databases and finally deployed in a Personal Identity Verification (PIV) scenario.