Masked face age and gender identification using CAFFE-modified MobileNetV2 on photo and real-time video images by transfer learning and deep learning techniques

One of the most challenging factors related to masked face age and gender identification (MFAGI) is developing a technique to quickly carry out identification and maintain accuracy without needing people to remove their masks. The eyes, forehead and important frontal features including the ears were the main focus of this investigation. In this study, we evaluated and compared ten (10) different MFAGI techniques, which include traditional methods such as MobileNet, MobileNetV2, ResNet101, DenseNet169, DenseNet201, EfficientNetB0, ResNet152V2, InceptionV3, Xception and our one proposed approach. These 10 models were used to classify masked faces of people among sixteen (16) classes. Combining the convolutional architecture of fast feature embedding (CAFFE) with the modified MobileNetV2 (MNV2) model developed a novel method known as the "CMNV2"model. To achieve the goal of model performance and classification accuracy for MFAGI of people's faces, this work introduces a unique CMNV2 model by developing a pre -trained model with eight (8) different layers. Using deep learning (DL) and a transfer learning approach, we developed a unique novel approach to image feature extraction that outperforms the nine (9) traditional methods for masked face classification (MFC), masked face age identification (MFAI) and masked face gender identification (MFGI). This approach developed a Python model that was trained and evaluated on 12,160 masked face images. This study demonstrates that an efficient CMNV2 method can be achieved by combining transfer learning with deep neural network (DNN) development. This proposed CMNV2 model achieved 96.54% accuracy and 3.46% error rate with less parameters and computational time exceeding the other 9 different traditional masked identification methods. It also shows that our proposed CMNV2 strategy is more successful than the other 9 models and can identify masked people's faces age and gender in photo and real-time video images. Overall, the results of this research indicate that DL methods have the capability to improve the accuracy and speed with which masked face classification and identification is performed. The artificially masked faces of images dataset created by us utilized in this work is publicly made available on the Kaggle website https: //www.kaggle.com/datasets/banilkumar20phd7071/masked-face-age-and-gender-identify-artificial-masking.