Light-Weight Face Shape Classifier for Real-Time Applications

Deep neural networks (DNNs) are memory and computationally intensive; hence they are difficult to apply to real-time systems with limited resources. Therefore, the DNN models need to be carefully optimized. The solution was a model based on a convolutional neural network (CNN) called MobileNet that decreases the computational and space complexities with classification precision loss by utilizing depthwise separable convolutions. This study uses MobileNet vl architecture to improve image classification complexities to reach an acceptable complexity that can be used in real-time applications that require a hasty response from the model. In this study, the MobileNet model was trained on a dataset consisting of 5000 images to be classified into the 5 human face shapes oval, square, heart, oblong, and round. The model has an F1-score of 0.781, recall of 0.782, precision of 0.78, and achieved an accuracy of 98.8%. With this level of accuracy, a real-time application that is result-driven would benefit significantly from this model.