Human Facial Pattern Shape Classification Using a Retraining Strategy and Convolutional Neural Network Architecture

Many shapes and patterns on the human body might be considered a person's uniqueness or feature since they differ significantly from one another, one of which is the shape of the face. In computer vision, the shape of a face is divided into five fundamental shapes. The experiment in this paper provides a model based on the final layer of the results of retraining InceptionV3, a Convolutional Neural Network (CNN) architecture for classifying human face photos. Inspired by human neural networks, this method generally works well for face recognition and computer vision research. This research begins with the stages of data acquisition, data exploration, classification, and evaluation. Retraining is performed to improve accuracy using the distance and angle of facial landmarks. The results are compared to other classification methods, including linear discriminant analysis (LDA), support vector machine with a linear kernel (SVM-LIN), support vector machine with a radial basis function kernel (SVM-RBF), artificial neural networks or multilayer perceptrons (MLP), and k-nearest neighbors. The facial dataset used consists of 747 photos, divided into five categories: oval, round, square, heart, and oblong. The Canny edge detector approach is utilized to enhance CNN accuracy, which has been effectively improved through training and testing. The maximum accuracy achieved was 91.7% based on training and testing at 85%-98%. This demonstrates that the outcomes of inceptionV3 retraining may appropriately adapt training data and outperform alternative classification techniques without the need to specify the function of certain features during the model training process.