Modelling and dense network model for moderate facial alignment prediction by feature representation

Deep learning approaches are extremely productive and accurate for predicting moderate face alignment and facial landmarks. The model learns sequential analysis during the training process to reduce discrimination between the ground truth value and the shape of the face based on feature representation. While testing, it employs feature representation to identify the shape factors iteratively. Also, when the facial directions and expressions change, the existing learning model cannot acquire superior performance based on the enormous variations among the target and the initial shape. This work proposes a novel multi-stage gradient descent with the ResNet-50 model to preserve higher prediction accuracy on training samples and enhance the testing data accuracy. One sample is provided during training, and multiple samples are provided with changing expressions. During testing, the distance among the face alignment landmarks is evaluated with the optimal selection of ligaments. The simulation is done in MATLAB 2020a environment.outcomes show that the anticipated model can enhance the conventional approaches' performance and show a better trade-off than other approaches.