An Optimization Strategy for Efficient Facial Landmark Detection Based on Improved Pixel-in-Pixel Net Model

Efficient facial landmark detection has been applied to various fields, such as driverless driving, facial beautification technology, facial expression analysis, etc. However, in specific practical tasks, there are still some situations where facial expression cannot be correctly recognized or analyzed. This paper proposes an improved MobileNetV2_re method to improve the loss of accuracy of key points of the problem of pixel-in-pixel Net (PIPNet) in the existing facial landmark detection task. We use the ghost module to replace part of the inverted residual block from the original model, build a new MobileNetV2_re network, and improve the accuracy of the model. It is proved that the situation where high NME and low AUC of PIPNet in the original network MobileNetV2 can be effectively improved by comparing the tested normalized mean error (NME) and the area under the curve (AUC) value and selecting a better network. Compared with MobileNetV2, Resnet18, Resnet50, and Resnet101, NME of MobileNetV2_re in PIPNET is reduced by about 14.07%, AUC of MobileNetV2_re in PIPNET is increased by about 7.52%, and it shows higher accuracy in efficient facial landmark detection.