The number of layers and the depth of the network are gradually increasing as the research on deep learning networks deepens and the accuracy of the network model improves, leading to an increase in computation. The lightweight, efficient and accurate network model becomes the key to research based on the need of deep learning model facial landmark detection for deployment on embedded devices. Therefore, an ultra-lightweight facial landmark detection network based on Ghost Model and Ghost Bottleneck is designed in this thesis to ensure the network accuracy while minimizing the network model size and reducing the computational effort. With a network width factor of 1X, the normalized mean error is reduced by 7% and the number of parameters is reduced by 36% compared to the best performing lightweight network model PFLD 1X; with a width factor of 0.25X, the proposed network model is only 420 KB in size, and the normalized mean error is reduced by 6.6% and the number of parameters is reduced by the average normalized error is reduced by 6.6% and the number of parameters is reduced by 25%.