Improving Head Pose Estimation with Neural Network Optimization on the AFLW2000 Dataset
Jinghong Duan · 2023
The problem of large regression loss values in the performance of the AlexNet model based on deep learning has been solved. The best combination of hyperparameters, including batch size, number of iterations, and learning rate, was obtained through comparative experiments while improving the structure of the neural network layers of the model. On the AFLW2000 dataset, the average absolute error loss value decreased from 59.89 to 15.86, and the prediction performance of Euler angles increased by 73.5%. The experiment made the AlexNet CNN model perform better on the AFLW2000 dataset, proving the superiority of feature regression in predicting head pose angles in all directions and improving the robustness of the AlexNet model.