Aiming at the problem of pose feature loss due to up and down sampling in the inference process of head pose estimation, a high-resolution feature maintained soft-stage regression algorithm for head pose estimation is proposed. The algorithm first utilizes the encoder HR-Net to encode multiscale features for high-resolution feature maintaining in raw face images, and TA dimension interaction module joined in its convolutional block to capture more spatial-channel interaction information. The decoder SSR-Net algorithm was then applied to decode the key parameters and soft-stage regression of head pose on the different scale features output from HR-Net, and the Efficient Channel Attention ECA is employed to enhance the information interaction between feature channels and reduce redundant features. The experimental results show that the proposed algorithm has excellent performance on both the public datasets AFLW2000 and BIWI, and its MAE is reduced to 4.19 and 3.00, respectively.