We propose an end-to-end model that estimates the exit pupil wavefront directly from phase diversity images using deep learning. The aim is to restore the exit pupil wavefront through zonal reconstruction to obtain more high-order modal aberrations, thereby improving the reconstruction quality of degraded images. Our simulated experimental results show that zonal reconstruction significantly outperforms modal reconstruction in restoring high-order aberrations. The ResNet50 model, which outputs Zernike modes of [3,28] orders, is limited by the order of reconstruction, and high-order aberrations lead to errors in these Zernike modes. Zonal reconstruction without mode restrictions, however, can obtain more high-order modal aberrations and achieve higher fitting accuracy across all aberrations. By comparing the reconstruction results of degraded images, as well as analyzing the normalized power spectrum curves and intensity profiles, we further validate the accuracy of the wavefront restoration by zonal reconstruction. The reconstructed images are richer in high spatial frequency details and provide more accurate reconstruction results.