Terahertz (THz) imaging based on Rydberg atom achieves high sensitivity and frame rates but faces challenges in spatial resolution due to diffraction, interference, and background noise. This study introduces a polarization filter and a deep learning-based method using a physically informed convolutional neural network to enhance resolution without pre-trained datasets. The technique reduces diffraction artifacts and achieves lens-free imaging with a resolution exceeding 1.25 lp/mm over a wide field of view. This advancement significantly improves the imaging quality of Rydberg atom-based sensor, expanding potential applications in THz imaging.