Single-pixel imaging is a technology that reconstructs images by measuring with a small number of pixels, possessing high flexibility. since each pixel is measured indirectly, the image reconstruction quality may not be as good as traditional direct imaging techniques. A single-pixel imaging method that combines the Discrete Cosine Transform (DCT) with an Efficient Sub-Pixel Convolutional Neural Network (ESPCN) is proposed. By using DCT to perform frequency domain decomposition on a small number of pixel measurements, low resolution frequency domain feature representations are obtained while preserving key frequency information. ESPCN maps low-frequency features to high-resolution images through sub-pixel convolutional layers, thereby achieving high-quality reconstruction. Simulation results show that the improved network not only enhances the signal-to-noise ratio and structural similarity of the image but also effectively improves the resolution, proving the feasibility of the network in practical applications.