Due to the limitations of seismic data acquisition and processing, real seismic data often suffer from low resolution, noise contamination, and other issues, which pose significant challenges for subsequent seismic interpretation. This paper proposes a visual perception-guided seismic image super-resolution reconstruction method to address the task of enhancing low-resolution seismic images to high resolution. The proposed method is based on the U-Net deep learning model and incorporates a perceptual loss function to guide the reconstruction process perceptually. Additionally, L1 loss and Total variable difference loss are combined to balance the quality and smoothness of seismic image reconstruction, thereby reducing high-frequency noise and checkerboard artifacts. Moreover, sub-pixel convolution is introduced to recover high-frequency detail information and mitigate the occurrence of checkerboard artifacts. To address the challenge of processing large-size images using the U-Net model, an overlapping patch cutting strategy is adopted to divide seismic images into multiple smaller patches with different features. Data augmentation techniques, such as random flipping and rotation, are applied to enhance the model's generalization ability and robustness. Experimental results show that, compared to existing super-resolution models, the proposed model demonstrates better performance, effectively recovering high-frequency details of seismic images, reducing noise, enhancing the continuity of seismic events and geological features such as faults, and facilitating subsequent manual annotation and automatic recognition.