Abstract: The Scanning Tunneling Microscope (STM), operating under ultra-high vacuum conditions, enables atomic-scale resolution imaging of material surfaces. However, STM images are often affected by various sources of noise, which degrades image quality. This paper proposes a deep learning model for STM image restoration, named MAED-CNN - Multi-scale Attention Encoder-Decoder Convolutional Neural Network. It uses artificially repaired STM images as references. The model leverages manually restored STM images as references and combines multi-scale convolution, attention modules, and an encoder-decoder U-Net architecture to transform noisy input images into high-quality, denoised outputs. Compared with several general deep learning models, the proposed model demonstrates superior performance in metrics such as PSNR, SSIM, and UQI. It effectively restores STM images and holds significant promise for advancing STM image restoration techniques and promoting research in imaging technologies.