This paper presents an integrated approach to enhancing low-resolution images that combines autoencoders for removing noise and sub-pixel convolutional neural networks for super-resolution. An autoencoder first denoises images by compressing them to capture the core features while removing noise; then, it reconstructs them clearly. This first stage ensures the image keeps the key information without unwanted distortions. After this, the sub-pixel CNN upscales the denoised images, converting the low-resolution pixels into high-resolution outputs with fine detail, avoiding the blurriness characteristic of many up-sampling methods. The sub-pixel CNN manages this by utilizing pixel-shifting techniques to recover fine details in images lost throughout simple interpolation processes. Low- and high-resolution image pairs for training are supplied with datasets such as Flicker and BSD500. These datasets contain diverse image types, which makes the model generalize well over many different real-world scenarios. Metrics of model evaluation PSNR and SSIM, which are used, ensure that output has the quality of the perceptual standards, proving that the model’s quality is the same on different types of images.