Dealing with noise in digital images is an important task in various scientific research processes. Noise often has different characteristics, with one of the key factors being its randomness. Distorted images many times are treated with techniques to remove distortion and improve image quality. However, in some cases, simulated distortion itself could be useful for testing hypothesis about the performance of the cleaning process or specific patterns. This paper presents an analysis of the performance of a convolutional autoencoder (CAE), proposed by us. The CAE reconstructs images from the MNIST dataset that were distorted with three increasing levels of binary noise: 10% (79 pixels), 20% (157 pixels), and 30% (235 pixels), respectively. In the validation stage, SSIM metric was used to compare the reconstructed images with original ones, showing that SSIM’s value decreased as distortion level increased: 0.918 (10%), 0.940 (20%), 0.900 (30%). Furthermore, the loss function throughout 150 epochs has shown: 0.008, 0.009, 0.113, respectively. In the experimental work with this CNN architecture images were reconstructed with very good visual quality, highlighting the potential for practical applications in imaging technology.