DEVELOPMENT AND TRAINING OF A NEURAL NETWORK AUTOEN- CODER MODEL FOR VISUAL DATA COLORIZATION

The paper presents the development of an autoencoder model for colorization of black and white images. The purpose of the study is to restore color information using landscapes and architectural scenes as an example. The model is implemented in Python using the Keras and TensorFlow libraries and trained on the dataset Landscape color and grayscale images (14,258 images, including 7,129 color and 7,129 converted to black and white). All images are reduced to 150x150 pixels in RGB format. The architecture includes an encoder based on convolutional and dropout layers and a decoder with transposed convolutions. Augmentations were used to improve the generalization ability. Experiments showed an accuracy of about 82.5 % on the validation sample. Additional metrics: MSE = 0.018, PSNR ≈ 27.6 dB , SSIM ≈ 0.83 . Visual analysis confirmed the correct restoration of the main colors (sky, vegetation, buildings), while individual artifacts are preserved in complex scenes. Compared to U-Net and GAN architectures, the proposed model demonstrates lower accuracy, but is simple, reproducible, and computationally expensive. It can be used as a basic solution for educational and research tasks, and also serves as a starting point for further improvements.

For citation: M. Urazgaliyeva, H.İ. Bülbül, B. Utenova, A. Mailybayeva, A. Mukhanbetkaliyeva. Development and training of a neural network autoencod- er model for visual data colorization//International journal of information and com- munication technologies. 2025. Vol. 6. No. 23. Pp. 303–322. (In Eng.). https://doi. org/10.54309/IJICT.2025.23.3.019.