METHODS FOR DESIGNING AND TRAINING NEURAL NETWORKS USING THE KERAS LIBRARY

Abstract

Modern methods of visual data processing require precise and automated reconstruction of color information in grayscale images, which makes the task of colorization highly relevant in the context of digital restoration, content generation, and archival preservation. The aim of this study is to develop and train a neural network autoencoder model based on convolutional neural networks (CNNs) to solve the problem of automatic image colorization. The research involved theoretical analysis of neural network architectures, practical implementation of the model using the Keras library, and training on a dedicated dataset of landscape images. As a result, a model was obtained that achieved an accuracy of 82.5% while maintaining visually acceptable colorization quality. The proposed architecture demonstrates high reproducibility, adaptability, and efficiency under local execution conditions. The findings confirm the scientific and practical significance of the proposed solution in the field of deep learning and image processing.