In a contemporary technological milieu characterized by an increasing demand for high-resolution images, concomitant with the rapid advancement of digital technology, this study presents an innovative deep learning approach for the high-resolution reconstruction of low-resolution images. As the necessity for high-quality images becomes more pronounced in diverse fields such as medical imaging, satellite photography, and security surveillance systems, there exists an imperative need to develop advanced technologies capable of surmounting the limitations of existing interpolation methods.

This investigation has developed a novel deep learning methodology that integrates convolutional neural networks (CNNs) and super-resolution models utilizing CIFAR-10 datasets. Transcending the limitations of detailed information restoration exhibited by traditional image upscaling methods, a deep neural network model has been designed to extract and learn progressively from low-level features to high-level features.

The core methodology of the study resides in the development of two complementary deep neural network models. The first classifier model performs effective classification of low-resolution images, while the second super-resolution model successfully upscales the low-resolution input image of 16x16 pixel size to a high-resolution image of 32x32 pixels. In this process, the performance of the model was optimized through the utilization of the Adam optimizer and various loss functions (sparse categorical cross-entropy, mean square error).

The experimental results demonstrate that the proposed model exhibited significantly improved image restoration performance compared to existing interpolation-based methods. In the performance evaluation based on the mean square error (MSE) and the mean absolute error (MAE), superior results were obtained across various image categories of the CIFAR-10 dataset, clearly demonstrating the practical applicability of artificial intelligence-based image restoration technology.

Limitations of the study include the current model's restriction to CIFAR-10 datasets and the need for improvement in computational complexity. Accordingly, future research endeavors aim to continuously develop image restoration technology through generalization performance verification, model weight reduction, and integration of generative adversarial neural networks (GANs) for other image datasets.

This study is anticipated to contribute to the advancement of computer vision and artificial intelligence technology by presenting a novel approach to the high-resolution reconstruction of low-resolution images.

Keywords

Deep learning, Convolutional neural network, Super-Resolution, Image restoration and CIFAR-10.