Tehlivets O. Methods for Improving Image Quality in Information and Communication Systems Based on Deep Learning Technologies

The dissertation solves the scientific and practical problem of developing image enhancement methods based on deep learning technologies for application in modern information and communication systems. With the rapid development of modern information and communication systems, ensuring high-quality visual data during transmission, storage, and processing in networked environments has become increasingly important. Noise, low contrast, blur, compression artifacts, and other types of distortion reduce the informational value of images and may lead to errors in automated systems. The continuous growth in data volume and the increasing demand for high-speed processing necessitate the development of effective methods for image quality assessment and automated image enhancement. The aim of the dissertation is to improve the efficiency of contrast assessment and digital image enhancement in information and communication systems by developing methods for generating test image datasets, applying neural networks, and performing a comparative analysis of expert and neural network assessments. The object of the research is the processes of digital image quality assessment and enhancement in visual information processing systems operating as components of modern information and communication systems. The subject of the research comprises deep learning methods, models, and techniques for digital image quality assessment and enhancement. Research methods. The dissertation employs digital image processing methods, deep learning based on convolutional neural networks, mathematical statistics, objective image quality assessment methods, subjective evaluation methods, statistical analysis, experimental methods, and software implementation techniques. The first chapter presents a comprehensive analysis of methods and algorithms for digital image quality assessment and enhancement in information and communication systems and substantiates the feasibility of integrating neural network approaches into the modular architecture of information and communication systems. Image quality assessment methods, contrast enhancement techniques, modern image quality metrics, and neural network-based image processing models are reviewed. The advantages of convolutional neural networks for digital image enhancement are identified. The second chapter develops and improves digital image quality assessment methods based on deep neural networks. Representative test datasets are constructed using the publicly available TID2013, CEED2016, and LOL databases. A specialized grayscale image database is created, and an automated image quality assessment algorithm based on a convolutional neural network is proposed. An operating algorithm for the deep neural network is developed, incorporating image preprocessing, multi-level feature extraction, local feature aggregation, and regression-based quality estimation. Experimental results confirm the effectiveness of the proposed model, demonstrating a high level of agreement with subjective assessments (up to 90% agreement in pairwise comparisons). The third chapter substantiates the selection of a deep neural network architecture for digital image enhancement and implements algorithms based on the MIRNet architecture and simplified convolutional neural networks. Experimental studies confirm the effectiveness of the proposed models and demonstrate that the simplified architecture reduces computational costs without significantly degrading the quality of the restored images. The fourth chapter presents a comprehensive experimental evaluation of the proposed image quality assessment and enhancement methods. A comparative analysis of conventional and neural network-based approaches confirms the effectiveness of the MIRNet architecture and simplified convolutional neural networks for deployment in resource-constrained information and communication systems. According to the SSIM metric, MIRNet achieves the best performance (up to 0.994), whereas conventional image editing tools provide SSIM values ranging from 0.938 to 0.988. The results demonstrate that manual and semi-automatic image enhancement requires substantial human intervention and exhibits lower consistency in batch processing. It is further shown that the GPU-based MIRNet approach processes a single image in 0.03-0.12 s, while conventional image editors require approximately 5-60 s per image. A comparative study of the proposed simplified CNN and the MIRNet architecture is performed on both full-color and grayscale images under different types of distortions, including reduced contrast and decreased brightness. The obtained results confirm the effectiveness of the proposed methods and justify the use of less complex neural network architectures for image enhancement tasks in resource-constrained information and communication systems.