IMAGE QUALITY ASSESSMENT: MODERN APPROACHES AND METRICS

Аннотация

Image quality assessment (IQA) plays a crucial role in digital imaging, compression, and processing. This research analyzes subjective and objective IQA methods, exploring advanced metrics such as PSNR, SSIM, MTF, and deep learning-based automated evaluation techniques. The study highlights the advantages and limitations of traditional and AI- driven assessment models, emphasizing their applicability in real-world scenarios like medical imaging, remote sensing, and video processing. Furthermore, we investigate the integration of quantum computing for IQA and its potential impact on high-dimensional image evaluation. The effectiveness of hybrid models, combining conventional and deep learning techniques, is also explored to optimize the trade-off between accuracy and computational efficiency. The findings suggest that hybrid IQA frameworks outperform single-method approaches, offering a balance between interpretability and automation. Experimental results demonstrate that AI-based IQA approaches significantly improve accuracy and reliability compared to conventional metrics. Finally, we discuss the implications of AI-based IQA models in real-time applications, including autonomous systems and augmented reality. Introduction. Image quality assessment (IQA) is a fundamental aspect of digital imaging, impacting applications ranging from photography and medical diagnostics to remote sensing and artificial intelligence. As digital imaging technologies evolve, the demand for high-quality visual data continues to grow. High-quality images are essential for accurate data interpretation, and IQA methods help quantify image degradation due to noise, compression, and transmission. Reliable IQA frameworks enable automated quality control in industrial imaging, reducing errors in defect detection and pattern recognition tasks. Traditional IQA techniques rely on subjective evaluation, where human perception plays a key role. However, subjective assessment is time-consuming, inconsistent, and impractical for large datasets. Moreover, subjective evaluation is influenced by external factors such as display quality, ambient lighting, and individual observer differences, leading to variability in results. Hence, objective IQA methods have been developed, incorporating mathematical models to evaluate image distortions systematically. These methods facilitate large-scale image analysis in domains like healthcare, security, and environmental monitoring. With the rise of artificial intelligence, deep learning-based IQA models have gained prominence, offering improved automation, adaptability, and accuracy. Deep learning has revolutionized IQA by enabling the prediction of perceptual image quality without reference images. This paper explores conventional and modern IQA techniques, comparing their effectiveness and potential applications. The paper also highlights the role of multi-modal data fusion in improving IQA accuracy for complex imaging systems. 78 RAQAMLI DUNYO: MATEMATIK VA INFORMATIK 2025 -Yil YONDASHUVLAR 13-Fevral Respublika ilmiy-uslubiy konferensiyasi Research Methodology. This study employs a comparative analysis of IQA methods, examining traditional mathematical metrics and AI-driven techniques. A hybrid evaluation framework is proposed, integrating classical and deep learning-based assessment models for enhanced accuracy. The research includes: 1. Objective IQA Metrics: Analyzing PSNR, SSIM, FSIM, VIF, and MTF. Each metric is evaluated for its sensitivity to different types of image degradation, such as blurring, compression, and noise. 2. Deep Learning-Based IQA Models: Evaluating CNNs, GANs, and transformer-based architectures. The architectures are tested on benchmark datasets to determine their generalizability across different imaging conditions. 3. Experimental Validation: Conducting tests on datasets from medical imaging, satellite imagery, and compressed media files. The datasets are preprocessed to normalize variations in resolution, contrast, and illumination. 4. Tools such as MATLAB, Python (OpenCV, TensorFlow), and IDL (Interactive Data Language) are used for data processing and metric analysis. Computational efficiency is measured for each method to assess feasibility in real-time applications. Traditional Image Quality Assessment Methods Subjective Assessment  Based on human perception, subjective methods involve rating image quality using mean opinion scores (MOS). MOS is commonly used in multimedia and video streaming services to gauge user satisfaction.  While reliable for human-centric applications, this approach lacks consistency and scalability. Variability among observers can introduce bias, reducing reproducibility.  The human visual system (HVS) has limitations in detecting subtle distortions that objective methods can quantify. For example, color fidelity assessment can be challenging without a standard reference.  Subjective evaluation often requires large-scale user studies, increasin