Designed and trained fuzzy systems in colour and imaging applications

Abstract

This thesis uses fuzzy logic to provide solutions to problems in the areas of image smoothing and display characterization. The research is divided into two main areas: image processing and display technology. In the realm of image processing, we present novel colour image denoising frameworks that integrate fuzzy inference systems (FIS) with eigenvector analysis. This approach addresses the challenge of removing Gaussian noise while preserving image quality and details. By converting images from the RGB domain to an eigenvectorbased space, we extract relevant local information to dynamically adjust the denoising process. The FIS uses this information to determine the appropriate intensity of smoothing, recognizing that homogeneous areas may require more aggressive treatment than detailed regions. The effectiveness of our methods is validated through various image quality metrics and visual comparisons against state-of-the-art techniques, demonstrating superior performance in noise removal while maintaining original image details. The second part of the thesis focuses on display characterization, investigating the application of trained fuzzy inference systems to accurately reproduce colours across LCD, OLED, and QLED displays. By utilizing device-dependent RGB data and device-independent XYZ coordinates, as well as the xyY colour space, we assess the performance of our models. Although the use of the xyY colour space showed lower performance, the focus remains on the XYZ colour space, where we develop models that not only achieve high accuracy but also offer interpretability, providing valuable insights into display behaviour. Although we include other machine learning methods like neural networks for experimental comparison, the focus remains on fuzzy models. We evaluate the effectiveness of our models using the ΔE00 visual error metric. Our findings demonstrate that the Fuzzy Modeling and Identification (FMID) method strikes an optimal balance between accuracy and interpretability, offering the potential for future display calibration and optimization strategies. We employ both designed fuzzy systems, based on expert knowledge in the image processing realm, and fuzzy systems trained from data, which are particularly useful when data is available instead of expert knowledge. We apply data-driven approaches specifically in display characterization. The concluding insights underscore the importance of fuzzy systems, not only for their accuracy but also for their interpretability, which offers valuable perspectives for future advancements in image processing and display characterization. This research contributes to the ongoing development of more efficient and accurate techniques for image enhancement and colour reproduction in modern display technologies.