Perceptual Error Logarithm (PEL): um eficiente e eficaz método analítico para avaliação da qualidade de imagem com referência completa

Abstract

The growing popularity of multimedia applications has fostered investments in technologies and services to ensure the processing, transmission, and storage of large volumes of data, notably high-quality images and videos. This guarantee required by such multimedia applications is commonly supported by techniques such as super-resolution (SR), compression, restoration, recovery, and object detection in images and videos. Although these techniques have received considerable attention from the academic and industrial communities recently, their processes often introduce artifacts that significantly degrade the quality of images and videos. To mitigate such degradations, researchers have been using perceptual methods for image quality assessment (IQA). However, it is important to emphasize that one of the great challenges currently existing is the development of perceptual methods that are simultaneously efficient and effective. In this sense, this thesis proposes an efficient and effective perceptual analytical method for evaluating the quality of full-reference images (FR-IQA), called the Perceptual Error Logarithm (PEL). Initially, the PEL method creates luminance and chrominance maps, which are generated after the reduction and spatial conversion of colors applied to the input images. The PEL method, through the use of chrominance and luminance maps, generates simple absolute difference maps, as well as special feature maps. The special maps are derived from perceptual feature extractions applied to gradient magnitude maps. These maps are generated immediately after the application of local standard deviation filtering to the luminance maps. Finally, to calculate the PEL score, a logarithmic function is applied to the perceptual error map, which combines the simple and special maps, in addition to adjusting this combination based on the weight of the superpixel similarity and local energy maps. Experimental results and discussions in eight image databases prove that the effectiveness of the proposed method, according to the application of statistical evaluation metrics such as Spearman Rank-Order Correlation Coefficient (SROCC), Kendall Rank-Order Correlation Coefficient (KROCC), Pearson Linear Correlation Coefficient (PLCC), and Root Mean Square Error (RMSE), is much more consistent with subjective evaluations than all nineteen IQA methods in the literature compared. Furthermore, the proposed method turns out to be efficient due to its low computational cost required to perform the average processing on all images. Thereby, PEL can be used as a method for FR-IQA or even as a loss function to support various analytical techniques used in multimedia applications or machine learning-based models, such as Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs), focused on image and video.