A bag of words description scheme for image quality assessment

• Autor(a) principal: Fernandes, Miguel Francisco Fidalgo
• Data de Publicação: 2016
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.6/7780
• topic: Bag of Words; Image Description; Image Quality Assessment; Machine Learning; Ssim; Svm; Domínio/Área Científica::Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática

Every day millions of images are obtained, processed, compressed, saved, transmitted and reproduced. All these operations can cause distortions that affect their quality. The quality of these images should be measured subjectively. However, that brings the disadvantage of achieving a considerable number of tests with individuals requested to provide a statistical analysis of an image’s perceptual quality. Several objective metrics have been developed, that try to model the human perception of quality. However, in most applications the representation of human quality perception given by these metrics is far from the desired representation. Therefore, this work proposes the usage of machine learning models that allow for a better approximation. In this work, definitions for image and quality are given and some of the difficulties of the study of image quality are mentioned. Moreover, three metrics are initially explained. One uses the image’s original quality has a reference (SSIM) while the other two are no reference (BRISQUE and QAC). A comparison is made, showing a large discrepancy of values between the two kinds of metrics. The database that is used for the tests is TID2013. This database was chosen due to its dimension and by the fact of considering a large number of distortions. A study of each type of distortion in this database is made. Furthermore, some concepts of machine learning are introduced along with algorithms relevant in the context of this dissertation, notably, K-means, KNN and SVM. Description aggregator algorithms like “bag of words” and “fisher-vectors” are also mentioned. This dissertation studies a new model that combines machine learning and a quality metric for quality estimation. This model is based on the division of images in cells, where a specific metric is computed. With this division, it is possible to obtain local quality descriptors that will be aggregated using “bag of words”. A SVM with an RBF kernel is trained and tested on the same database and the results of the model are evaluated using cross-validation. The results are analysed using Pearson, Spearman and Kendall correlations and the RMSE to evaluate the representation of the model when compared with the subjective results. The model improves the results of the metric that was used and shows a new path to apply machine learning for quality evaluation.