Image quality assessment for Digital documents

This thesis focuses on No Reference (NR) methods for Image Quality Assessment (IQA). A review of the IQA field is presented in Chapter 2; where the different IQA methods are described and classified. In particular, the application of IQA methods within a workflow chain is discussed. In Chapter 3 we focus on NR metrics for JPEG-blockiness and noise artifacts. It is in general assumed that subjective methods produce an actual estimate of the perceived quality while objective methods produce values that should be correlated with human perceptions as best as possible. From the analysis of the regression curves that correlate objective and subjective data we have found that in some cases the metric's predictions are not in correspondence with the subjective scores. After reviewing the available databases, we realize that the distortion ranges considered are not in general representative of real case applications. Therefore, in Chapter 4 the Imaging and Vision Lab (IVL) database is introduced. It was generated with the aim of assessing the quality of images corrupted by JPEG and noise. In Chapter 5 we approach the NR-IQA field by focusing on a classification problem. A framework based on machine learning classification is proposed that let us evaluate how images can be classified within different groups or classes, according to their quality. NR metrics are considered as features and the assigned classes are obtained from the psychovisual data. For the JPEG distortion case, the feature space of the classifiers is built using each NR metric as single feature and also a pool of eleven NR metrics. Classification within five and three classes was addressed. In the former case, the five classes are in correspondence to the five categories recommended by the ITU (excellent, good, fair, poor, and bad) when designing image quality experiments. In the latter case we were interested in classifying images as high, medium or low quality ones. The classifiers are trained and tested on different databases. The classifier obtained using the pool of metrics outperforms each single metric classifier. Better performance is obtained in the case of three classes. Considering an image as the combining of two signals, content and distortion, we note that the crosstalk between both signals influences both subjective and objective quality assessment. We address this problem in Chapter 6 where our working hypothesis is that regression can be improved if performed within a group of images that present similar contents in terms of low level features. The criteria chosen to divide the images in different groups is the image complexity. The proposed strategy consists on two steps: the images (of a given database) are first classified in three groups of low, medium and high complexity. In a second step, regression is performed within each of these groups separately. The strategy is tested for different NR metrics for JPEG-blockiness and noise artifacts, different databases are considered. Correlation coefficients are computed and statistical significance tests are applied. The gain in performance depends on the metric and distortion considered. Summarizing, the two main proposals of this research work, i.e. the classification approach that combines several NR metrics and the grouping strategy, are able to outperform the correlation between subjective and objective data for the case of JPEG-blockiness. Both strategies can be extended to consider other type of distortions.