Selection of Features by a Machine Learning Expert to design a Color Image Quality Metric

Introduction With the first glance on an object, the human observer is able to say if its sight is pleasant for him or not. He makes then neither more nor less than one classification of the perception of this object according to the feeling gotten and felt in two categories: ”I like” or ”I don’t like”. Such an aptitude to classify the visual feelings is indisputably to put in relation with the inherent conscience of each human being. The conscience is related so that Freud calls ”the perception-conscience system”. It concerns a peripheral function of the psychic apparatus which receives information of the external world and those coming from the memories and the internal feelings of pleasure or displeasure. The immediate character of this perceptive function involves an impossibility for the conscience of keeping a durable trace of this information. It communicates them to preconscious, place of a first setting in memory. The conscience perceives and transmits significant qualities. Freud employs formulas like ”index of perception, of quality, of reality ” to describe the content of the operations of the perception-conscience system. Thus, perception is to be regarded as one of the internal scales of a process driving to an overall quality assessment of an object or an image. From a more pragmatic point of view, the quality of an image is one of the concepts on which research in image processing takes a dominating part. All the problem is about characterizing the quality of an image, in the same way done by a human observer. Consequently, we should dissociate the two types of measurements: 1) fidelity measurement and 2) Quality measurement. The fidelity measurement mainly allows to know if the reproduction of the image is faithful or not to the original one. In this case, the measurement set up calculates the distance between the two images. This distance numerically symbolizes the variation existing between the two reproductions of the image. The quality measurement is close to what a human observer naturally and instinctively does in front of any new work: it gives him an appreciation according to its conscience with respect to features. Thus, the study of the mechanisms allows to apprehend the internal scales used for the quality evaluation by a human observer became an imposing research field. Instead of using classical approaches to measure quality of images, one propose to mesaure the quality based on a learned classification process in order to respect the one of human observers. Instead of computing a final note, our method classifies the quality using the quality scale recommended by the UIT. This quality scale contains 5 ranks ordered from 1 (the worst quality) to 5 (the best quality). The selected class of the proposed method represents the opinion score OS. In that way, a machine learning expert, providing a final class number and its associated confidence probability, is designed. Furthermore, one are able to measure the importance of embedded features using by a human observer to judge the final quality with respect to the recommendations given by the UIT [1], where the human observers have to choose a quality class from a scale containing five notes. Those notes caracterize the quality of the reconstructed images. In that way, the human observers make then neither more nor less one classification.