No-Reference Image Quality Assessment using Level-of-Detail

Traditionally, image quality assessment has involved the comparison of a corrupted image with an “original” or perfect version of that given image. In many practical settings, this perfect image is not available. This research introduces a new metric that measures the perceived visual quality of a single given image. Operating in this no-reference framework, the new method is ideally suited for real-world applications, including television monitoring and digital camera quality sensing. Much of the theoretical basis of this work centers on the notion of level-of-detail. Knowing whether an image is highly detailed or very smooth is important in both the detection and assessment of errors. At this time, three types of errors that commonly arise in practice are considered, namely noise, blur and compression. Each given image is assigned a score reflecting its perceived quality. Human test cases validate the new techniques.