Designing a reduced reference objective image quality metric under bandwidth constraints of the side-channel

In this thesis a reduced-reference (RR) metric for Image Quality Assessment is designed by transforming the full-reference (FR) structural similarity index (SSIM) into a RR metric. Such RR quality metrics are required when the bandwidth of the transmission channel is limited, which is the case for most applications. The research is based on the partial reconstruction of a modified reference image from some DCT coefficients in order to predict the quality using the SSIM at the receiver side. Different techniques have been studied and incorporated, such as the random selection of features and the localization of impairments in the DCT domain. The block-DCT has been proven to be an adequate domain to localize impairments and the final metric combines this technique with a random selection. The constructed RR quality metric is validated on three different databases (LIVE, CSIQ, and TID). As it turns out, the new metric reduces the required side information by a factor of 200 maintaining the performance of the original SSIM. Furthermore, the validation tests showed the importance of well-designed databases to minimize the bias of the test results. Current available databases do not focus enough on non-uniformly distributed impairments and will need to be extended in the future to improve the reliability of the validation.