No Reference Image Quality Assessment Based on DCT and SOM Clustering

The quality of images and videos is an important research topic due to their wide applications. The research should match the human subjective evaluation of quality. There are three types of objective image quality assessment: full-reference, reduced-reference, and no-reference. No-reference image quality assessment is the most realistic because distorted images often have no reference. In this article, we propose an algorithm that uses the discrete cosine transform (DCT) of patches to extract a feature vector. We apply the Self-Organizing Map (SOM) at two levels. In the first level, we classify the patches into three types: good quality, noisy, and blurred. In the second level, we further classify the noisy and blurred patches into weak and severe distortions using two separate SOMs. Then, we used a straightforward neural network with supervised back-propagation to adjust the number of distorted patches of five classes in one image to assign a quality score. Our method of training Self-Organizing Maps (SOMs) with reference, noisy, and blurred image patches allowed us to estimate the quality scores of other types of distortions that matched subjective scores equally well. The Spearman Rank Order Correlation Coefficient (SROCC) performance of our method, when measured against the subjective scores of degraded images that were used in training, lies in the range of 0.88-0.92. Similarly, for other degradation types that were examined (10 in total), the SROCC performance is in the range of 0.76-0.89, which is higher than other methods. The experiments show that our scores are consistent with the subjective scores.