ANFIS Tuned No-Reference Quality Prediction of Distorted/Decompressed Images featuring Wavelet Entropy

Abstract

Assessing quality of distorted/decompressed images without reference to the original image is difficult due to vagueness in extracted features and complex relation between features and visual quality of images. The paper aims at assessing the quality of distorted/decompressed images without any reference to the original image by developing an adaptive network based fuzzy inference system (ANFIS). First level Haar approximation entropies of test images from LIVE database and region based features extracted from the benchmark images are considered as inputs while mean opinion score (MOS) based quality of the images used as output to the fuzzy inference system (FIS). The input-output variables of the FIS are expressed using linguistic variables and fuzzified to measure the vagueness in extracted features. Takagi-Sugeno-Kang (TSK) inference rule has been applied to the FIS to predict the quality of a new distorted/decompressed image. The FIS has been trained to tune the parameters of the membership functions of the fuzzy sets that assess quality of the image more accurately. Quality of decompressed and various noise incorporated distorted test images are predicted using the proposed method producing output comparable with other existing no reference techniques. Results are validated with the objective and subjective image quality measures.