No-Reference Natural Image/Video Quality Assessment of Noisy, Blurry, or Compressed Images/Videos Based on Hybrid Curvelet, Wavelet and Cosine Transforms

In this thesis, we first propose a new Image Quality Assessment (IQA) method based on a hybrid of curvelet, wavelet, and cosine transforms, called the Hybrid No-reference (HNR) model. From the properties of natural scene statistics, the peak coordinates of the transformed coefficient histogram of filtered natural images occupy well-defined clusters in peak coordinate space, which makes no-reference possible. Compared to other methods, HNR has three benefits: (1) It is a no-reference method applicable to arbitrary images without compromising the prediction accuracy of full-reference methods; (2) To the best of our knowledge, it is the only general no-reference method well-suited for four types of image filters: noise, blur, JPEG2000 and JPEG compression; (3) It has excellent performance for additional applications such as the classification of images with subtle differences, hard to detect by the human visual system, the classification of image filter types, and prediction of the noise or blur level of a compressed image. HNR was tested on VIVID (our image library) and LIVE(a public library). When tested against VIVID, HNR has an image quality prediction accuracy above 0.97 measured using correlation coefficients with an average RMS below 7%. Despite the fact that HNR does not use reference images, it compares favorably (except JPEG) to state-of-the-art full-reference methods such as PSNR, SSIM, VIF, when tested on the LIVE image database. HNR also predicts noisy or blurry compressed images with a correlation above 0.98. In addition, we extend our image quality assessment methodology to three video quality assessment models. Video-HNR (VHNR) uses 3D curvelet and cosine transforms to study the relation between the extracted features and video quality. Velocity-Video-HNR (V-VHNR) considers video motion speed to further improve the accuracy of the metric. Frame-HNR defines the video quality as the average of the image quality of each video frame. These metrics perform much better than PSNR, the most widely used algorithm.