Blind Image Quality Prediction Using Joint Statistics of Gradient Magnitude and Laplacian Features

Description

Existing BIQA models typically decompose an image into different frequency and orientation bands, and then extract statistical features from the decomposed coefficients to learn a quality prediction model. However, few BIQA models explicitly exploit simple image contrast features such as the gradient magnitude (GM) and Laplacian of Gaussian (LOG) responses, although LOG responses share similarities to human receptive field responses. Here we made the first attempt to use GM and LOG features to conduct high performance BIQA. 

To alleviate the effects of image content variations, we applied a joint adaptive normalization procedure to normalize the GM and LOG features and whiten the image data. Since GM and LOG features are not independent and the interaction between them can reflect local quality prediction on natural images, we proposed a simple index, called independency distribution, to measure the joint statistics of them.

The proposed BIQA models employ the marginal distributions and the independency distributions of GM and LOG, and they lead to highly competitive performance with many state-of-the-art BIQA methods in terms of quality prediction accuracy, generalization ability, robustness (i.e., across-database prediction capability) and computational complexity. Encouraged by the state-of-the-art BIQA results obtained in this paper, in future work we will investigate how to use the GM and LOG features for blind local quality map estimation, which is a very useful yet very challenging research problem.

ref :

Xue, Wufeng, Xuanqin Mou, Lei Zhang, Alan C. Bovik, and Xiangchu Feng. “Blind image quality assessment using joint statistics of gradient magnitude and Laplacian features.” IEEE Transactions on Image Processing 23, no. 11 (2014): 4850-4862.

Network traffic prediction by hybrid model (MLP and ARIMA)

Network traffic prediction by Elman neural network