Blind image quality assessment using a general regression neural network
C.Li, A.Bovik and X.Wu,IEEE Transactions on Neural Networks, vol 22, no 5, pp:793-799,2011
This paper also deals with assessing the quality of the image without having a reference image. It is done by using a general regression neural network. Four features are used for calculating QA.
(I) MPC - Mean value of the phase congruency image
(II)EPC- Entropy of the phase congruency image
(III)EDIS - Entropy of the distorted image
(IV) MGDIS - Mean of the gradient magitude of distorted image.
The quality of the image is calculted by approximating functional relationship between these features and subjective mean opinion scores using General regression Neural Network(GRNN).
The GRNN architecture consists of 4 layers: input layer(the four features act as input),pattern layer, summation layer and the output layer. The number of inputs is equal to the number of features. And the units in the pattern layer represents training pattern. The summation layer includes 2 units, the first unit sums all the outputs from the pattern layer and is equal to yi. While the second unit is equal to unity. The output unit computes the two outputs of the summation layer yielding the predicted value of the dependent feature.
This method tests 5 types of distortion including JPEG, JPEG 2000, white noise, Blur and fast fading. Differential mean opinion squares(DMOS) are obtained for each distorted image. The algorithm is evaluated by using three performance measures : Spearman rank order correlation coefficient(SROCC) which measures the prediction monotonicity of the quality index, Linear Correlation Coefficient(LCC) and RMSE.
(I) MPC - Mean value of the phase congruency image
(II)EPC- Entropy of the phase congruency image
(III)EDIS - Entropy of the distorted image
(IV) MGDIS - Mean of the gradient magitude of distorted image.
The quality of the image is calculted by approximating functional relationship between these features and subjective mean opinion scores using General regression Neural Network(GRNN).
The GRNN architecture consists of 4 layers: input layer(the four features act as input),pattern layer, summation layer and the output layer. The number of inputs is equal to the number of features. And the units in the pattern layer represents training pattern. The summation layer includes 2 units, the first unit sums all the outputs from the pattern layer and is equal to yi. While the second unit is equal to unity. The output unit computes the two outputs of the summation layer yielding the predicted value of the dependent feature.
This method tests 5 types of distortion including JPEG, JPEG 2000, white noise, Blur and fast fading. Differential mean opinion squares(DMOS) are obtained for each distorted image. The algorithm is evaluated by using three performance measures : Spearman rank order correlation coefficient(SROCC) which measures the prediction monotonicity of the quality index, Linear Correlation Coefficient(LCC) and RMSE.