Blind Image Quality Assessment Using Joint Features Of Spatial-frequency Domains In The Non-flat Region

As an important carrier of information transmission,the image plays an important role in the information exchange of current social.The amount of information depending on the quality of the image,the image quality assessment has received wide attention and the no-reference methods is becoming a hotspot with low cost,high real-time and no original image required.To quantify image degradation level,a blind image quality assessment method which uses joint features of spatial-frequency domains in the non-flat region of image is proposed and the method proposed is improved in this thesis.The main work and research results are as follows:1.As the low attention to the monochrome background area of the image,the segmentation of the image flat region from non-flat region are introduced to the measurement system.Ten image features which have important influence on the image quality are extracted from the non-flat region of images,which include the spatial domain and frequency domain features.By segmenting the non-flat region of the image,1.1028% predictive consistency and 2.3975% monotonicity of the quality assessment method are improved,and the prediction error is reduced by 32.7484%.2.Towards ameliorating the predicted model against the impact of the visual mask due to the difference size of features,the equalization layer which is on the basis of the existing artificial neural network is introduced and the equalization general regression neural network is proposed in this thesis.By introducing the equalization layer,the quality assessment method is improved by 6.068% of the prediction consistency and 5.4391% of the monotonicity,reducing the prediction error of 8.7737%.3.Based on the non-flat region of the image,the space-frequency domain features and the equalization general regression neural network,a new no-reference quality assessment method is proposed in this thesis.The experimental result demonstrates the no-reference quality assessment method proposed in this thesis can predict the quality of the image effectively and reliably,and the correlation between predicted score and human subjective score is 0.88.4.In order to predict the image score more accurately,improved schemes are proposed to the no-reference quality assessment method.By analyzing the regular patterns of the discrete cosine transform coefficient statistic,the change of the statistical law caused by the image distortion is described by the shape parameter of the generalized Gaussian distribution.Three image features which reflect the texture of image are increased.The distortion-classification method is introduced to increase image recognition of the system and select the appropriate distortion classifier,which improves the prediction consistency of 7.2492% and the monotonicity of 6.0031% and reduces the prediction error of 19.5543%.5.The predict performance of improved method is tested on Database Release2 from LIVE.The experimental result demonstrates that,compared with the original method,the improved method improves the prediction consistency of 8.3968% and the monotonicity of 6.1141%,reduces the prediction error by 29.6631%.Compared with the partial full-reference and noreference image quality assessment method,the improved non-reference quality assessment method is higher than the other approaches on accuracy,monotonicity and reliability.