Studies On Space-Domain No-Reference Image Quality Assessment

With the rapid growth of mobile imaging equipment,digital images have become an important carrier of information representation and interaction.However,digital images are likely to cause varying degrees of quality attenuation throughout the life cycle.These fades will result in loss of visual information,resulting in a sensory experience that does not satisfy the observer,and will also cause difficulties in subsequent image processing and analysis steps.The evaluation of image quality plays an important role in a large number of image/video processing and machine vision applications,including image compression,transmission,acquisition,display,enhancement,denoising,segmentation,detection,photo classification and so on.The quality of the image includes both fidelity and intelligibility,where fidelity measures the deviation between the image to be measured and the standard image;the intelligibility represents the ability of a human or machine to obtain relevant information from the image.Image quality evaluation is generally divided into subjective evaluation and objective evaluation.The objective of subjective evaluation is to obtain the subjective evaluation score of the human by the design experiment,and the objective of the objective evaluation is to design the calculation model accurately and automatically to predict the quality of the distorted image associated with human perception.Non reference image quality evaluation refers to the ability to evaluate the quality of the image to be measured without knowing any reference to the image information,as compared to the full reference and the semi-reference image quality evaluation in the objective evaluation.This is also the most challenging and difficult The largest objective evaluation method.In recent years,the rapid development of deep learning has led to the field of machine vision has made many remarkable breakthrough results,such as image recognition,image classification,target detection,target tracking,image segmentation,image super-resolution reconstruction,image denoising,Image quality evaluation and so on.Depth learning has greatly improved the performance of image quality evaluation field,which greatly promoted the development of this field.However,compared with traditional research methods,the theory of depth learning is not yet complete,and its performance analysis can only be carried out from experiment To verify the angle.The traditional research method is simple,the theory is complete,the biological correspondence function is clear,the training is flexible,the advantage of the small sample training makes it still has the vigorous vitality in the field.If you can extract the appropriate characteristics of human visual system information processing methods,the traditional method can still achieve excellent performance.The effects of network depth and convolution kernel number on the performance of Convolutional Neural Network(CNN)are discussed.Taking into account the characteristics of the human visual system to the depth of learning methods to add,this paper first proposed a joint gradient information and CNN non-reference image quality evaluation algorithm.The original CNN framework has been improved by introducing the relevant characteristics of the human visual system,such as the edge and contour information in the image.By calculating the gradient map of the segmented image,and then summing the gradient amplitude in the image block,the weight of each small block in the original image is obtained.The final obtained image quality fraction is the quality prediction score of each image block obtained by the CNN frame,multiplied by the corresponding small block weight value,and finally the mean value of the small block fraction obtained after weighting represents the mass fraction of the original image.The algorithm proposed in this paper can correct the image quality of each small block in the image,and not only improve the subjective and objective consistency of the image evaluation through the union gradient information only for the whole image.Secondly,a non-reference image quality evaluation algorithm based on CNN and Support Vector Regression(SVR)hybrid model is proposed.CNN framework has excellent ability of self-learning,and traditional artificial design features are usually a lengthy and time-consuming process,and the extraction of these features is mainly concentrated in the transformation domain,rarely directly from the original image block extraction.In this paper,CNN framework is used as feature extractor,and then SVR is used to replace the whole connection layer in CNN framework.The mapping relationship between quality sensitive feature and subjective evaluation score is established.The experimental results show that the hybrid model can improve the subjective and objective consistency of image quality evaluation.Thirdly,through the introduction of the depth of the new technology proposed in recent years,such as Dropout,local corresponding normalization,the experimental results show that the introduction of new technology to improve the image quality evaluation of the subjective and objective consistency.As a single image quality prediction can not completely describe the distortion of the image,but also to determine the type of image distortion,in the actual research and engineering practice has important significance.On this basis,a multi-task CNN framework is proposed to predict both the image quality and the distortion type.The experimental results show that the proposed multitasking CNN framework can predict the quality of the image and the type of distortion it belongs at the same time,and improve the subjective and objective consistency with respect to the single task CNN framework which only predicts the image quality.At the same time,based on the research of human visual system on information processing,this paper proposes a multi-channel,multi-directional and multi-scale natural Deviation Statistics(NDS)feature extraction framework,which is based on visual characteristics.The asymmetric generalized Gaussian distribution(AGGD)model is used as the NDS feature.Finally,the NDS features and images extracted by SVR are studied by using SVR.Subjective evaluation of the mapping between scores.The experimental results show that the proposed algorithm framework has good subjective and objective consistency of image quality evaluation,and has high time complexity,which can meet the requirements of real-time system.Finally,this paper extends the non-reference image quality evaluation method to the actual project requirements.In the actual defense project,the image compression detection system for different compression ratio image quality evaluation is a key measure of system performance indicators.The traditional methods are based on the reference image is known,the direct use of the classic reference image quality evaluation of the method to assess it.On the basis of this paper,the traditional idea is extended,and the CNN framework and the natural scene statistical model are used to evaluate the method without using the reference image.The experimental results show that the evaluation models studied by the two methods have good generalization ability and can evaluate the images in the original and the different compression ratios based on the original image effectively and accurately.