Research On No-reference Stereo Image Quality Evaluation Method Based On Deep Learnin

With the development of information communication industry and multimedia network,stereo vision technology is widely used and occupies a crucial position in the fields such as film and television,medical,military,aerial survey and remote sensing.With the deepening of applications,the requirements for stereoscopic image quality in various fields are also increasing.However,under the influence of environmental factors and hardware conditions,these stereoscopic images will inevitably suffer from various types of noise,resulting in the degradation of image quality and unable to meet the needs of their subsequent applications.Therefore,in order to measure the quality of stereoscopic images and provide objective quality assessment indicators and basis for production,application and academic research,it is particularly important to study efficient and accurate objective stereoscopic image quality assessment algorithms.The objective evaluation model is divided into three types: full reference,reduced reference and no reference.Since the first two models both rely on the original image information,and the reference amount is limited in the actual scene,the current research direction of the SIQA algorithm is mainly focused on the no reference evaluation field.For the NR-SIQA,the research content of this paper includes the following two aspects:1.A no-reference stereoscopic image quality assessment algorithm based on the fusion of monocular and binocular visual features is proposed.The algorithm combines a variety of image content features,including the traditional hand-extracted visual features such as color,texture,and edge,as well as the deep visual features automatically learned by convolution neural network models.By integrating the two features,a richer combined feature vector can be extracted from the monocular view.Meanwhile,the algorithm considers the difference between 3D images and 2D images,so the binocular perception feature of the disparity image is used as the depth information,which is then fused with monocular view features.Finally,the regression prediction of distorted image quality score is realized by the fully connected layer.The experimental results on the LIVE 3D-IQA databases show that the prediction performance of the proposed algorithm is highly consistent with the subjective perception of the human eye.2.A no-reference stereoscopic image quality assessment method based on interactive convolution neural network is proposed.The algorithm is inspired by the route of human visual perception.First,based on binocular sum and difference channel theory,the left and right views are synthesized into binocular images,and the low-level visual features are extracted and integrated into the output signal of the primary visual cortex through the shallow CNN model.Then a dual-stream deep CNN structure is designed to simulate the HVS visual pathway to characterize the complex process of depth feature perception of image signals in high-level visual cortex,which additionally adds interactive connections.The experimental results show that compared with some mainstream SIQA algorithms,the proposed algorithm has better overall performance and individual distortion type evaluation performance,which be used as a general and efficient NR-SIQA algorithm.