Research On Stereoscopic Image Quality Assessment Based On Deep Learning

With the continuous development of virtual reality technology and multimedia technology in recent years,3D stereoscopic images have been widely used in enhancing visual communication and the visual experience in image systems.At the same time,people's requirements for image clarity and fidelity are also increasing.However,in the course of image processing,it is difficult to avoid distortion,which has an impact on human understanding and utilization of image information.Therefore,designing a system like the Human Vision System(HVS)that can evaluate the quality of visual information has significant research value and application prospects.Deep Learning(DL)is a research hotspot in the field of artificial intelligence.With the advent of big data,it has shown great advantages in areas such as image recognition and detection,speech recognition,and natural language processing.The Alpha Go Zero program which is based on deep learning is beyond the level of professional players.Compared with previous shallow neural networks,the essence of deep learning is to construct a neural network model with multiple hidden layers and use big data to train models to learn more useful features and more complex knowledge to improve classification and identification accuracy.Traditional image quality assessment methods need to design and extract features that can evaluate image quality based on the experience of researchers.The quality of the designed features directly affects the quality of the image quality assessment system.The image quality evaluation algorithm based on deep learning eliminates the tedious work of manually extracting features and can accurately predicts the image quality score.The main contents are as follows:(1)A non-reference stereoscopic image quality assessment(SIQA)algorithm based on deep convolutional neural network(CNN)is proposed.The model uses three channels as input,namely left view,right view and their disparity map of the raw color image patches.Each channel consists of a 12-layer depth network,and the local natural scene statistical features of the stereoscopic characteristics are gained by the multi-layer stacking of the convolution layer and the max pool.Finally,the feature vector learned from 3-channels is linearly combined,and the quality score is obtained by regression of the full connection layer.Experimental results show that this method has a better subjective perception consistency in several distortion types.(2)According to the special attention of the human visual system to the salient regions of the image,a stereoscopic image quality evaluation method based on the stereoscopic visual significance is proposed.The method obtains the color saliency map,the center region saliency map and the band-pass filter map from the cyclopean map,and then merges the three saliency maps to form the saliency map of the stereoscopic image.Finally,the saliency map and the left and right distorted images are input into the deep convolutional neural network to train the quality scores of the predicted stereoscopic image pairs.The experimental results show that this method has good effects on various types of distortion and is consistent with human subjective evaluation of image quality.3.In the human visual system,depth and disparity information also have an effect on the significance of stereoscopic images.A depth-perceived stereoscopic image saliency map algorithm is proposed.The algorithm combines the depth saliency map of a stereoscopic image with its 2D semantic significance information,and uses the generated saliency map as an input of a stereoscopic image quality assessment depth convolutional neural network to predict the stereoscopic image quality score.The experimental results show that the method has better subjective consistency than other methods reported in the literature.