Stereoscopic Image Quality Assessment Based On Multi-channel Convolutional Neural Network

With the rapid development of imaging technology,how to evaluate the quality of stereoscopic images efficiently and reliably has always been one of the key issues of stereo imaging technology.The stereo image quality assessment algorithm is generally divided into subjective evaluation and objective evaluation.The subjective assessment method is the process that testers give the stereoscopic image score by their subjective feelings.this method is time-consuming and laborious,and it is easily affected by the emotion of testers and test environment.The objective evaluation method is to achieve the evaluation of the quality of stereoscopic images through related algorithms,and has the characteristics of high efficiency and stability.This paper introduces stereoscopic image quality assessment theory,research status and development trend,and proposes objective evaluation model of stereo image quality based on convolutional neural network.Deep learning method has more advantages than some traditional machine learning methods,but it is limited to the size of the input image.Simple segmentation of input images may lose part of the image structure information.To solve the above problem,the principal component analysis and image segmentation are combined to preprocess the data in this paper,and the quality of the stereo image is evaluated by constructing the multi-channel convolutional neural network to automatically extract the image features.Firstly,two different preprocessing methods are used to obtain the different sizes data sets.Then,the appropriate channel structures are designed for different data sets,where each data set corresponds to two channels with the same network structure.Finally,a multi-channel convolutional neural network model is constructed by combining the six channels corresponding to the three data sets.The model has higher recognition accuracy and better stability.In this paper,400 stereoscopic images of different quality levels are provided by the wireless broadband communication and stereo imaging Institute of Tianjin University,and extended to 800 images.The appropriate 400 are selected as the training data set,and the other 400 are used as the test data set.The experimental results show that the accuracy of the proposed model on the test data set is 97.25%.In addition,the influence of different network structure,activation function and some optimization methods(dropout,LRN)on the performance of the model are analysed,and compared the model with other methods(ELM,SVM).What's more,the stability of modeling method is verified on LIVE stereo database,and the pearson linear correlation is 0.9505.