Research On Evaluation Methods Of No-reference Stereoscopic Images Based On Convolutional Neural Network

In recent years,with the development of multimedia technology,stereoscopic image has attracted more and more attention.Stereoscopic image will degrade in the process of acquisition,compression,transmission and display.The quality of stereoscopic image will directly affect people’s viewing experience.Therefore,an effective stereoscopic image quality evaluation method is urgently needed.The quality of stereoscopic image is evaluated,and the practical performance of various image processing algorithms related to stereoscopic image is described.Deep learning,especially convolutional neural networks,has been rapidly developed in recent years,and has been well applied in the fields of image classification and object detection.The convolutional neural network simulates human brain cognitive system.It extracts deep feature information by deep mining of massive data,and solves the target in autonomous learning way,thus effectively solving related problems.Therefore,this paper proposes to construct a class of no-reference stereoscopic image quality evaluation algorithm using convolutional neural networks.The paper mainly carries out the following two aspects of work.Firstly,we propose a no-reference stereoscopic image quality assessment method based on cyclopean images.In order to better simulate the process of the human brain processing stereoscopic images,a cyclopean image algorithm is proposed that firstly fuses the left and right views and then processes them.The cyclopean images are overlapped and cut to serve as the input of the migration learning network;the migration learning network model converges faster than traditional convolutional neural networks and has better initial weights.At the end of this paper,we use weighted features to weight the image blocks to better simulate the visual salient features of human eyes.Then,we propose an end-to-end two-channel deep convolutional neural network(DCNN)through local to global regression for no-reference(NR)stereoscopic images quality assessment(SIQA).Firstly,in most deep learning based methods,they use the given subjective mean opinion score(MOS)or the differential MOS(DMOS)value to adjust the network parameters.But it is unreasonable,especially for the asymmetrical distortion stereoscopic image.To alleviate the problem,we propose to use feature similarity index(FSIM)to provide pseudo labels for the left and right view respectively,named local regression,so that the left and right channels are trained better.Then,we use the given DMOS to finetune the locally trained model parameters,named global regression.More specifically,we combine the high level features of left and right view images by concatenation layer,and then employ Squeeze-and-Excitation block(SE-block)to simulate the binocular rivalry and use two convolutional operations to simulate binocular integration in Human Visual System(HVS).Finally,the quality score of the image is output through the three fully connected layers.The paper is tested on the published LIVE3 D phase-I and LIVE3 D phase-II databases.The results show that the proposed method achieves better results than other methods both in symmetric and asymmetrical stereoscopic image databases,and keeps good consistency with human subjective perception.