|Journal of New Media |
No-Reference Stereo Image Quality Assessment Based on Transfer Learning
1Jiangsu Key Construction Laboratory of IoT Application Technology (Wuxi Taihu University), Wuxi, 214000, China
2Pactera Yuanhui Technology (Wuxi) Co., LTD., Wuxi, 214000, China
3Jiangnan University, Wuxi, 214000, China
*Corresponding Author: Lixiu Wu. Email: firstname.lastname@example.org
Received: 12 January 2022; Accepted: 26 April 2022
Abstract: In order to apply the deep learning to the stereo image quality evaluation, two problems need to be solved: The first one is that we have a bit of training samples, another is how to input the dimensional image’s left view or right view. In this paper, we transfer the 2D image quality evaluation model to the stereo image quality evaluation, and this method solves the first problem; use the method of principal component analysis is used to fuse the left and right views into an input image in order to solve the second problem. At the same time, the input image is preprocessed by phase congruency transformation, which further improves the performance of the algorithm. The structure of the deep convolution neural network consists of four convolution layers and three maximum pooling layers and two fully connected layers. The experimental results on LIVE3D image database show that the prediction quality score of the model is in good agreement with the subjective evaluation value.
Keywords: No-reference; stereo image quality assessment; convolution neural network; transfer learning; phase congruency transformation; image fusion
In 2009, the release of 3D movie Avatar brought a wave of movie watching across the world, creating a new record in the history of movie taking USD28 billion at the box-office globally, which amazed the world, according to the statistics of Motion Picture Association of America, 3D movie mushroomed by 50% each year, currently, more than half of movies provide 3D copies. Starting 2010, 3D of several brands had been launched in the mainland China in succession. On 1st January 2012, CCTV cooperated with many television sets in our country, jointly rolling out the first 3D trial channel of China. The reason why the 3D picture and video are so popular is mainly because the production of 3D videos applies the principle that the parallax angle that people use two eyes to observe objects can tell the distance, then stereoscopic vision is generated, which makes the audiences have the feelings that they are in the scene personally. Recently, 3D technology has been the foundation of innovation and new vantage point of product design, manufacture, management, marketing, consumption, and so on, 3D three-dimensional picture and video technique have been the trendy research field, among them include quality assessment of 3D image.
The algorithm of three-dimensional image quality assessment can be divided into three types according to degree of dependence that distorted image to original image: Full Reference (FR), Reduced Reference (RR) and No Reference (NR) image quality assessment. Some image quality database and subjective mean opinion score (MOS) are republished in accordance with a serial of suggestions of video quality expert group , meanwhile, the image evaluation methods based on various effective models are brought out. In the past, peak signal to noise ratio (PRNR) was deemed as the best image quality assessment method before structural similarity (SSIM)  was emerged, on the assumption that human visual perception system is easy to extract structure information of image, the document  verify SSIM can have better performance than traditional PSNR on the LIVE database. At present, the research on the flat image quality evaluation has made significant achievement, but the research regarding stereoscopic image quality evaluation has certain challenge. In the method of FR image quality assessment, such as the document [4–7], in which, You et al.  and others use the method of FR image quality assessment to the method of flat image quality assessment directly in the evaluation of stereoscopic image, but this kind of method cannot take the differences between 3D image and flat image into consideration, which cannot fully imitate the complicate vision perception mechanism of human, performance of evaluation is below the average. The document  puts forward one type of FR stereo picture quality assessment model on the foundation of binocular energy quality metric (BEQM) by studying the binocular competence, the influence of binocular visual perception on the stereo image quality assessment is considered when they are producing this model, so it is consistent with the subjective evaluation of stereo image, but it is not sufficient for this model to know the binocular visual perception, and this model is relatively complicated, which is not suitable for NR stereo image quality assessment. The document  puts forward a method of FR quality assessment method based on colorful stereo image, due to the complex of human vision system, this method has not had a reasonable vision model, in addition, FR stereo image quality assessment need gain all the information of reference image, so this method has certain limitation. The document  No reference quality assessment for stereoscopic images by statistical features. No-reference Stereoscopic Image Quality Assessment Based on Binocular Fusion and Binocular Rivalry  and No-reference stereoscopic image quality assessment based on deep feature learning . The document  put forwards a no-reference stereoscopic image quality assessment method based on wavelet transform. The document  put forwards a method of NR stereo image quality assessment based on in-depth learning, which firstly left and right views by Gabor filter, the statistical features of different sizes and directions as the monocular, then according to the binocular competence feature of HVS, merge the left and right views, extract the histogram with the fusion of direction and gradient as the binocular feature, through deep belief network (DBN) to train regression model, and at last, predict the quality scores of left view and right views and gain the quality score of the whole image by combining the above two scores.
Deep learning has been succeeded in image identification, voice processing, understanding of natural language, and other fields since the Hilton and Salakhutdinow were proposed in 2006. Using the deep learning can solve the problem of stereo image quality assessment, but the biggest obstacle lies on the mass data required in the course of training models. The requirement of such a large number of data is because the machine will encounter enormous parameter in the learning. However, the relation gained in one model training targeting at certain type of data can easily apply to different problems in the same field, which can be called transfer learning . This article is to transfer the quality assessment models of the flat image to the ones of stereo image, and achieved a great assessment indicator.
2 Image Fusion and Phase Congruency
2.1 Image Fusion
Image fusion is to merge complementary information of two and more than two images into one new image, fitting the visual perception of human more and more convenient for the image to have further processing and analysis. This article applies image fusion algorithm  of PCA (principal component analysis), which is a multidimensional orthogonal linear transformation based on statistical feature. In the fusion, firstly take the left and right views of stereo pictures as the matrix input, then transfer into one-dimensional vector X, Y, count covariance matrix of two according to Eq. (1), secondly, take count the eigenvalue and eigenvector of covariance matrix, finally by comparing the eigenvalue, confirm the fusion coefficient of image according to eigenvector, according to Eq. (2),
In the formula, X, Y respectively represent serialization vector of left and right views, n represents the length of vector.
According to the above calculation, select a random image pair from LIVE3D image library, carry out the left and right views with PCA fusion processing, the effect is shown in the Fig. 1:
We can see from (a), (b), (c) in the Fig 1, the image after PCA fusion is a little vague and double image visually, and other problems, this is because in the fusion of image, it not only collects the flat information of stereo image, but also collects the in-depth information specific for stereo pictures. For the convenience of extracting features of images, this article the left and right views of stereo image are combined into one image by PCA fusion method.
2.2 Phase Congruency
Phase Congruency (PC) is raised by Morrone et al.  and more people when they find that the image features are always happened at the big overlapping area in the research of extracting the image feature. This model is not affected by the change of contrast ratio of the image, and can have better image visual perception when the luminance change, which is more suitable for the visual mechanism of human. Its application in the image quality assessment also works well, such as in the document [16,17], they use the PC in the image quality assessment, among them,  is first to conduct PC conversion to distorted images, and get the images of PC, and images of PC with the largest and smallest covariance, then calculate the gradient entropy of these three images and refer to the gradient entropy and gradient mean value of images, finally through the training of GRNN, predict the quality score of images; the document  is first to measure the distorted degree of distorted images in the similarity of local subdomain through the PC of distorted images and reference images, use the mean value and extreme value of PC of images in each subdomain as the weighted coefficient PC of images, calculate the weighted mean of relevant coefficients of each subdomain, and get the prediction quality of images.
Given is the one-dimensional signal, and respectively represent even symmetric filter and odd symmetric filter of n scale, the convolution of it and I can represent:
So the calculated mode of PC of one-dimensional signal is as follows:
In it: is a relatively small positive number, to avoid the denominator to be zero, is one-dimensional Hilbert transformation of , it can also be used in the two-dimensional ones, the calculated mode of PC is as follows；
In the ; represents noise compensating factor; represents a very small positive number to prevent the denominator from being zero.
This article selects randomly one reference image and one distorted image from LIVE3D image library, we can see PC can shake off the not important background information in the image the effect is shown in the Fig. 2, remain the structure features of images well, it is easy to discover that structure outline of (d) is more complicated than that of (b) from comparing (b) PC of reference image with (d) PC of distorted image, with the reason that there are various distorted elements in the distorted images, such as blur, noise, JPEG and JPEG2K compression, etc.
3 Based on the Internet Frame of Transfer Learning
3.1 Transferring Learning
Transfer learning is to learn new field by using the knowledge you already had. The knowledge used for transferring mainly include the transfer of data, feature and model parametric. Among them, data transfer is to extract the data suitable for task two in the task one, and combine the other data aiming at task two and one to get training; the feature transfer is to use the same features in the old task and new task to the training and learning of new task; model parameter transfer, similar to feature transfer, which not only transfers the feature into the similar tasks, but also the model can be transferred, such as the document  transfer ImageNet network structure. Transfer learning can put the trained model in the small new dataset and similar tasks, which can be used in small-scale dataset, and reduce the training time as well, mostly important, it works well.
LIVE3D stereo picture dataset has 725 pairs of distorted images, which are typical small training samples, it is not easy to have good performance index with these samples to train deep Convolutional Neural Network (CNN). The lab used the expanded 80 thousand flat distorted image a few days ago to train a flat quality evaluation model of good performance in deep CNN. This article transfers this flat model in the stereo picture quality evaluation using the technique of transfer learning, continuing to train stereoscopic distorted image on the trained model parameter, change of the Solver parameters test_iter from 1000 to 500, learning rate base_lr change from 0.01 to 0.001.
3.2 Network Architecture of CNN (Convolutional Neural Network)
CNN is one of widely used models in deep learning, which is developed from multilayer perceptron, designing for image identification and other problems originally. But now CNN is used widely on the image and video fields, and sound signal, textual data, etc. CNN normally is comprised of several convolutions, and the image extract the local features of each image in convolution through filters of different convolution cores and bias, every convolution core can reflect a new image, then output result of filter in the convolution cores are activated by non-linear activation function, finally the pooling function are done for the result of activation function (i.e., down sampling). In most practical circumstances, choosing the features by hand is not stable to a large extent, and waste time and energy. CNN just overcome these weaknesses, the weight of CNN is similar to the traditional BP neural network, all of them use backward propagation, so CNN can be directly to deal with input image, avoiding complicated preprocessing of the images. The big advantage of CNN is the sharing of weights, shrinking the parameter of neural network, which can prevent overfitting and makes the model of neural network easier.
This article uses the Caffe frame , refers to AlexNet  network architecture, uses the model parameter that already trained on the flat library, the network models transferred in the article is a convolution neural network containing 4 convolution layers and 2 fully connected layers (FC layer), and use built-in Euclidean Loss Layer in Caffe to solve the regression problem the effect is shown in the Fig. 3. Euclidean loss calculation methods are seen in Eq. (8):
In it, represents loss value, represents prediction values of () as the shapes, represents the original value of (), N represents the number of images, C represents the number of image channels, H represents the height of image, W the width.
In this article we use Rectified Linear Units (ReLUs  to replace traditional Sigmoid and Tanh as for activation functions, the document  has certified that in the deep training of CNN, the use of ReLUs is many times faster than the use of Tanh.
4 Experimental Result and Performance Analysis
4.1 Experimental Database
This article selects the LIVE3D image database  Phase I and Phase II released by University of Texas at Austin, among them, Phase I database has 20 pairs of reference images, 365 pairs of distorted images totally, with each pair of the image containing left view and right view of symmetrical distortion. The distorted type includes JPEG compressed distortion, JP2K compressed distortion, white gaussian noise (WN), Gaussian Blur (BLUR) and fast fading (FF) distortion; Phase II database includes 8 pairs of reference images, the difference of Phase I and Phase II database is: Phase II distorted image library contains symmetrical distortion and asymmetrical distortion, and the left and right views are spliced together for storage. The distribution of LIVE3D image library can be seen in Tab. 1:
Select 80% of images from Phase I and Phase II distorted image libraries respectively as training set in the experiment of this article, 10% of verification set and 10% of test set, with no overlapping between test set and training set. The image block conversed by phase congruency as input in the course of training. The final prediction score of image quality is gained by obtain the average scores of all the subblocks of each image. For the convenience of processing data, this article normalize formula min-max for Difference Mean Opinion Score (DMOS) that the dataset presents, DMOS value is difference quality mean score of reference image and distorted image, the large the DMOS value, the large the degree of image distortion.
Among them, is the data after normalization, X is original value, and are the minimum value and maximum value of original value.
4.2 Performance Index of Quality Evaluation
In order to measure the property of objective image quality evaluation method, the objective prediction score and objective evaluation of images need to be compared, at present days, the correlation of objective prediction value and objective evaluation are detected by Linear Pearson Correlation Coefficient (LPCC) and Spearman Rank Order Correlation Coefficient (SROCC). See Eqs. (10) and (11):
In (10): and represent objective evaluation value and subjective prediction value of I image respectively, and mean value of objective evaluation value and subjective prediction value, N represents the number of images. LPCC evaluate the precision of prediction, that is degree of accuracy, value range is within [0, 1], the closer to 1, the higher the degree of accuracy is, vice versa, the lower.
In (11), N represents the number of images, represents the difference between the objective evaluation value and subjective prediction value of image. SROCC evaluate the monotonicity of prediction, the value range is within [−1, 1], The absolute value is close to 1, the better the monotonicity.
4.3 Experimental Result and Analysis
To verify the effectiveness of the methods in the article, compare the experimental results on the LIVE3D, relatively new methods of FR image quality evaluation method with NR ones, the Tabs. 2–5 represent result comparison of different methods FR and NR stereo picture quality evaluation on the Phase I and Phase II image library. The models in this article are mainly targeting at JPEG, JP2K, WN, Blur, FF and various distorted types, we can see the methods in the articles are better than other stereo image quality evaluation model from the Tabs. 2–5. We can find from the documents [4–7] and the documents [9–12] that these evaluation models are mostly aiming at distortion of single type, having bad effect on the 5 distorted types combined. In the real life, the distorted image cannot always be pointed out clearly which type of distortion can be, this article, through carry outing comprehensive training on these 5 distorted types of JPEG, JP2K, WN, Blur, FF, no need to extract the inspection figure, structure features, high-frequency energy, and others of stereo image, we can see LPCC is obviously excel other evaluation model overall. But the models have a better evaluation effectiveness of JP2K, WN, Blur, FF of distorted types than JPEG distorted types, excel other stereo image quality evaluation models by and large.
This article puts forward a NR stereo picture quality evaluation model based on transfer learning, integrate the left view with the right view through PCA, and applies training features of CNN and regression models among image DMOS values, which effectively increase the performance of image quality evaluation model. Different from traditional learning, this article trains the progression model of stereo picture by trained parameter in the flat image library through transfer learning, overcoming the disadvantages of extracting features by hand, poor generalization ability and others in traditional machine learning. The experimental result shows that the models in this article is superior to some existing stereo picture quality evaluation model, but they are not effective on the JPEG distortion type, need improving, additionally, as for the convolution neutral network parameter setting of in-depth learning need further study.
Acknowledgement: Thanks to the teacher of my team for their guidance in the process of completing this article.
Funding Statement: The authors received no specific funding for this study.
Conflicts of Interest: The authors declare that they have no conflicts of interest to report regarding the present study.
|This work is licensed under a Creative Commons Attribution 4.0 International License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.|