Aiming at the solution of color image quality assessment, this paper proposes a full⁃reference color image quality assessment algorithm based on siamese network. We adopt the double convolution block as the feature extraction unit. The double convolution block extracts features on different dimensions of the reference image patch and the distorted image patch. Then, we fuse the extracted image features via the convolution and the pooling operations. After that, the final quality score is obtained through two levels of the fully connected layers and the average pooling layer. We performed training-testing experiments of the algorithm on the LIVE and the TID2013 datasets. Furthermore, the cross-database evaluation between the two databases is also conducted. Both achieved prominent results. The experiments show that the proposed algorithm has a good subjective and objective evaluation consistency for the evaluation of color image quality, indicating that the algorithm has exceptional performance and generalization ability. In addition, the algorithm has a simple architecture and fewer parameters. Since the network can be trained end-to-end, it is promising to exploit further usage other than image quality assessment. For an instance, the network can be leveraged as a feedback network for extended applications such as image denoising.
为此,本文提出了一种基于孪生网络的全参考彩色图像质量评价算法(image quality assessment based on siamese network, SNIQA)。该算法的输入为彩色RGB图像,参考图像与失真图像双支路训练参数共享,可以实现输入图像到输出图像质量分数的端到端训练。SNIQA网络参数适中,存储容量低,可以同时作为基准量化指标和反馈量加入到图像处理应用中。
本文所讨论的IQA数据集选用LIVE(laboratory for image & video engineering)[20]和TID2013(tampere image database 2013)[21],如表1所示。LIVE数据集提供了29张原始参考图像和779张对应的失真图像。图像失真类型包括JPEG2000压缩、JPEG压缩、白噪声、高斯模糊和快速瑞利衰退等。数据集提供位于之间的图像DMOS,DMOS越低,表示图像质量越好。TID2013数据集包含3 025幅图像,其中有25幅分辨率为的彩色参考图像,其余为参考图像对应的失真图像。失真类型包括加性高斯噪声、JPEG压缩、JPEG2000压缩、高斯模糊等24种,每种类型包含5个不同的失真等级。TID2013数据集提供来自5个国家的971位测试人员进行主观评分统计得到的所有测试图像MOS。MOS取值范围为,MOS越大,表示图像质量越好。
2.2 性能指标
为了客观衡量算法性能,本文选取3个通用的性能指标作为量化标准,分别为:Spearman 秩相关系数(Spearman’s rank-order correlation coefficient,SROCC)、Pearson线性相关系数(Pearson linear correlation coefficient,PLCC)和均方根误差(root mean square error,RMSE)[22]。这3个指标均衡量算法的客观评价和人类视觉系统(human visual system,HVS)主观评价的一致性,不同指标评价算法性能时各有侧重。RMSE评价算法输出和主观评价结果的平均误差;SROCC评价算法输出和主观评价的单调性;PLCC评价算法输出和主观评价的线性相关程度。一般而言,RMSE越小或SROCC/PLCC越大,代表算法性能越好;相反,RMSE越大或SROCC/PLCC越小,代表算法性能越差。
2.3 模型训练
在TID2013数据集上,随机选取15张参考图像及其对应的失真图像作为训练集,5张参考图像及其对应的失真图像作为验证集,剩下的图像作为测试集。模型搭建后以随机数初始化,随后实验的所有模型在训练集上训练,每完整遍历一次训练集(称为一个epoch)后,在验证集上评估效果,以便选择最好的模型。TID2013数据集中第15种图像失真类型为不同程度的局部块失真(local block-wise distortions of different intensity)。实验发现,对这一类失真类型的图像进行分块处理后,图像块之间的失真程度极其不均匀,有些图像块失真严重,而其他图像块无明显失真,因此该类图像不适合输入SNIQA模型中训练,需要在训练时剔除。
CHUJ, CHENQ, YANGX C. Review on full reference image quality assessment algorithms[J]. Application Research of Computers, 2014,31(1):13-22. DOI: 10.3969/j.issn.1001-3695.2014.01.003(Ch ).
WANGZ M. Review of no-reference image quality assessment [J]. Acta Automatica Sinica, 2015, 41(6):1062-1079. DOI: 10.16383/j.aas.2015.c140404(Ch ).
[5]
HOUW L, GAOX B, TAOD C, et al. Blind image quality assessment via deep learning[J]. IEEE Transactions on Neural Networks and Learning Systems, 2015, 26(6):1275-1286. DOI: 10.1109/TNNLS.2014.2336852 .
[6]
LIC F, BOVIKA C, WUX J, et al. Blind image quality assessment using a general regression neural network [J]. IEEE Transactions on Neural Networks, 2011, 22(5): 793-799. DOI: 10.1109/TNN.2011.2120620 .
[7]
BOSSES, MANIRYD, MULLERK R, et al. Deep neural networks for no-reference and full-reference image quality assessment [J]. IEEE Transactions on Image Processing, 2018, 27(1):206-219. DOI: 10.1109/TIP.2017.2760518 .
[8]
LINK Y, WANGG X. Hallucinated-IQA: No-reference image quality assessment via adversarial learning [C]// 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York :IEEE Press,2018: 732-741. DOI: 10.1109/CVPR.2018.00083 .
[9]
KANGL, YEP, LIY, et al. Convolutional neural networks for no-reference image quality assessment [C]// Proceedings of the 2014 IEEE Conference on Computer Vision and Pattern Recognition. New York :IEEE Press,2014: 1733-1740. DOI: 10.1109/CVPR.2014.224 .
[10]
PEIS C, CHENL H. Image quality assessment using human visual DOG model fused with random forest[J]. IEEE Transactions on Image Processing, 2015, 24(11): 3282-3292. DOI: 10.1109/TIP.2015.2440172 .
[11]
LIUX L, WEIJERJ, BAGDANOVA D. RankIQA: Learning from rankings for no-reference image quality assessment[C]// Proceedings of the 2017 IEEE International Conference on Computer Vision. New York :IEEE Press,2017: 1040-1049. DOI: 10.1109/ICCV.2017.118 .
[12]
GUK, ZHAIG T, YANGX K, et al. Deep learning network for blind image quality assessment[C]// Proceedings of the 2014 IEEE International Conference on Image Processing. New York:IEEE Press,2014: 511-515. DOI: 10.1109/ICIP.2014.7025102 .
[13]
LIM H, KIMH G, RA Y M, et al. VR IQA NET: Deep virtual reality image quality assessment using adversarial learning [C]// Proceedings of the 2018 IEEE International Conference on Acoustics, Speech, and Signal Processing. New York :IEEE Press,2018: 6737-6741. DOI: 10.1109/icassp.2018.8461317 .
[14]
RENH Y, CHEND Q, WANGY Z, et al. RAN4IQA: Restorative adversarial nets for no-reference image quality assessment [C]// Proceedings of the IEEE Inational Conference on Artificial Intelligence. New York:IEEE Press,2018: 7308-7314.
[15]
CHOPRAS, HADSELLR, LECUNY . Learning a similarity metric discriminatively, with application to face verification [C]// Proceedings of the 2005 IEEE Conference on Computer Vision and Pattern Recognition. New York :IEEE Press,2005: 539-546. DOI: 10.1109/CVPR.2005.202 .
[16]
SIMONYANK, ZISSERMANA. Very deep convolutional networks for large-scale image recognition[C]// Proceedings of the International Conference on Learning Representations. San Diego: ArXiv E-Prints,2015:1-14.
[17]
KIMJ, LEE S. Fully deep blind image quality predictor[J]. IEEE Journal of Selected Topics in Signal Processing, 2017, 11(1): 206-220. DOI: 10.1109/JSTSP.2016.2639328 .
[18]
GUANJ, YIS, ZENGX, et al. Visual importance and distortion guided deep image quality assessment framework[J]. IEEE Transactions on Multimedia, 2017, 19(11): 2505-2520. DOI: 10.1109/TMM.2017.2703148 .
[19]
MAK D, LIUW T, ZHANGK, et al. End-to-end blind image quality assessment using deep neural networks [J]. IEEE Transactions on Image Processing, 2018, 27(3): 1202-1213. DOI: 10.1109/TIP.2017.2774045 .
[20]
杨光义. 图像质量评价及其在图像去噪中的应用研究[D]. 武汉: 武汉大学, 2018.
[21]
YANGG Y. Image Quality Assessment and Its Application in Image Denoising[D]. Wuhan: Wuhan University, 2018 (Ch).
[22]
KINGMAD P, BAJ. Adam: A method for stochastic optimization[C]// Proceedings of the International Conference on Learning Representations. San Diego: ArXiv E-Prints, 2015:1-15.
ZHANGS F, ZHANGC, ZHANGT, et al. Review on universal no-reference image quality assessment algorithm. Computer Engineering and Applications, 2015, 51(19): 13-23+151. DOI: 10.3778/j.issn.1002-8331.1507-0044(Ch ).
[25]
SHEIKHH R, SABIRM, BOVIKA C, et al. A statistical evaluation of recent full reference image quality assessment algorithms [J]. IEEE Transactions on Image Processing, 2006,15(11): 3440-3451. DOI: 10.1109/TIP.2006.881959 .
[26]
PONOMARENKON N, JINL, IEREMEIEVO, et al. Image database TID2013: Peculiarities, results and perspectives [J]. Signal Processing:Image Communication, 2015,30(1): 57-77. DOI: 10.1016/j.image.2014.10.009 .
[27]
WANGZ, BOVIKA C, SHEIKHH R, et al. Image quality assessment: From error visibility to structural similarity [J]. IEEE Transactions on Image Processing, 2004, 13(4): 600-612. DOI: 10.1109/TIP.2003.819861 .
[28]
WANGZ, SIMONCELLIE P, BOVIKA C. Multiscale structural similarity for image quality assessment [C]// Proceedings of the IEEE Conference on Signals, Systems and Computers. New York:IEEE Press,2003: 1398-1402. DOI: 10.1109/ACSSC.2003.1292216 .
[29]
WANGZ, LIQ. Information content weighting for perceptual image quality assessment [J]. IEEE Transactions on Image Processing, 2011, 20(5): 1185-1198. DOI: 10.1109/ACSSC.2003.1292216 .
[30]
LARSONE C, CHANDLERD M. Most apparent distortion: Full-reference image quality assessment and the role of strategy [J]. Journal of Electronic Imaging, 2010, 19(1): 1-21. DOI: 10.1117/1.3267105 .
[31]
SHEIKHH R, BOVIKA C, DE VECIANAG. An information fidelity criterion for image quality assessment using natural scene statistics [J]. IEEE Transactions on Image Processing, 2005, 14(12): 2117-2128. DOI: 10.1109/TIP.2005.859389 .
[32]
SHEIKHH R, BOVIKAC . Image information and visual quality [J]. IEEE Transactions on Image Processing, 2006, 15(2):430-444. DOI: 10.1109/TIP.2005.859378 .
[33]
ZHANGL, ZHANGL, MOUX, et al. FSIM: A feature similarity index for image quality assessment[J]. IEEE transactions on Image Processing, 2011, 20(8): 2378-2386. DOI: 10.1109/TIP.2011.2109730 .
[34]
ZHANGL, SHENY, LIH, et al. VSI: A visual saliency-induced index for perceptual image quality assessment[J]. IEEE Transactions on Image Processing, 2014, 23(10): 4270-4281. DOI: 10.1109/TIP.2014.2346028 .
[35]
BAE S, KIMM. A novel image quality assessment with globally and locally consilient visual quality perception[J]. IEEE Transactions on Image Processing, 2016, 25(5): 2392-2406. DOI: 10.1109/TIP.2016.2545863 .
[36]
NAFCHIH Z, SHAHKOLAEIA, HEDJAMR, et al. Mean deviation similarity index: Efficient and reliable full-reference image quality evaluator [J]. IEEE Access, 2016,4: 5579-5590. DOI: 10.1109/ACCESS.2016.2604042 .