工程科学学报，第40卷，第8期：996-1004,2018年8月 Chinese Journal of Engineering,Vol.40,No.8:996-1004,August 2018 DOI:10.13374/j.issn2095-9389.2018.08.014;http://journals.ustb.edu.cn 联合多种边缘检测算子的无参考质量评价算法 沈丽丽区，杭宁 天津大学电气自动化与信息工程学院，天津300072 ☒通信作者，E-mail:sll@u.cu.cn 摘要提出了一种联合多种边缘检测算子的无参考质量评价算法，同时考虑一阶和二阶边缘算子，避免了单一算子的局限 性.该方法首先将彩色图像转换为灰度图像，然后计算灰度图像的梯度，相对梯度以及LOG特征.本文所使用的特征分为两 部分，一部分提取相对梯度方向的标准差，另一部分利用条件嫡来量化不同特征之间的相似性和相互关系，并且考虑到人眼 特性进行多尺度计算，最后使用自适应增强(AdaBoost)神经网络进行训练和预测.在公共数据库LIVE和TD2OO8上进行实 验，结果表明新方法对失真图像的预测评分与主观评分有较高的一致性，能很好地反映图像质量的视觉感知效果，仅使用10 维特征，性能优于现有的主流无参考质量评价算法 关键词无参考图像质量评价：梯度：LOG特征：条件嫡：AdaBoost神经网络 分类号TN911.73 No-reference image quality assessment using joint multiple edge detection SHEN Li-i,HANG Ning School of Electrical and Information Engineering,Tianjin University,Tianjin 300072,China Corresponding author,E-mail:sll@tju.edu.cn ABSTRACT Before digital images become available to consumers,they usually undergo several stages of processing,which include acquisition,compression,transmission,and presentation.Unfortunately,each stage introduces certain types of distortion,such as white noise,Gaussian blur,and compression distortion,which may degrade the perceptual quality of the final image.Therefore,it is important to design an effective and robust image quality assessment method to automatically evaluate distortions in image quality.Im- age quality assessment is widely used in image compression,image deblur,image enhancement,and other image processing domains. In general,no-reference image quality assessment methods have profound practical significance and broad application value;hence,it remains the main focus of many researchers.At present,many image quality assessment methods extract features and predict image quality using single edge detection operations such as gradient or local binary pattern.However,it is difficult for a single edge detection operation to represent the whole perceptual quality of distorted images,and hence,their predictions may not be satisfactory.To elimi- nate the limitations of single edge detection operation,this paper proposes a new no-reference image quality assessment method based on a multiple edge detection operation.The paper considers first-order and second-order derivative information and utilize their similari- ty to predict image quality.The proposed method first converted color images to grayscale images,and calculated the gradient magni- tude GM),relative gradient magnitude RM),relative gradient orientation (RO),and Laplacian of Gaussian LOG)of the grayscale images.The feature vectors extracted from the maps were divided into two parts,where one part was the standard deviation of RO,and the second part utilized conditional entropy to quantify the similarity and relationship of GM,RM,and LOG.The images were naturally multiscale,and distortions affected the image structures across scales.Hence,all features at two scales were extracted:the o- riginal image scale and at a reduced resolution low-pass filtered and down sampled by a factor of 2).Lastly,an AdaBoost back-propa- gation network was used to train and establish a regression model to predict the image quality.The experiment of the proposed method 收稿日期：2017-08-22 基金项目：国家自然科学基金资助项目(61520106002,61471262)工程科学学报,第 40 卷,第 8 期:996鄄鄄1004,2018 年 8 月 Chinese Journal of Engineering, Vol. 40, No. 8: 996鄄鄄1004, August 2018 DOI: 10. 13374 / j. issn2095鄄鄄9389. 2018. 08. 014; http: / / journals. ustb. edu. cn 联合多种边缘检测算子的无参考质量评价算法 沈丽丽苣 , 杭 宁 天津大学电气自动化与信息工程学院, 天津 300072 苣 通信作者, E鄄mail:sll@ tju. edu. cn 摘 要 提出了一种联合多种边缘检测算子的无参考质量评价算法,同时考虑一阶和二阶边缘算子,避免了单一算子的局限 性. 该方法首先将彩色图像转换为灰度图像,然后计算灰度图像的梯度,相对梯度以及 LOG 特征. 本文所使用的特征分为两 部分,一部分提取相对梯度方向的标准差,另一部分利用条件熵来量化不同特征之间的相似性和相互关系,并且考虑到人眼 特性进行多尺度计算,最后使用自适应增强(AdaBoost)神经网络进行训练和预测. 在公共数据库 LIVE 和 TID2008 上进行实 验,结果表明新方法对失真图像的预测评分与主观评分有较高的一致性,能很好地反映图像质量的视觉感知效果,仅使用 10 维特征,性能优于现有的主流无参考质量评价算法. 关键词 无参考图像质量评价; 梯度; LOG 特征; 条件熵; AdaBoost 神经网络 分类号 TN911郾 73 收稿日期: 2017鄄鄄08鄄鄄22 基金项目: 国家自然科学基金资助项目( 61520106002, 61471262) No鄄reference image quality assessment using joint multiple edge detection SHEN Li鄄li 苣 , HANG Ning School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China 苣 Corresponding author, E鄄mail: sll@ tju. edu. cn ABSTRACT Before digital images become available to consumers, they usually undergo several stages of processing, which include acquisition, compression, transmission, and presentation. Unfortunately, each stage introduces certain types of distortion, such as white noise, Gaussian blur, and compression distortion, which may degrade the perceptual quality of the final image. Therefore, it is important to design an effective and robust image quality assessment method to automatically evaluate distortions in image quality. Im鄄 age quality assessment is widely used in image compression, image deblur, image enhancement, and other image processing domains. In general, no鄄reference image quality assessment methods have profound practical significance and broad application value; hence, it remains the main focus of many researchers. At present, many image quality assessment methods extract features and predict image quality using single edge detection operations such as gradient or local binary pattern. However, it is difficult for a single edge detection operation to represent the whole perceptual quality of distorted images, and hence, their predictions may not be satisfactory. To elimi鄄 nate the limitations of single edge detection operation, this paper proposes a new no鄄reference image quality assessment method based on a multiple edge detection operation. The paper considers first鄄order and second鄄order derivative information and utilize their similari鄄 ty to predict image quality. The proposed method first converted color images to grayscale images, and calculated the gradient magni鄄 tude (GM), relative gradient magnitude ( RM), relative gradient orientation ( RO), and Laplacian of Gaussian ( LOG) of the grayscale images. The feature vectors extracted from the maps were divided into two parts, where one part was the standard deviation of RO, and the second part utilized conditional entropy to quantify the similarity and relationship of GM, RM, and LOG. The images were naturally multiscale, and distortions affected the image structures across scales. Hence, all features at two scales were extracted: the o鄄 riginal image scale and at a reduced resolution (low鄄pass filtered and down sampled by a factor of 2). Lastly, an AdaBoost back鄄propa鄄 gation network was used to train and establish a regression model to predict the image quality. The experiment of the proposed method