【作者】 陈晨；

【导师】 孟繁杰；

【作者基本信息】 西安电子科技大学 ， 检测技术与自动化装置， 2019， 硕士

【摘要】 在信息技术日新月异的发展下,计算机视觉领域在新理论的支撑下飞速发展,开拓了新的研究方法和思路。图像作为人类生产生活中必要的信息来源,提供了丰富的知识,图像处理技术在网络电子化的21世纪发挥着重要的作用,也是人类社会创新进步的催化剂。图像质量是图像研究领域中最重要的环节之一,图像的质量为很多计算机视觉相关的研究工作提供了量化图像及视频性能的重要依据,图像是人类获取信息直接又高效的方式,虽然主观的质量评价分数由很多专业的观察者提供,具有准确的统计学意义,但是此方法需要耗费大量的人力物力,也有不能嵌入系统进行实时处理的问题,因此通过建立计算模型模拟人类视觉系统对图像质量进行评价的客观质量评估方法,在实际应用中具有非常重要的应用价值。本文提出了一种基于人类视觉系统的全参考图像质量评价方法,充分考虑了人眼视觉系统的相关特性,利用卷积神经网络计算图像的特征图,并结合图像的显著图及轮廓图的特征对图像进行有效的质量评价。随着深度学习技术的日益成熟,卷积神经网络等深度网络在图像领域的分类识别任务中产生了较好的效果,其中代表性的VGG-NET作为一种卷积神经网络,能够自动的提取到图像的特征,具有很强的鲁棒性。本文使用VGG-NET对图像特征进行提取,考虑到关于质量评价的图像库包含的数据量较小,首先在大数据集ImageNet进行自然图像的网络训练,然后将训练好的网络迁移到图像质量评价的数据库中进行图像特征图的计算。相关研究表明,人类在观察图像时往往更关注图像中感兴趣的区域,图像的显著性检测符合人眼视觉特性,强调了图像中的重点信息同时忽略了很多不被人眼注视的冗余信息。图像的亮度信息在人类视觉系统中非常重要,颜色信息能够一定程度的反应出人眼更注意的区域,因此本文对一种图像显著性检测的方法进行改进,对于颜色直方图的检测算法没有考虑亮度及距离信息的不足之处,本文提出了基于中心距离的颜色亮度显著性检测方法,通过对图像中亮度和颜色的计算得到参考图像和失真图像的显著图,基于图像颜色和亮度的显著性检测方法能有效地辅助图像质量的量化。同时因为图像的纹理轮廓根据图像的失真程度会产生一定的损失,本文使用Mean-Shift和高斯金字塔对图像的轮廓进行检测,生成图像的边缘轮廓图,作为图像质量评价重要的参考信息。本文将原始图像和这三种方式产生的特征图有机的结合起来,并将神经网络输出的特征图作为质量测量的权重,提出了一种符合人眼视觉系统特性的全参考图像质量检测算法。实验结果显示本文提出的算法和人类视觉系统有较高的一致性,相比于其他经典全参考的图像质量评价方法,本文提出的算法在评价指标上表现的更好。在TID,CSIQ等数据集上本文的算法都表现出与人眼视觉系统具有很高的一致性。更多还原

【Abstract】 Under the ever-changing development of information technology,the field of computer vision has developed rapidly under the support of new theories,opening up new research methods and ideas.As a necessary source of information in human production and life,images provide a wealth of knowledge.Image processing technology plays an important role in the 21 st century of electronic network and is also a catalyst for innovation and progress in human society.Image quality is one of the most significant aspects in the field of image research.Image quality provides an important basis for quantifying image and video performance for many computer vision-related research work.Image is a direct and efficient way for humans to obtain information,the subjective quality evaluation score which provided by many professional observers has accurate statistical significance.However,this method not only requires a lot of manpower and material resources,but also face the problem that it cannot work in the real-time processing system.Therefore,the objective quality assessment method which by constructing a computational model to simulate the human visual system has very important application value in practical applications.With the deepening of deep learning technology,deep networks such as convolutional neural networks have produced notable results in the image domain of the classification and recognition tasks.The representative features can be extracted automatically by VGG-NET which belongs to convolutional neural network.It is very robust in the field of compute vision.In this paper,a full reference image quality evaluation method based on human visual system is proposed.The most relevant characteristics of human visual system are considered.The convolutional neural network is used to calculate the feature map of the image,with the characteristics of the image and the contour map are combined to effectively evaluation quality of images.In this paper,VGG-NET is used to extract image features.Considering that the image library for quality evaluation contains a small amount of data,firstly,the natural image network training is performed in the big data set Image-Net,and then the trained network is migrated to the image quality evaluation.The calculation of the image feature map is performed in the database.Related studies have shown that humans tend to pay more attention to the region of interest in the image when observing the image.The saliency detection of the image conforms to the visual characteristics of the human eye,emphasizing the key information in the image while ignoring many redundant information that is not eye-catching.The brightness information of the image is very important in the human visual system.The color information can reflect the area that the human eyes pay more attention to.Therefore,this paper aims to improve the method of image saliency detection by combine the luminance feature with the detection algorithm of color feature.Insufficient brightness and distance information,this paper proposes a color brightness saliency detection method based on center distance.By calculating the brightness and color in the image,the saliency map of the reference image and the distorted image is obtained,based on the saliency of the image color and brightness.The detection method can effectively assist in the quantification of image quality.At the same time,because the texture contour of the image will have a certain loss according to the degree of distortion of the image,this paper uses Mean-Shift and Gaussian pyramid to detect the contour of the image and generate the edge contour map of the image as an important reference information for image quality evaluation.In this paper,the original image and the feature map generated by these three methods are organically combined,and the feature map output by the neural network is used as the weight of the quality measurement.A full reference image quality detection algorithm conforming to the characteristics of the human visual system is proposed.The experimental results show that the proposed algorithm and the human visual system have higher consistency.Compared with classical ful-reference image quality evaluation methods,the proposed algorithm performs better on the evaluation index such as SROCC,PLCC and RMSE.The algorithms in this paper on TID,CSIQ and other datasets also show high consistency with the human visual system.更多还原