一种数码成像设备视觉质量无参考评价方法的研究

Digital image is one important transport of information storages and transmission, and objective assessing the quality of images is now an important challenging task. Based on the key issues of “Non-reference visual quality assessment for digital imaging device”, we develop our research. Exploring the visual quality of digital imaging device, it is different from the visual distortion during the transmission, compression, storage and processing of image,and capturing the right and good quality image is gained more attention in imaging quality,thus in our view, the key factors which impacts the visual quality can be are divided into four ones: blur, backlight, color and exposure level. And based on the four factors, a new visual quality assessment model is proposed. In particular, we will research on out of focus blur and motion blur which happens in digital imaging device, and the distortion model of image structural features with the two kind of blur. The research result of this project will provide the technical support for many applications, such as the development of imaging devices, sorting the mass image data and evaluation of the image algorithms.

数字图像日益成为信息存储和传播的重要载体，客观图像质量评价是一个重要的研究课题。本项目围绕“数字设备成像视觉质量无参考量评价”这个关键问题展开研究。从数码成像设备成像质量的角度出发，不同于图像传输、压缩、存储以及处理造成的质量失真，成像质量更多关注是否拍摄到了正确和高质量的图像，因此本项目将影响人眼视觉质量的关键因素理解为：模糊、背光、颜色以及曝光度四个关键特性，从而建立新的视觉质量评价模型。特别的，还将研究数码成像设备成像中造成的失焦模糊和运动模糊，研究在这两种模糊失真下图像结构特征信息的研究模型。本项目的研究成果将为数字成像设备的研制，海量图像数据的筛选，图像算法的评测等提供技术保障。

数字图像日益成为信息存储和传播的重要载体，图像质量客观评价是一个重要的研究课题。本项目围绕“数字设备成像视觉质量无参考量评价”这个关键问题展开研究，不同于图像传输、压缩、存储以及处理造成的质量失真，数码设备图像视觉质量直接来自于图像成像阶段，所以对数字图像的关键特征的分析成为关键。针对图像质量关键特征，研究了图像频谱以及灰度直方图统计信息，提出了一种基于图像直方图的曝光程度检测方法；研究了关于柯西评估在图像识别中的应用，提出了一种子空间学习的方法用于人脸识别，并且应用在了RGB-D场景识别中；针对在线移动图像分类问题，提出一种基于稀疏相关熵的学习框架；针对深度学习在特征分析上的优势，重点研究了Deep two-View Feature Learning、Deep Multi-View Feature Learning、Generative Adversarial Network、Convolution Neural Networks，分别在图像识别和修复进行了应用。