【作者】 胡安洲；

【导师】 吴一戎； 张荣；

【作者基本信息】 中国科学技术大学 ， 信号与信息处理， 2014， 博士

【摘要】 面向图像失真的感知质量评价是图像处理领域中一项基础而富有挑战性的工作。可靠的图像质量评价标准可用于监控网络视频的服务状况、分配压缩算法中的比特率参数、校准图像采集和处理系统等。图像质量评价最准确的方式是通过人眼判读。然而主观评价的方法由于预算昂贵操作复杂,无法在实际中广泛开展；同时,传统的信号误差统计方法如PSNR和MSE等的评价结果并不能很好的符合人眼的主观感受。因此为了进一步推动图像处理技术的发展,需要研究基于人眼视觉感知的图像感知质量评价方法。本文针对图像质量评价领域中近年来兴起的工程学客观评价框架,结合传统仿生学评价模型中使用的HVS感知特性,探讨研究了工程学框架中的局部失真度量和失真特征汇集的方法设计,并提出了一种基于内容信息提取的新的工程学评价算法；同时本文还探究了主客观一致的SAR图像压缩视觉质量的评价问题,主要工作包括：(1)概括了生理学和心理物理学领域关于HVS的研究成果,总结了与生物视觉相对应的信号与信息处理技术,为设计高效的图像感知质量评价算法提供了依据。(2)基于HVS空间-频率的交互敏感性,设计了依照局部图像复杂度选择不同多通道评价模型的局部失真度量方法。首先,采用图像区域划分算法将图像划分为复杂区域和简单区域；然后对不同区域,分别使用空域和小波域SSIM算法度量局部失真；最后综合得到全图的质量评价结果。实验结果表明,综合使用HVS空域和频域感知特性的局部失真度量算法在评价准确性方面优于只单纯使用空间域或频率域HVS特性的改进SSIM算法。(3)针对工程学评价框架中的失真信息汇集问题,分别研究了局部质量特征在空域和频域的汇集模型。空域方面,首先提出一种考虑视觉侧抑制现象的结构熵权值算法；然后以结构熵为基础设计了模拟视觉信息非均匀采样特性的变尺度空间汇集模型。频域方面,本文通过引入机器学习思想,实现了基于人工神经网络的频谱多通道失真汇集模型。实验结果表明,在相同局部失真评价算法下,本文提出的汇集方法性能优于传统基于视觉注意机制和CSF加权的汇集方法。(4)根据图像的失真会造成图像内容的改变和丢失,提出了基于内容信息提取的普通图像质量评价方法。首先使用SIFT算子提取局部图像内容关键点,然后通过比较关键点的匹配与相似性评估图像整体和细节的内容失真,最后自适应加权图像内容失真度量和结构失真度量得到最终的评价结果。实验结果显示,提出的基于图像内容信息提取的评价方法在多个数据库上的评价准确性优于当前的最优工程学方法VIF。(5)针对将含噪图像有损压缩思想应用于SAR图像压缩的合理性问题,开展了主客观一致的SAR图像压缩视觉质量评价研究。首先通过设计组织主观评价实验获取了经4种压缩算法压缩后的300幅测试图像的主观质量；然后综合考虑SAR图像目视特性和HVS敏感性,提出了一种基于图像内容分解和支撑矢量回归的SAR图像压缩专用客观评价方法。实验结果表明,由于相干斑的存在,特定条件下有损压缩会导致SAR图像视觉质量出现一定的提高；同时提出的专用客观评价方法不仅在评价准确性上优于常见的保真度评价指标,而且能准确地预测压缩后SAR图像质量可能上升的特殊现象。综上所示,本文基于人眼视觉感知特性,结合图像信号表示、视觉注意机制、图像理解、机器学习理论和主观质量评价实验,提出了多种自然图像质量评价算法和专用的SAR图像压缩质量评价算法。实验结果表明。提出算法均具有较好的主观感受一致性,能够为图像处理算法的比较、改进和优化提供可靠的性能标准。更多还原

【Abstract】 Perceptual image quality assessment (IQA), which aims at the evaluation of quality degration due to image distortions, has become a fundamental and challenging task in the field of image processing. A reliable IQA method can be used to dynamically monitor image/video quality, choose the parameters in a coding system, and benchmark image acquisition and processing systems. The most accurate IQA method is the subjective viewing test. However, subjective evaluation s is as cumbersome in practice as it is expensive, complex, and time-consuming. On the other hand, the traditional mathematical statistic-based error measurements, such as PSNR and MSE, do not match well with subjective perception. Therefore, in order to further promote the development of image processing technology, it is need to develop IQA metrics based on human vision and perception.Based on the engineering assessment framework that received widespread attention for lower complexity and better performance in recent years and human visual system (HVS) peculiarities that used in the conventional vision bionics methods, this thesis investigates the design of efficient IQA algorithms by employing the two stages of the engineering framework:local distortion measure and distortion feature pooling, and proposed a new engineering algorithm by using the idea of content information extraction. In addition, this thesis also studies the perceptual quality assessment of synthetic aperture radar (SAR) image compression. The main work is detailed as follows:(1) The HVS peculiarities from the physiology and psychophysics researches have been summarized as well as the corresponding signal and information processing technologies. It provides the theoretical basis for designing effective perceptual IQA algorithms.(2) A local distortion measure algorithm has been designed by incorporating the interaction of spatial and spectral sensitivities of HVS. The general idea of the proposal is using different local quality measurements according to specific image region. First, a region partition algorithm is utilized to segment the image into complex and smooth areas. Then, a simple SSIM or a wavelet-based SSIM is chosen for each local region according to the result of region partition, to obtain local image qualities. Finally, the local image qualities are merged into a single quality score. Experiment results show that the proposal outperforms the improved SSIM algorithms only considering the spatial or spectral sensitivities of HVS.(3) The distortion pooling model has been studied for spatial and spectral domain, respectively. For spatial domain, first, a structural entropy weighting algorithm is developed by considering the visual phenomenon of lateral inhibition. Secondly, a structural entropy-based variable-scale spatial pooling model is proposed by mimicking the space-variant sampling nature of HVS. For spectral domain, an artificial neural network-based multi-channel evaluation pooling strategy is realized by introducing the machine learning techniques. Experiment results show that, in the case of the same local distortion assessment scheme, the proposed pooling models outperforms traditional models based on visual attention and CSF weighting.(4) A general IQA algorithm based on image content information extraction is proposed inspired by the justification that image distortions will lead to the change and loss of image content. First, the scale invariant feature transform (SIFT) is employed to extract the local content keypoints. Secondly, by comparing the feature matching results and feature similarity between the reference and the test images, the global and detailed content distortions are evaluated. Then, the content distortion measures and the conventional structual distortion measure are integrated into the final IQA result though adaptive weighting. Experiment results show that the proposal outperforms the top engineering method, VIF, on multiple publicly available databases.(5) The study of visual quality assessment of synthetic aperture radar (SAR) image compression in accord with subjective perception has been carried out. First, a psychometric study that contained four SAR image compression techniques and a total of300test images was carried out to obtain subjective evaluation results. Then, for SAR image compression, a special IQA algorithm based on image content classification and support vector regression is proposed by taking into account the characteristics of the SAR image and HVS. Experiment results show that the visual quality of the compressed images may be slightly better than the original images due to the presence of speckle. And the proposed objective method has a perfect and robust performance for predicting the perceptual quality of SAR image compression.To sum up, according to the perceptual properties of human vision, this thesis has presented several IQA algorithms for natural image and SAR image compression based on multi-channel decomposition, visual attention, image understanding, machine learning, and subjective assessment experiment, and the provided experiments have demonstrated their reliability and effectiveness.更多还原