Visually Optimized Supra Threshold Wavelet Image Compression

图像传输的增加，目前是通过互联网，未来是通过无线个人通信服务，在数字图像压缩方面提出了新的问题。通过高带宽的通信链路，几乎可以立即下载高质量的图像。但是，随着带宽的减少，下载时间急剧增加。在某个时刻，用户可以根据可用的图像信息来决定是继续传输还是停止传输，而不是等待整个高质量图像。如果传输在任何一点停止，理想情况下，用户应该拥有该数据量的最高视觉质量的图像。生成这样的图像需要理解人类视觉系统（HVS），特别是HVS如何感知低质量图像中的压缩伪影。然而，迄今为止，HVS的研究只集中在理解HVS如何感知高质量图像中的伪影。本研究的目的是量化具有可见伪影的小波编码图像中压缩伪影的感知。由此产生的算法将允许以比目前已知算法更低的比特率进行图像识别、对象识别和数据提取，因此成像任务将需要更少的数据传输或下载。提出的工作将实验量化人类对小波压缩图像中超阈值压缩伪影的敏感性。由于对可见伪影的灵敏度正在评估中，因此采用了自顶向下的测试，其中实际图像是扭曲的，而不是传统的自底向上的测试，测量对单个基函数的灵敏度。将均匀的量化噪声添加到各个子带中，观察器识别产生明显差异的步长。根据其特定的视觉特征选择的15幅图像和150多个主题将提供足够大的数据集，以进行一般特征描述，而不是产生特定于图像和观察者的结果。颗粒失真和过载失真的影响将被分别表征，这些实验结果将被纳入传统和嵌入式小波编码器的量化策略和压缩算法中。

CCR-9902796HemamiThe increase in image transmission, currently over the internet and in the future via wireless personal-communication-services, poses new problems with respect to digital image compression. High-quality images can be nearly instantaneously downloaded over a communications link with high bandwidth. However, as bandwidth decreases, the download time increases dramatically. At some point, the user may decide whether to continue or to stop the transmission based on the image information available, rather than waiting for the entire high-quality image. If the transmission is stopped at any point, theuser should ideally have an image of the highest possible visual quality for that amount of data. Generation of such images requires an understanding of the human visual system (HVS), and specifically how the HVS perceives compression artifacts in low-quality images. However, HVS research to date has focussed only on understanding how the HVS perceives artifacts in high-quality images. The goal of this research is to quantify the perception of compression artifacts in wavelet-coded images with visible artifacts. The resultingalgorithms will allow image recognition, object identification, and data extraction at lower bit rates than algorithms known to date, and hence imaging tasks will require that less data be transmitted or downloaded.The proposed work will experimentally quantify human sensitivity to supra-threshold compression artifacts in wavelet-compressed images. Because sensitivity to visible artifacts is under evaluation, a top-down test in which actual images are distorted is employed, rather than the traditional bottom-up test which measures sensitivity to individual basis functions. Uniform quantization noise is added to individual subbands and observers identify the step sizes that produce noticeable differences. Fifteen images selected for their specific visual characteristics and over 150 subjects will provide a data set large enough to make general characterizations rather than producing image- and observer-specific results. The effects of granular andoverload distortion will be separately characterized, and these experimental results will be incorporated into quantization strategies and compression algorithms, both for traditional and embedded wavelet coders.