ES Postdoctoral Associate: Image Compression and the Evaluation of Quality

小行星9405058 有损图像压缩方法，有利于图像存储和通信的成本和时间效率，不完美地重建原始图像。 评价所得到的图像质量是重要而复杂的。 最近的算法执行同时压缩和分类，提出了进一步的挑战，确定图像质量。 质量通常通过对观看者评级的统计分析来主观地量化，并且通过信噪比或通过模拟图像的实际使用来客观地量化。 由于收集模拟实际使用的观看者分数是昂贵且耗时的奋进，因此期望找到与更昂贵的质量测量强烈相关或准确预测更昂贵的质量测量的图像质量的可计算测量。 更好的可计算失真测量将允许更好地筛选候选编码技术，并且可以通过将它们结合到用于设计代码的优化算法中来提供改进的代码设计。 该研究将寻求开发统计方法，以量化图像质量的可计算测量在预测从人类观察者昂贵获得的测量中的有用性。 这些方法将基于拟合一般线性模型，并使用现代计算机密集型统计样本再利用技术评估重要性。 将实施各种基于感知的可计算测量的变化，并比较它们的预测能力。 这些可计算的失真和统计技术的措施，然后可以应用于评估和设计优化的矢量量化算法的压缩和联合压缩/分类目前正在开发。 ***

9405058 Gray Lossy image compression methods, useful for cost and time efficiency of image storage and communication, imperfectly reconstruct the original image. Evaluating the resulting image quality is important and complex. Recent algorithms perform simultaneous compression and classification, presenting a further challenge for determining image quality. Quality is typically quantified subjectively by statistical analyses of viewer ratings, and objectively by signal-to-noise ratios, or by simulation of the actual use of the images. As gathering viewers scores of simulating actual use is an expensive and time-consuming endeavor, it is desirable to find computable measures of image quality that strongly correlate with or accurately predict the more expensive quality measures. Better computable distortion measures will permit better screening of candidate coding techniques and may provide improved code designs by incorporating them into the optimization algorithms used to design the codes. The research would seek to develop statistical methods to quantify the usefulness of computable measures of image quality in predicting the expensively obtained measures from human observers. The methods would be based on fitting general linear models, and evaluating significance using modern computer-intensive statistical sample reuse techniques. Variations on a variety of perceptually based computable measures and compare their predictive power will be implemented. These computable measures of distortion and statistical techniques could then be applied both for the evaluation of and design optimization of vector quantization algorithms for compression and joint compression/classification currently being developed. ***