CIF:Small:Visual Information Measures for Task-Based Imaging Applications

许多数字图像不是为了文件证据或存档而获取的，而是最终由人类用于评估情况，做出决定或得出结论的中间信息。通常，图像不需要提供“完美的图像质量”，以便人类成功地执行任务-例如，考虑对随身行李的安全检查。许多任务可以很容易地执行的图像，将被认为是严重退化，在纯粹的美学意义上。 然而，迄今为止，图像评估算法主要集中在评估美学质量上。 这些算法通常是为相对高质量的图像而设计的，它们在高度退化的图像上表现不佳，这些图像表现出较低的美学质量，但仍然对人类执行的任务有用。该表征是根据图像的属性而不是根据高级视觉对图像的响应来执行的，并且是基于大量的主观实验，该主观实验量化了高度退化图像的感知效用并确定了图像的识别阈值-最大程度退化的图像仍然允许识别。 开发了一种实用措施，其将原始图像和失真图像作为输入，并输出量化图像中相对于识别阈值（在低端）和原始或原始视觉无损表示（在高端）的退化量的距离。 通过将该措施集成到成像应用程序（包括压缩和增强）中，证明了该措施的使用。

Many digital images are not acquired for documentary evidence or for archival use, but instead are intermediate pieces of information to be ultimately used by a human to assess a situation, make a decision, or reach a conclusion. Often, the image need not provide "picture-perfect quality" in order for a human to successfully perform a task - consider security screening of carry-on baggage, for example. Many tasks can be easily performed with images that would be considered to be severely degraded in a purely aesthetic sense. To date, however, image assessment algorithms have been primarily focused on evaluating aesthetic quality. These algorithms are typically designed for relatively high-quality images, and they do not perform well on highly degraded images which exhibit low aesthetic quality but remain useful for human-performed tasks.This research characterizes the suitability of an image for recognition and the perceived utility in terms of conveying information about content. The characterization is performed in terms of properties of the image, rather than in terms of the response of higher-level vision to the image, and is based upon extensive subjective experiments quantifing the perceived utility of highly degraded images and identifing recognition thresholds for images - maximally degraded images that still allow recognition. A utility measure is developed which takes as input the original image and the distorted image and outputs a distance quantifying the amount of degradation in the image relative to both the recognition threshold (at the low end) and to the original or a visually lossless representation of the original (at the high end). Use of this measure is demonstrated by integration this measure into imaging applications including compression and enhancement.