Space Object Detection: Receiver Operating Characteristics for Poisson and Normally Distributed Data

摘要

Object detection algorithms typically implement a Likelihood Ratio Test (LRT) for determining if an image-pixel contains a star. Implementing a LRT requires prior knowledge about the statistics of the data, such as the mean and variance. Assuming the background noise follows a Gaussian distribution, the mean and the variance have to be calculated separately. If the background data follows a Poisson distribution, the mean is the only calculation needed, as mean and variance of the Poisson distribution equal each other. This paper will compare the possible detection improvements when using a Poisson assumption. Many star detection LRTs will use a windowing technique to limit the amount of background data that is being tested. Various window sizes will also be tested to determine possible detection improvements that can be realized.