基于ROC曲线的目标识别性能评估方法

A New Performance Evaluation Method Based on ROC Curve

在自动目标识别(ATR)领域,评估目标识别算法性能的指标常用的有分类准确度、精确度、检测概率、混淆矩阵等,但这些指标都存在固有的局限性,对类别先验概率不具有稳健性。近几年来采用的基于雷达接收机工作特性曲线即ROC曲线的评估方法得到的评估结论不敏感于类别先验概率,从根本上克服了以上指标的缺陷。同时,该评估方法可以在错误分类代价未知的情况下进行,并能对识别算法进行多门限评估,因而在分类器识别算法性能评估中得到了广泛应用。文中首先叙述了ROC曲线和ROC曲线建立的方法,然后详细论述了基于ROC曲线的评估方法中常用的性能评估指标。

The most commonly used evaluation measures in automatic target recognition（ATR） are classification accuracy（CA）, precision, the probability of deteetion（PD）, confusion matrix and so on. However, all of these measures have the shortcomings inherently. They are not robust to the prior probability of classes. Recently the ROC curve evaluation method is widely used an＇d developed, because the evaluation conclusions based on the ROC curve is insensitive to the prior probability of classes and it overcomes the defects of the previous commonly used evaluation measures. It can be performed even without knowledge of the cost of incorrect classification, and most attractively it can give a multiple threshold evaluation. We firstly talk about the concept of the ROC curve and the way to plot the curve, and then talk about the measures based on the ROC curve in detail.