基于双极偏好的滑动窗口参数优化方法

Sliding Window Parameter Optimization Method Based on Bipolar Preferences

滑动窗口是形状匹配中的常用检测方法,可以检测图像中不同尺度不同位置的多个物体。检测效果采用检测率和误检率来衡量。在传统的滑动窗口检测方法中,通常基于经验选取滑动步长和图像缩放规模这两个参数值,来获得较高的检验率和较低的误检率。然而这是典型的两目标优化问题,传统方法未考虑决策者对检验率与误检率的不同偏好。根据实际情况,考虑到决策者的正偏好(高检验率与低误检率)及负偏好(低检验率和高误检率),引入双极偏好控制策略,提出基于双极偏好的多目标粒子群算法(BPMOPSO)的滑动窗口参数优化方法。通过Leeds Cows图像数据集上图像的检测实验结果表明,与传统算法相比,该算法显著改善了滑动窗口检测中的检验率和误检率,且大大提高了运行效率。

One of the commonly-used detection methods in shape matching is the sliding window,in which multiple objects,different in size and position,can be detected.The detection performance is generally measured by detection rate and false positive rate.The two parameters in sliding window detection method,sliding step and the scale step are empirically selected for high detection rate and low false positive rate.However,those two factors can be formulated as a typical two-objective optimization problem,while the empirical selection shows no consideration over decision-makers＇different preferences regarding detection rate and false positive rate.Given the fact that decision makers＇ positive preferences are represented by high detection rate and low false positive rate,and negative preferences,by low detection rate and high false positive rate,the paper introduced the bipolar control strategy and then proposed a new way to optimize the sliding window parameters based on Bipolar Preference Multi-objective Particle Swarm Optimization （BPMOPSO）.The new method was applied in the detection experiment on Leeds Cows image datasets.The experiment results show that the performance of the new method is largely irnproved,i e.,the false positive rate declines sharply and the detection rate improves significantly,and that the efficiency of the algorithm ameliorates considerably as well.