一种非线性的无监督对象检测模型

现有的无监督对象检测模型采用线性模型引入自顶向下的对象信息。由于对象的多变性及背景的复杂性,线性模型无法很好地刻画局部区域的对象信息。本文采用非线性模型学习引入对象性,同时采用了一种结合的策略引入对象的显著信息,以实现对象的检测。我们采用著名的Pascal图像库以提供广泛的对象样本,基于核的支持向量机则用于非线性模型的学习。实验结果,表明本文方法能够改善对象检测的性能。

A Nonlinear Model Based on Unsupervised Object Detection Method

The existing unsupervised object detection model uses the linear model to introduce the top-down object information.Since the variations of the objects and the complexities of the backgrounds,the linear model may not describe the ＂objectness＂ of the local region very well.In this paper,we used non-linear model to learn and introduce the object information.Meanwhile,we combined several saliency detection methods to obtain more accurate saliency map.The well-known PASCAL dataset was used to provide kinds of object samples,and the kernel based support vector machine was used for the non-linear model learning.The experimental results demonstrated that the non-linear model can improve the object detection performance.