基于Census变换和Boosting方法的陨石坑区域检测

Crater Region Detection Based on Census Transform and Boosting

提出了一种基于图像的陨石坑区域检测技术。首先,通过对原始图像进行Census变换并获取变换后的直方图;其次,利用主分量分析的方法对直方图空间进行降维压缩,并将压缩后的直方图作为特征向量,同时利用主分量分析的重构误差设定阈值构建第一层分类器;再次,基于Boosting原则对训练样本集进行选择,并利用支持向量机构建第二、三层分类器;最后,将测试样本依次送入这3层分类器确定该样本是否含有陨石坑。在实验过程中,通过对原始图像进行连续缩放,并遍历所有大小为20×20的子图像,以检测大小不一的陨石坑区域,并研究了虚警阈值对于最终检测结果的重要意义。相关实验研究表明:提出的方法可以从俯视摄像机拍摄的陨石坑图像中有效地检测出陨石坑区域。

An algorithm for image-based crater region detection is presented. Firstly, original images are transformed by Census transform and histograms of transformed images are obtained. Secondly, the principal component analysis is used for compressing histograms and the compressed histograms are used as feature vectors. Synchronously, the first classification is built by reconstruction error based on principal component analysis. Thirdly, train samples are selected based on Boosting principle, and the second and third classifications are built by support vector machine. Finally, test samples are detected by three classifications to confirm crater regions. In the experiment, different size craters are detected by reducing and magnifying original images. Research shows that the false alarm plays an important part in crater region detection. Experimental study demonstrates that the algorithm can detect crater region by images obtained from the downlooking camera.