基于局部Haar和PHOG特征的月球撞击坑综合检测方法

基于月貌图像的撞击坑的检测需要采用合理的特征选择和机器学习策略，我们提出了一种基于区域局部灰度和梯度分布特征与机器学习方法相结合的撞击坑检测方法．这种方法将Haar特征与AdaBoost结合，使候选撞击坑区域的定位更加快捷，采用局部区域的塔式梯度方向直方图（PHOG）与高效的支持向量机学习算法相结合的方法用来精确地对撞击坑候选区域进行分类．考虑到Haar特征数的繁多而采用AdaBoost作为特征提取和分类方法，并由于PHOG特征的每一项都对分类起作用，将撞击坑区域统一预处理为不含阴阳面的各向梯度向量基本一致的圆形模糊边界，使圆形撞击坑的正样本特征具备更多的稳定性．文中还讨论了几种特征和分类方法的机理和集成，以及参数调整对撞击坑检测的效率分析．

A Compositive detection method of lunar crater based on partial Haar and PHOG feature

It should adopt a reasonable strategy of selecting features, which are comparatively easy to achieve and can be exactly used for lunar crater detection, and the machine learning methods, which are of high efficiency or of high accuracy, to promote the efficiency of lunar crater detection. We present an integrated method based on the local gray level, the gradient distribution and the machine learning for higher recognition efficiency in high resolution lunar terrain images. The combination of the Haar feature and the AdaBoost classification method provides faster and higher accuracy of crater candidate area detection, and the combination of the local Pyramid Histogram of Oriented Gradients feature and the Support Vector Machine gains an accurate geometric orientation and verification for the candidate craters. It adapts the AdaBoost algorithm as the both feature selection and classification method considering the miscellaneous Haar features. Whereas every item of the Pyramid Histogram of Oriented Gradients feature having influence on classification, it preprocess all crater region image into that each oriented gradient has almost same bin＇s modulus without shadow and highlight pairs. The paper has discussed the mechanism and integration of features selection, classification methods, parameters adjustment and recognition efficacy analysis.