基于高斯混合模型的叶片检测分割算法

摘要

In order to tackle the problems to segmentation of leaf images arising from variation of illumination and appearance of a leaf as well as cluttered background,a combined segmentation method was presented.The method used a sliding window scanning method to detect the blades in the image at a plurality of scales.The initial region of the center of the leaf was used as the initial foreground,and the area outside the leaf window was used as the initial background,and the initial probability model was established for the foreground and background with the Gaussian mixture model (GMM).The segmentation was carried out by the iterative method in which the standard Graph Cut,combing with GMMs which was gotten in previous round,was used to segment the leaves in each round,and the results to update GMMs of the foreground and background.For leaves detection,HOG feature was exploited which could describe the appearance and shape of a leaf.To solve challenges of variations of appearance and viewpoint to accurate detection,we adopted a strategy of multiple-subcategory classifiers for the leaves.We took the grape leaf as an example to evaluate the performance of the proposed segmentation method.The experimental results showed that our method worked very well on condition of aforementioned challenges,exhibiting robustness and accuracy.Furthermore,the proposed method can fulfill auto-segmentation completely.%为解决光照变化、叶片自身表观变化和复杂背景对植物叶片图像准确分割所造成的困扰,提出一种组合式分割方法.该方法在多个尺度上采用滑动窗口扫描方式检测图像中的叶片;对检测到的叶片区域中心区域像素为初始前景,而叶片窗口之外的区域为初始背景,用高斯混合模型(GMM)分别对前景和背景建立初始概率模型;采用迭代法完成叶片分割,在每一轮迭代中,用标准的图割算法和上一轮GMM模型分割前景和背景,根据新的分割结果重新估计前景和背景的GMM;迭代过程在能量函数收敛时结束.叶片检测时,以能描述叶片的外观和形状的HOG特征为检测依据;为了应对实际叶片图像中叶片形态、角度变化较多的挑战,采用多子类分类器策略.以葡萄叶片为例,用该方法进行了分割实验.结果表明,该方法对上述复杂条件下叶片图像的分割具有较好的鲁棒性和较高的精度,能实现分割过程的完全自动化.