年轮图像早晚材的准确分割是树木年轮计数和间距测量的前提条件.为解决年轮本身生长的复杂性、采伐过程中的锯痕干扰、早晚材图像灰度差别较小等因素造成的分割难题,提出了一种基于随机森林( random forest,RF)算法的分类模型,可实现年轮图像的准确分割.首先,通过变换图像的颜色域空间,提取出样本图像在RGB、HSV和L?a?b?模型下的9个颜色分量,基于灰度共生矩阵提取样本图像的对比度、相关性、能量和熵的均值与标准差共8个纹理特征.然后,根据早晚材颜色与纹理特征的差异,基于随机森林算法构建像素分类器,实现年轮图像的早晚材的初步分割.为了提高分割图像的质量和准确度,对分割后的图像使用形态学方法消除孤立和黏连噪声,以得到最终分割图像.最后,将该方法与 K-均值聚类( K-means)算法和支持向量机(support vector machine,SVM)算法进行对比.结果表明:所采用基于RF算法的分类模型分割年轮晚材的正确识别率为95%左右,错误识别率在6%左右,图像分割效果明显优于其他两种算法.%The annual rings are the archives of trees, which record the age of trees and the growth environment of trees, including annual rainfall, temperature changes, and forest fires. The precise and high quality image segmenta-tion of early wood and late wood of tree ring is a necessary prerequisite for annual ring number counting and the annu-al rings width measurement. However, there are usually some difficulties that affect image data processing to obtain correct and useful trees’ information mentioned above. So, an optimal way was explored to solve the problems of im-age segmentation of tree rings caused by the complexity and variety of annual ring image, the interference of saw marks in cutting process and subtle grayscale difference between early wood and late wood. A kind of effective seg-mentation method based on Random Forest ( RF) algorithm, a flexible method of machine learning which could a-chieve desired result, was proposed. The specific implementation process is as follows. First of all, 9 color compo-nents of sample images data in RGB, HSV and L?a?b?model and 8 texture features based on Gray Level Co-occur-rence Matrix, namely mean values and standard deviations of contrast, correlation, energy and entropy from sample tree ring images, were extracted. And then according to the colors and textures difference between early wood and late wood, an image pixel classifier based on RF algorithm was constructed for preliminary segmentation of early wood and late wood. Thirdly, a preliminary segmentation image of tree ring image was obtained by inputting the extracted features of color and texture data into the trained image pixel classification model. However, the preliminary image segmentation results still had some isolated and adhesive noise, and a method named morphological was used to elimi-nate burrs and miscellaneous points, then a more accurate segmentation image was obtained. Eventually, the final im-age segmentation results of this image segmentation method based on RF algorithm were compared with those obtained by K-means and SVM algorithm. The results of the experiment confirmed that the accuracy rate of late wood recogni-tion based on random forest algorithm was about 95%, and the error recognition rate was about 6%, which was obvi-ously better than K-means algorithm and support vector machine algorithm.
展开▼
