基于机器学习XGBoost、引导滤波及矩的自由落体岩石块度在线检测

Online detection of free falling rock fragments based on machine learning XGBoost,guided filtering and moment

基于采石场获取岩石块度的自由落体图像特性,在自由落体中获取在线后置光岩石块度图像,基于此视频图像,设计了新的岩石块度在线检测方法.用引导滤波增强图像,减少因光照不均、目标运动过快或抖动引起的图像模糊;用基于力学矩的启发式多阈值分割算法对灰度图像二值化;将机器学习中的XGBoost理论和方法首次用于该领域,制定新的准则判定待检测目标是否为多块度粘连体;对检测到的粘连体的凸包区域,由最大凹点处进行基于凹凸点的多边形近似,在多边形轮廓上检测凹点,选取两分割点对.通过对300多幅图像的测试并与分水岭、模糊聚类、最小生成树及水平集等传统算法进行对比,结果表明,该方法对岩石块度粘连体具有良好的图像分割效果,精度达到95%.

Based on the characteristics of free falling image of rock fragments obtained from a quarry,a method of acquiring rear light image in a free falling stream was studied;a new online detection method of rock fragments was proposed,which mainly includes four algorithms using the guided filter to enhance the image to reduce the image blur caused by uneven illumination,fast motion speed or jitter.A heuristic multi threshold segmentation algorithm based on mechanical moment was studied to binarize the gray scale image.Based on the characteristics of fragment images,the XGBoost method in machine learning was applied for the first time in this area,and was utilized in this research field to develop new criteria for determining fragment touching clusters.The polygon approximation algorithm based on the concave convex point was carried out from the deepest concave point in the convex hull region of the detected cluster.According to the specially defined concave point criterion,the concave points were detected on the polygon contour,and two points were selected as a segmentation point pair.By testing more than 300 images and comparing several traditional algorithms,such as watershed,level set,fuzzy cmeans and minimum spanning tree,the results showed that the new method had a good image segmentation effect on the touching fragments,and its accuracy reached 95%.