Automatic Segmentation for Plant Leaves via Multiview Stereo Reconstruction

Abstract

This paper presented a new method for automatic plant point cloud acquisition and leaves segmentation. Quasi-dense point cloud of the plant is obtained from multiview stereo reconstruction based on surface expansion. In order to overcome the negative effects from complex natural light changes and to obtain a more accurate plant point cloud, the Adaptive Normalized Cross-Correlation algorithm is used in calculating the matching cost between two images, which is robust to radiometric factors and can reduce the fattening effect around boundaries. In the stage of segmentation for each single leaf, an improved region growing method based on fully connected conditional random field (CRF) is proposed to separate the connected leaves with similar color. The method has three steps: boundary erosion, initial segmentation, and segmentation refinement. First, the edge of each leaf point cloud is eroded to remove the connectivity between leaves. Then leaves will be initially segmented by region growing algorithm based on local surface normal and curvature. Finally an efficient CRF inference method based on mean field approximation is applied to remove small isolated regions. Experimental results show that our multiview stereo reconstruction method is robust to illumination changes and can obtain accurate color point clouds. And the improved region growing method based on CRF can effectively separate the connected leaves in obtained point cloud.