Space Plant Image Segmentation via Multi-Scale Deep Feature Fusion

Abstract

As a key research in space science, plant experiments are persistently arranged and massive plant images are acquired, which are then almost manually processed for further analysis in tradition. This paper proposes a novel image segmentation method based on multi-scale deep feature fusion for automatic plant image segmentation. The method applies a deep convolutional neural network to extract multiple levels of features, and a skip architecture hierarchically fuses high-level to low-level features extracted by a deep layer, a middle layer and a shallow layer to recognize plant on the pixel level. The hierarchical feature fusion combines semantic information, middle-level information and geometrical features, which improves pixel-wise segmentation accuracy. Experiments demonstrate that our method performs well in segmentation accuracy, and can be applied to automatically extract useful information in space plant experiments.