Abstract: As a key research in space science, space plant experiment usually obtains massive plant sequence images. The traditional processing methods are mostly observed manually for further analysis. This paper proposes a space plant image segmentation algorithm based on multi-scale deep feature fusion. This method uses a full-convolution deep neural network to extract multi-scale features, and hierarchically fuses features from deep to shallow to achieve pixel-level segmentation of plants. The hierarchical features fuse semantic information, middle layer information, and geometric features to improve segmentation accuracy. Experiments demonstrate that the method performs well in segmentation accuracy and can automatically extract useful information in space plant experiments.

Key words: image segmentation, full convolutional neural network, multi-scale feature fusion, plant