RT-Net: Plant Phenotype Semantic Segmentation Network Based on Advanced Deep Learning Framework

Abstract

Quantitatively deriving plant phenotypes from plant images in a non-contact manner is a very challenging task that relies heavily on the accurate segmentation of plant images. Previous methods mainly used the U-Net network structure and attention mechanism to obtain the corresponding plant phenotype segmentation results. However, the U-Net structure and attention mechanism are relatively outdated, and its method can only achieve a Dice score of 98.47% on the open source dataset, which is still insufficient for the recent plant phenotype segmentation task and needs to be further improved for more detailed research. Therefore, in view of the low segmentation performance of existing plant phenotype semantic segmentation models, this paper proposes a semantic segmentation network RT-Net based on an advanced deep learning framework. Specifically, the network mainly adopts the encoder-decoder network structure of deeplabv3+, and the encoding part of the network adopts the more efficient RepVGG as the backbone network for local feature extraction. At the same time, compared with the traditional Atrous Spatial Pyramid Pooling (ASPP), this paper designs the (Atrous Spatial Pyramid Pooling Based Transformer)ASPPBT module to extract more global feature information through a global adaptive method to obtain denser plant phenotypes. The decoding part performs feature fusion on the output of the encoding part, and then uses upsampling to restore the scale, and finally obtains the semantic segmentation result. The experimental results show that the proposed network has achieved a Dice score of 99.33% on the Arabidopsis plant dataset released by the CVPPP14 competition, and has better segmentation ability compared with other advanced plant field segmentation algorithms