MCC-Net: A class attention-enhanced multi-scale model for internal structure segmentation of rice seedling stem

Abstract

Internal structural parameters of rice seedling stems are of great significance for rice growth detection, rice selection, breeding, and damage examination. Aiming at the problems of non-repeatability and low detection accuracy in the existing plant internal structure phenotypic traits detection methods, this paper presents a non-destructive segmentation method for examining the internal structure of rice seedling stems based on deep learning. We use a standard X-ray CT imaging technology to obtain non-destructive tomographic images of rice seedling stems and then design a class attention-enhanced multi-scale segmentation model (MCC-Net), where UNet is used as the backbone network. Specifically, the proposed MCC-Net mainly consists of three core components: multi-scale convolutional block (MCB), coordinate spatial attention (CSA) module, and class attention enhancement (CAE) module. MCB is the main component of the encoder to improve the feature extraction ability of the model for regions of different sizes in the internal structure. CSA is embedded into the UNet skip connections to enhance the expression of effective features and automatically locate the regions with different structures of rice seedling stems. CAE is designed to calculate the dependencies between image pixels and categories, which can enhance the feature expression from the perspective of categories and correct the category errors in the segmentation results. The experimental results show that MIOU, average dice coefficient and average precision of our proposed MCC-Net model on the self-built rice seedling stem CT image dataset are 92.56%, 96.33% and 96.59% respectively. Compared with several state of the art models, the proposed model achieves better segmentation performance on the rice seedling stem CT image dataset.