Semantic segmentation model of cotton roots in-situ image based on attention mechanism

Abstract

The growth and distribution of root system in the soil has an important influence on the growth of plants and is one of the important factors affecting crop production. However, the root system of plants is located in the dark and closed soil. Even if we can obtain high-definition root image from the complex soils, the interference of the soil particles on root system and the small difference of color between them will pose challenges for further root segmentation. In this experiment, the cotton mature root system is used as the research object. Based on the introduction of sub-pixel convolution DeepLabv3+ semantic segmentation model, we further added the attention mechanism to the model, assigning more weight to the pixel points of fine roots and their root hairs, and designed a semantic segmentation model of cotton roots in-situ image based on the attention mechanism. The experimental results show that the model has higher segmentation accuracy and operational efficiency than only introduces sub-pixel convolution DeepLabv3+ model, U-Net model and SegNet model. The precision value, recall value and F1-score are 0.9971, 0.9984 and 0.9937 respectively, and the IoU value of 161 untrained root image segmentation tasks was 0.9875. At the same time, we also performed segmentation experiments on the early cotton root images. The results show that the DeepLabv3+ model which only introduces sub-pixel convolution, U-Net model and SegNet model have poor segmentation effects. The semantic segmentation model based on attention mechanism proposed in this paper can be segmented accurately. The above results show that the proposed model can distinguish the cotton root system from the complex soil background accurately and has good segmentation effect. It can realize the accurate segmentation of root image in early and mature period in the process of cotton root growth, and provide important theoretical value and practical application reference for deep learning in plant root segmentation.