Root hair image processing based on deep learning and prior knowledge

Abstract

In situ measurement of root traits is very essential for better understanding of root development. However, difficulty in acquisition of root traits, fine roots and root hairs in particular, from low contrast background presents challenges for root traits segmentation. In this paper, an in-ground device termed microrhizotron was used for local root area observation and capture. An efficient and accurate model was proposed for segmenting local detailed root (fine roots and root hairs) image by constructing region of interest and adding prior knowledge in convolutional neural network. In order to reduce complexity of root image, regional growth result was used to position the root area and thus construct the region of interest. Transfer learning was applied to pre-train models on relevant dataset as initial parameters. Root axes and root hairs were separated using a pruning method, and root hair parameters were extracted based on it. The result showed that pre-trained features on relevant dataset could somewhat overcome small dataset, and prior knowledge reduced mislearning of background. The P-T-U-Net model (U-Net base on prior knowledge and transfer learning) had the best performance in root segmentation among all the applied models. The Intersection over Union (IoU), Pixel accuracy (PA), and F1 was 0.869, 0.977, and 0.872 respectively for pepper root hairs. It succeed in root traits segmentation with an average F1 score over 0.9. Plant species has little influence on the segmentation. It was able to figure out details on crossing and overlapping roots, and was able to focus on local root areas over time.