High-Quality Wheat Stalk Cross Section Micrograph Measurement via Semantic Segmentation Network

Abstract

Sclerenchyma and vascular bundle are two major compositions of the wheat stalk, and their measurement is of great importance for the study on wheat growth. This article investigates a novel convolution neural network (CNN)-based method for high-quality measurement of sclerenchyma and vascular bundle in the wheat stalk cross section micrograph (i.e., WSCSM). Compared with other types of images, WSCSMs include more small-scale objects and suffer from color feature degeneration. To tackle these issues, the proposed network consists of two branches. The first introduces attention mechanism to enhance discriminative information extraction ability of hierarchical feature maps, which improves the output resolution of objects without damaging their semantic information. The second branch uses the distribution feature of every tissue to provide prior information and thus refines the segmentation results. Meanwhile, to successfully train and evaluate a deep model, we build and release a dataset of WSCSM with annotations for benchmark evaluation. In addition, to address the image shortage and class imbalance during training, we also develop a new image cropping method to greatly augment the valid training data. Extensive experiments on this dataset demonstrate that our method can outperform the current state-of-the-arts significantly, more than 5.7% mIoU improvement.