A wheat spike image segmentation method based on improved U-Net

Abstract

The segmentation of wheat spike images is a prerequisite for conducting research on wheat spike diseases and yield estimation. To address issues such as small color differences in the background of wheat spike images in the field and low segmentation accuracy, this article proposes a wheat spike segmentation method called SAU-Net (Striped Pooling and Attention Mechanism optimized U-Net). Firstly, based on the U-Net model, residual network 50 (ResNet50) is selected as the backbone network of U-Net to reduce feature loss. Then, the stripe pooling block (SPB) and multi-scale dilated convolution (MSDC) are used to obtain local and global features, enhancing the accuracy of wheat spike feature extraction. Meanwhile, the convolutional block attention module (CBAM) is adopted to capture the dependency between channels and space, strengthen the focus on important features, and reduce the influence of background on segmentation results. Finally, a joint loss function is employed to further optimize the network performance. The results show that the average Intersection over Union (IoU) of the improved SAU-Net model is 88.57%, which is 5.29 percentage points higher than the improved U-Net model before. Compared with Pyramid Scene Parsing Network (PSPNet), Deep Convolutional Lab v3 (DeepLabv3), fully convolutional network (FCN), and Lite Residual Atrous Spatial Pyramid Pooling (Lraspp) network models, the SAU-Net model has the best segmentation accuracy. This study achieves wheat spike segmentation under complex backgrounds, providing technical support for wheat spike disease diagnosis and crop phenotype analysis.

Plant phenotyping relevance

コムギ穂画像から穂を抽出する画像セグメンテーション手法を開発・比較評価しており、植物器官の表現型取得が中心である。

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.