Weighted skip-connection feature fusion: A method for augmenting UAV oriented rice panicle image segmentation

Abstract

Unmanned aerial vehicles (UAVs) have the potential to reduce manual interventions in digitizing farmland and improve the accuracy and efficiency of data collection. Measuring crop yields with UAVs is a critical step in achieving precision agriculture on unmanned farms. The key to estimating rice yield is to distinguish the panicle region from the non-panicle region based on semantic segmentation. However, the traditional semantic segmentation model, such as the UNet, has an inferior segmentation performance on UAV images. In addition, UAVs are unable to segment the rice panicle region in real time due to the limited edge computing capabilities for some improved UNet models. To address this issue, this study proposes a new method for augmenting UAV rice panicle image segmentation called weighted skip-connection feature fusion (WSFF). Furthermore, a novel model WSUNet is constructed by combining WSFF and UNet model, which aims to enhance the performance of rice panicle segmentation without additional computational cost. Two datasets of rice panicle images taken with UAV are constructed. These two and cell nuclei dataset are used to compare the performance of UNet, WSUNet and UNet++. Mean intersection over union (mIOU) and mean pixel accuracy (mPA) are adopted as evaluation metrics. In terms of mIOU, the results indicate that WSUNet outperforms the UNet on all datasets, with a maximum increase of 2.84 on the rice panicle dataset. And the average inferencing speed (AIS) of WSUNet on CPU is 2.1 times that of UNet++. Additionally, in order to verify the role of WSFF, ¹WSUNet and ²WSUNet are constructed based on WSFF with two different skip-connection modes. By observing the training scalars of ¹WSUNet, ²WSUNet, WSUNet, and UNet, it can be seen that the model set with WSFF has a more competitive learning ability than UNet, and the segmentation performance of the model could be further improved with the increase of the amount of skip-connection.

Plant phenotyping relevance

UAV画像からイネ穂領域を抽出するセグメンテーション手法を開発し、複数データセットで性能比較・検証しているため、植物表現型取得法が中心である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.