Field Grape Berry Counting Algorithm Based on Spatial Attention Mechanism

Automated grape yield estimation is one of the important aspects of intelligent vineyard management. In order to meet requirements of fast estimation and large-scale grape yield estimation in complex field environments, this paper proposes a Spatial Attention Grape Berry Counting Network (SAGBCNet). Firstly, the Spatial Feature Network (SFNet) is used to extract more spatial information of grape berries in image. Then attention mechanism is used to adaptively fit the long-distance context features of location and channel, so as to effectively distinguish the grape berry from background. Finally, through Feature Fusion Network, feature maps of different scales are fused to generate the predicted density map, and counting results of grape berries in the field are obtained. The experiment on the field grape counting dataset showed that SAGBCNet obtained a lower counting error for grape berries. The mean absolute error (MAE) and mean squared error (MSE) were 35.39 and 47.39 respectively.

I. Introduction

In recent years, China’s grape planting area and output have increased. Traditional vineyard management has begun to change to a more efficient and intelligent way. Grape automatic yield estimation is one of the important links to realize intelligent management of vineyards. At the same time, it is of great significance to reduce production costs and improve agricultural efficiency. Grape berry count is a key step in yield estimation process. Currently, the counting of grape berries still relies mainly on traditional manual observation methods. This method has low efficiency, strong subjectivity and high labor cost. Some efforts have been made to automate the counting of grape berries [1], [2]. However, it has been studied in a single context, with a high degree of individual differentiability and a low number of grape bunches. It is just as inadequate for rapid estimation and large-scale grape yield estimation in complex field environments [1]. Therefore, field grape berry counting needs a gradual shift from tedious, limiting working methods to low-cost, efficient and robust approaches. The computer vision technology and deep learning get quickly development in the decades, object counting method based on computer vision is expected to become an effective means of field grape counting automation.