Robust counting for multi-species plants based on Few-Shot learning

Abstract

Listen

Translate article

Accurate counting of plant organs at various growth stages is crucial for crop growth monitoring, phenotypic assessment, and yield prediction. Traditional plant counting models, typically designed for specific plant types, often exhibit poor generalization capabilities. In this research, an improved model, Field-CounTR, was proposed to accurately count multiple plant organs. The Field-Count dataset, utilizing a Few-Shot method, was introduced to enhance accuracy in plant counting. Additionally, a mixed salient module was developed to improve the feature fusion capability of sample images by exploiting their high similarity in Few-Shot counting tasks. During the downsampling process, Shuffle Attention was integrated to reduce redundant information in the feature fusion process. Furthermore, a hybrid convolution module combining Do-Conv and dilated convolutions was developed to increase the speed of convolutional inference and expand the receptive field through over-parameterized and dilation operations. To assess the effectiveness of the proposed approach, tests were conducted using the Field-Count dataset, which includes Sorghum, Wheat, Maize, and Rice. The results demonstrated a mean absolute error (MAE) of 14.49 and a root mean square error (RMSE) of 21.14. Compared with the CounTR model, the Field-CounTR model reduced the MAE and RMSE by 2.01 and 1.33, respectively. The enhanced Field-CounTR model exhibited superior feature extraction performance, high detection accuracy, and excellent generalization capabilities. This model can accurately count multiple plant types in complex field or orchard conditions and offers a wide range of applications.