Three-dimensional image recognition of soybean canopy based on improved multi-view network

Rapid and effective identification and diagnosis of soybean drought conditions is crucial for soybean yield and quality. Due to the complexity and diversity of agricultural environments, deep learning models based on threedimensional data suffer from low accuracy and slow efficiency in practical applications, this paper proposes a three-dimensional image recognition method for soybean canopy based on an improved multi-view network. A lightweight network Res2net was used to reconstruct the feature extraction skeleton network in the MVCNN model, and the group convolution module of the network was optimized by embedding the ECA attention mechanism to propose a new three-dimensional image recognition model based on multi-view network (ECAMVRes2net). In the study, drought soybeans were used as an example to obtain projected images of soybean canopy in six viewpoints using three-dimensional rotation and image feature theory, and the proposed ECAMVRes2net was applied to carry out three-dimensional image recognition experiments of drought soybeans, and its recognition accuracy, F1 value and Kappa coefficient reached 96.665 %, 96.7 % and 0.924, respectively, compared with MVCNN, MVResnet, Pointnet++ and PointConv models with 3 evaluation metrics average improved by 17.289 %, 17.43 % and 0.356, respectively. The result realized a lightweight fast and accurate network model suitable for three-dimensional image recognition, which provides a theoretical foundation and technical support for the rapid recognition and accurate management of crops based on three-dimensional image processing.