An Improved Transformer Network With Multi-Scale Convolution for Weed Identification in Sugarcane Field

Automated weeding equipment is urgently needed to deal with weeds in farmlands in the context of the rapid development of Intelligent Agriculture. A system for accurately identifying crops and weeds in images is crucial component of automated weeding equipment. However, in the field environment, crops and weeds grow intertwined, and weeds are very similar to sugarcane leaves, making it difficult to accurately segment crops, weeds, and their boundaries from images. In this paper, we proposed a novel network that fully utilizes low-level semantic information to accurately segment crops and weeds in images, improving the accuracy of crop and weed segmentation while reducing the need for training weight parameters and improving speed in the prediction stage. Specifically, we made three important modifications for crop and weed identification. First, a Multi-scale Feature Extraction and Fusion module (MFEF) was designed to capture abundant low-level semantic feature information. Afterward, we introduce a Global Response Normalization (GRN) block to select more useful feature information. Finally, a series of residual attention transformer layers are designed to transmit the long-range dependency information extracted between layers. Numerous experimental results confirmed that our proposed network achieved excellent performance in segmenting sugarcane and weed images. Specifically, (1) the mean accuracy and, Mean Intersection of Union (MIoU) reached 96.97% and 94.13%, respectively, (2) the training parameters of the model have been reduced by more than 25%, improving the Frames Per Second (FPS) value of the prediction process, and (3) is also effective on the publicly available BoniRob Dataset, indicating that the proposed model has considerable generalization ability. This study provides an accurate weed identification map and has reference significance for subsequent systems as mechanical weeding equipment.