Weed Recognition Method based on Hybrid CNN-Transformer Model

Abstract

As an important task in precision agriculture, weed recognition plays a crucial role in crop management and yield increase. However, achieving high accuracy and efficiency at the same time remains a challenge. To address the balance between accuracy and timeliness in weed recognition, this paper proposes a hybrid CNN-Transformer model for weed recognition. The model uses a combination of convolutional neural network (CNN) and Transformer structures for feature extraction and classification, taking into account both global and local information. In addition, the proposed Transformer Block incorporates the SDTA (Segmentation Depth Transpose Attention) mechanism to improve timeliness. Furthermore, this paper improves the original ViT model to enhance its accuracy. Experimental results on the Deep Weeds dataset by Olsen et al. show that the proposed hybrid model outperforms the original Vision Transformer model in weed recognition accuracy (89.43% vs. 96.08%). This research provides an effective solution for weed recognition using a hybrid model, with high practical value and application prospects.