Crop weed image recognition of UAV based on improved HRNet-OCRNet

Abstract

Aiming at the problem of low model recognition accuracy caused by high similarity and mutual occlusion between crops and weeds in Unmanned Aerial Vehicle (UAV) images, a pixel-level weed recognition method based on improved HRNet-OCRNet is proposed. In this method, a multi-stage and multi-scale feature fusion method is added to HRNet to preserve more details and enhance semantic information at different levels, to solve the problem of high similarity between crops and weeds. The spatial self-attention module of Polarized Self-Attention (PSA) is integrated to HRNet, enhance the network's learning of important features, and reduce the false identification caused by mutual occlusion of crops and weeds. The expansion prediction method is used to generate an accurate distribution map of crop weeds. Compared with Deeplabv3+, GCNet and K-Net, the experimental results show that the proposed method has higher recognition accuracy for crop weeds, and mean intersection over union (mIoU) reaches 85.76%.