Weed Semantic Segmentation Network Combining Hybrid Attention Mechanism

Weed removal plays a crucial role in agricultural production. As the foundation for future intelligent weeding, precise weeding relies heavily on image recognition and segmentation. This paper proposes an enhanced hybrid attention mechanism, cbam, to be integrated into a semantic segmentation model using Unet as the model framework and resnet50 as the backbone extraction network. The aim is to address the challenges of low recognition rate and poor segmentation effect of maize seedling and weed images in complex environments. The results demonstrate that this attention mechanism significantly enhances the performance of the network model. Through adjusting the embedding position, the average intersection ratio of image segmentation reaches 90.01%, with an average pixel accuracy of 98.25%. These findings indicate that the segmentation model excels in actual segmentation tasks, meeting the requirements for seedling and grass recognition segmentation tasks effectively.

I. Introduction

Weeds have a significant impact on the yield and quality of crops, making weed removal an essential aspect of agricultural production. The advancement of machine vision technology has demonstrated great potential in replacing human eye recognition for weed detection. Through image processing and analysis techniques, machine vision can quickly and accurately identify field weeds for targeted weed control. Louargant et al. successfully achieved early weed detection in row crops by integrating spatial and spectral information. Wang et al. utilized various features such as biological ecology, spectral characteristics, visual texture, and spatial background to develop different classification methods based on color index, thresholding, and learning for effective weed detection. Furthermore, modern sensors with excellent temporal, spatial, and spectral resolution are widely employed in weed identification. Khan et al. relied on hyperspectral imaging technology using spectral cameras to capture field images; Lammie et al. implemented weed image acquisition and AI acceleration based on FPGA (Field Programmable Gate Arrays)^[1].