ATT-UNet: Pixel-wise Staircase Attention for Weed and Crop Detection

Abstract

Deep learning has been widely applied to image segmentation, effectively locating crops and weeds, reducing herbicide use, and avoiding unnecessary production costs and pollution. Weed recognition using traditional deep learning methods has been around for some time, but there are challenges in extracting, detecting, and segmenting weeds. A real-time segmentation method based on the UNet++ network is proposed in this paper. The Attention module is integrated into the UNet++ upsampling process to effectively suppress external noise interference by adopting UNet++ as the backbone network for extracting multi-scale information fusion. The UNet++ model with the attention mechanism module achieves better results with a higher IOU than the UNet++ model that is commonly used in medical image analysis. This method is effective in detecting crop-weeds in complex backgrounds and can serve as a reference for precise robot weeding.