A modified U-Net with a specific data argumentation method for semantic segmentation of weed images in the field

Abstract Weeds are harmful to crop yield. The segmentation of weeds in images is of great significance for precise weeding and reducing herbicide pollution. However, in the field environment, crops and weeds are similar, so it is difficult to accurately segment weed from complex field images. In this paper, an algorithm based on deep learning was proposed to segment weeds from images. This algorithm can segment weeds from the soil and crops in images. This semantic segmentation algorithm was developed with a simplified U-net. Due to the difficulty of image labeling for the semantic segmentation of weeds, an image augmentation method was proposed. The semantic segmentation network was trained by a two-stage training method composed of pre-training and fine-tuning. After training, the intersection over union (IoU) of this method was 92.91% and the average segmentation time of a single image (ST) was 51.71 ms. The results demonstrated that the modified U-Net was able to effectively segment weeds from images with a significant amount of other plants. The weed-targeted image segmentation method proposed in this paper can accurately segment weeds in complex field environments. It has a relatively wide range of applicability.