Weeds pose a serious threat to agricultural production. Their intelligent identification and precise root localisation are essential for improving productivity and prowmoting sustainable development. This study focuses on achieving high-precision, rapid recognition and root localisation of multiple crops (e.g., maize) and weeds (e.g., quinoa, Poa pratensis, thistle, sedge) under complex field conditions Traditional methods are often inefficient and mainly address single-leaf or single-background scenarios, resulting in reduced accuracy in complex environments with multiple leaves and diverse backgrounds. To overcome these limitations, a deep learning-based crop–weed recognition system is proposed, combining YOLOv12 for object detection with DeepLab V3+ for root segmentation. To improve generalisation and robustness, systematic preprocessing was applied, including data augmentation (random rotation and cropping), Gaussian denoising, and conversion to grayscale and HSV colour spaces. The experimental results show that the TipDM Intelligent Technology public dataset delivers strong detection performance, achieving an overall precision of 84.0%, recall of 84.7%, and F1-score of 84.3%. The ablation experiment showed that we achieved good results on YOLOv12 and DeepLab V3 +. Compared with other model such as RoWeeder, TIA-YOLOv5 and VGG-16, our model outperforms these models. The improved YOLOv12 model achieved superior multi-category detection in complex field conditions. The results confirm the feasibility and effectiveness of deep learning models for weed detection and root localization, offering an efficient, accurate, and eco-friendly solution for intelligent weed management in modern agriculture.