YOLOv12n-RMB: An Improved Dual Target Recognition Network for Seedling Maize and Field Weeds in Cold Regions

Accurate identification of maize seedlings and weeds at the seedling stage in dry farmland of cold regions serves as an essential prerequisite for field variable weeding and intelligent plant protection operations. Existing YOLO networks exhibit prominent limitations in feature extraction for fine weeds that grow close to the ground and feature fusion across different scales, and show poor adaptability to complex field conditions including illumination fluctuations, tiny weed interference, and mixed soil textures. To address these issues, this study focuses on maize fields in the cold region at a high latitude in Heilongjiang and collects RGB field images of maize plants at the two to five leaf stage. Using YOLOv12n as the baseline model, the internal convolutional blocks within the C3k2 structure are replaced with RepViTBlock to construct the C3k2-R module. The Multi-Frequency in Multi-Scale Attention (MFMSA) module is embedded into A2C2f to obtain A2C2f-M. BiFPN is adopted to replace the original concatenation block to form Concat-B. By integrating the above three optimized modules, the improved model YOLOv12n-RMB is developed. The results demonstrate that YOLOv12n-RMB achieves an AP@0.5 of 0.9745 for maize and 0.7820 for weeds, with an overall mAP@0.5 improvement of 0.0421 relative to the original YOLOv12n. The proposed model can perform robustly under diverse field working conditions such as strong light, weak light, and varying seedling ages. This research provides visual technical support for field weed density evaluation and intelligent decision-making for variable weeding equipment.