Authors

Abstract

Accurately identifying rice growth stages in cold regions is challenging due to subtle morphological differences between key stages (e.g., tillering, jointing-booting) and complex field conditions, often leading to inefficient water use. This paper proposes an improved YOLOv12n-based method for rice growth stage recognition. First, a Deformable Large Kernel Attention (D-LKA) mechanism is introduced to enhance the model's ability to capture multi-scale morphological features of rice plants through large receptive field convolution and adaptive sampling grids. Second, a weighted Bidirectional Feature Pyramid Network (BiFPN) and a Dynamic Upsampling module (DySample) are employed to construct an efficient multi-scale feature interaction pathway, improving the model's perception of details in key plant parts. Finally, the Normalized Wasserstein Distance (NWD) is adopted to optimize the small-target detection strategy, effectively mitigating the under-detection of small-scale features. A cold-region rice growth stage image dataset was constructed for training and evaluation. Results show that the proposed YOLOv12n-DBD model achieves a precision of 93.4%, a recall of 89.7%, a mean Average Precision (mAP@0.5) of 94.4%, and an inference speed of 121.9 FPS. The mAP@0.5 represents an improvement of 4.9 percentage points over the baseline model, outperforming current mainstream detection models while maintaining real-time performance. A mobile recognition system was also developed to provide a convenient solution. The proposed YOLOv12n-DBD model effectively balances recognition accuracy and computational efficiency in the complex environments of cold-region rice fields, offering reliable technical support for growth-stage-specific field management.

Abstract in Chinese

针对寒地水稻关键生育时期（如分蘖期、拔节孕穗期）形态特征存在细微差异、田间环境复杂导致识别难度大进而造成的水资源浪费问题。本文提出基于改进YOLOv12n模型的水稻生育时期识别方法。首先，引入可变形大核注意力机制（D-LKA），通过大感受野卷积与自适应采样网格增强模型对水稻植株多尺度形态特征的捕获能力；其次，采用加权双向特征金字塔网络（BiFPN）与动态上采样模块（DySample），构建高效的多尺度特征交互通路，提升模型对水稻关键部位的细节感知能力；最后，通过归一化Wasserstein距离（NWD）优化小目标检测策略，有效缓解小尺度特征的漏检问题。构建寒地水稻生育期图像数据集，对改进后的模型进行训练与验证。结果表明：YOLOv12n-DBD模型的准确率（P）、召回率（R）、平均精度均值（mAP）和推理速度分别为93.4%、89.7%、94.4%和121.9fps，其中mAP@0.5较基线模型提升4.9个百分点。在保持实时性的同时优于当前主流检测模型。构建移动端识别系统，可提供便捷识别方案。本文提出的YOLOv12n-DBD模型有效平衡了寒地水稻复杂环境下的识别精度与计算效率，能够为寒地水稻的生育期田间管理提供技术支持。