【摘要】 稻穗准确计数对估算水稻产量至关重要。然而在实际生产场景中，由于稻田环境复杂、水稻品种繁多、穗部形态多异等原因，现有检测方法存在精度不足、模型参数量大的问题。为此本研究提出一种轻量级稻穗检测模型CBLP-YOLO 11n。首先，在Backbone中，选择C3k2-CFCGLU替换原C3k2模块，增强模型对稻穗的特征提取和表达能力；其次，使用加权双向特征金字塔BiFPN实现多尺度特征信息融合，通过跳跃连接和删除冗余节点的方式，在提高稻穗识别精度的同时，有效降低模型浮点运算量；然后，设计轻量细节增强共享检测头(Lightweight detail-enhanced shared detection head, LDSDH),通过共享卷积降低检测头复杂度；最后采用Powerful-IoUv2损失函数替换原有的CIoU损失函数，加快模型收敛速度，并优化模型对稻穗的定位准确性。实验结果表明：CBLP-YOLO 11n模型的检测精确率为88.2%,召回率为87.9%,平均精度均值为93.9%。与YOLO 11n相比，CBLP-YOLO 11n的精确率提高1.9个百分点，召回率提高1.1个百分点，平均精度均值提高1.3个百分点，参数量下降23.7%,浮点运算量下降40.6%。相比于其它主流模型，CBLP-YOLO 11n模型的平均精度均值最高，内存占用量最小，仅为3.78 MB。所提模型实现了对稻穗的精确识别，可部署在无人机等资源受限设备中，为复杂田间背景下稻穗识别计数提供技术支撑。更多还原

【Abstract】 Rice is a globally important food crop and accurate counting of rice panicles is crucial for estimating rice production. However, in actual production scenarios, due to reasons such as complex paddy field environment, diversity of rice varieties, and their panicle morphological features, existing detection methods have insufficient accuracy and a large number of model parameters. To this end, a lightweight rice panicles detection model CBLP-YOLO 11n was proposed. Firstly, in Backbone, the original C3k2 module was replaced with C3k2-CFCGLU in order to enhance the model’s feature extraction and expression ability for rice panicles. Secondly, bidirectional feature pyramid network(BiFPN) was used to achieve multi-scale feature information fusion. By means of skip connections and deletion of redundant nodes, the recognition accuracy of rice panicles was improved while effectively reducing the computational complexity of the model. Then a lightweight detail-enhanced shared detection head(LDSDH) was designed and the complexity of detection head was reduced through shared convolution. Finally, the original CIoU loss function was replaced by Powerful-IoUv2 loss function to accelerate the convergence speed of the model and optimize positioning accuracy of the model for rice panicles. The experimental results showed that the detection precision of the CBLP-YOLO 11n model was 88.2%, the recall was 87.9% and the mean average precision was 93.9%. Compared with YOLO 11n, the CBLP-YOLO 11n model showed improvements in the precision, recall, mAP by 1.9, 1.1 and 1.3 percentage points, respectively. Meanwhile, the CBLP-YOLO 11n reduced the parameter quantity by 23.7% and the computational quantity by 40.6%. Compared with other mainstream models, the CBLP-YOLO 11n model had the highest average detection accuracy and the smallest memory usage, which was only 3.78 MB. The proposed model realized accurate identification of rice panicles and can be deployed in resource-constrained devices such as drones, providing technical support for identification and counting of rice panicles in complex field backgrounds.更多还原