基于轻量化MLCE-RTMDet的人工去雄后玉米雄穗检测算法

Maize Tassel Detection Algorithm after Artificial Emasculation Based on Lightweight MLCE-RTMDet

摘要:

玉米制种田遗漏雄穗检测是实现人工去雄质量评估的关键。针对现有玉米雄穗检测模型面临的参数量大、检测效率低和精度差等问题，提出一种基于RTMDet-tiny的轻量级雄穗检测模型MLCE-RTMDet。模型采用轻量级的MobileNetv3作为主干特征提取网络，有效降低模型参数量；在特征提取网络中引入CBAM注意力模块，增强对雄穗目标的多尺度特征提取能力，克服引入轻量化网络可能带来的性能损失。同时，使用EIOU Loss替代GIOU Loss，进一步提高雄穗检测精度。在自建数据集上的试验表明，改进的MLCE-RTMDet模型参数量缩减至3.9×106，浮点运算数降至5.3×109，参数量和浮点运算数分别比原模型减少20.4%和34.6%。测试集上模型平均精度均值增至92.2%，较原模型提高1.2个百分点；同时，推理速度达到41.9f/s，增幅达12.6%。与YOLO v6、YOLO v8、YOLO X等当前主流模型相比，MLCE-RTMDet表现出更好的综合检测性能。改进后的高精度轻量化模型可为实现玉米制种田人工去雄后的遗漏雄穗检测提供技术支撑。

Abstract:

Detecting missed tassels is crucial for assessing the quality of aritificial emasculation in maize seed production fields. Aiming at the problems of large parameter quantity, low detection efficiency and poor accuracy of the existing maize tassel detection models, a lightweight tassel detection model based on RTMDet-tiny, named MLCE-RTMDet, was proposed. The model used the lightweight MobileNetv3 as the feature extraction network to effectively reduce the model parameters. The CBAM attention module in the neck network was integrated to enhance multi-scale feature extraction capability for tassel objects, overcoming potential performance losses caused by the lightweight networks. Simultaneously, the EIOU Loss was adopted, replacing the GIOU Loss, which further improved the accuracy of tassel detection. Experiments on the self-built dataset showed that the improved MLCE-RTMDet model reduced model parameters to 3.9×106, while the number of floating point operations was lowered to 5.3×109, resulting in a 20.4% reduction in parameters and a 34.6% decrease in computational complexity compared with that of the original model. When evaluated on the test set, the model’s mean average precision (mAP) reached 92.2%, reflecting a 1.2 percentage points improvement over the original model. The inference speed was increased to 41.9 frames per second (FPS), representing a 12.6% enhancement. Compared with current mainstream detection models such as YOLO v6, YOLO v8, and YOLO X, MLCE-RTMDet demonstrated superior overall detection performance. The improved high-accuracy lightweight model offered technical support for tassel re-inspection and emasculation quality assessment in maize seed production fields following artificial emasculation.