基于轻量化高效层聚合网络的黄花成熟度检测方法

针对黄花传统人工识别效率低，辨识标准不统一的问题，提出基于轻量化和高效层聚合过渡网络的黄花成熟度识别方法LSEB YOLO v7。首先，引入轻量化卷积对高效层聚合网络和过渡模块进行轻量化处理，减少模型计算量。其次，在特征提取与特征融合网络之间增加通道注意力机制，提升模型检测性能。最后，在特征融合网络中，优化通道信息融合方式，使用双向特征金字塔网络替换Concatenate,增加信息融合通道，持续提升模型性能。实验结果表明：与原始模型相比，在黄花成熟度检测中，改进后的LSEB YOLO v7模型参数量和浮点运算量分别减少约2.0×10⁶和7.7×10⁹。训练时长由8.025 h降低至7.746 h,模型体积压缩约4 MB。同时，训练精确率和召回率分别提升约0.64个百分点和0.14个百分点，mAP@0.5和mAP@0.5:0.95分别提升约1.84个百分点和1.02个百分点。此外，调和均值性能保持不变，均为84.00%。LSEB YOLO v7算法可均衡模型复杂性与性能，为黄花成熟度检测和智能化采摘设备提供技术支持。

Maturity Detection Method for Hemerocallis citrina baroni Based on Lightweight and Efficient Layer Aggregation Network

To address the problems of low efficiency of traditional manual identification and inconsistent identification standards, a ripening identification method for Hemerocallis citrina baroni based on lightweight and efficient layer aggregation network LSEB YOLO v7 was proposed. Firstly, lightweight convolution was introduced to lighten the efficient layer aggregation network and transition module to reduce the model computation. Secondly, the channel attention mechanism was added between the feature extraction and feature fusion networks to improve the model detection performance. Finally, in the feature fusion network, the channel information fusion method was optimized, and the bi-directional feature pyramid network was used to replace concatenate to increase the information fusion channels and continuously improve the model performance. The experimental results showed that compared with the original algorithm, in the Hemerocallis citrina baroni maturity detection, the number of parameters and floating-point operations of the improved LSEB YOLO v7 algorithm were reduced by about 2.0×106 and 7.7×109, respectively, and the training time was reduced from 8.025h to 7.746h, and the model volume was compressed by about 4MB. Meanwhile, the training precision and recall were improved by about 0.64 percentage and 0.14 percentage, respectively. The mAP@0.5 and mAP@0.5：0.95 were improved by about 1.84 percentages and 1.02 percentages, respectively. In addition, the harmonized mean remained unchanged at 84.00%. It was evident that the proposed LSEB YOLO v7 algorithm solved the problem of the paradox between model complexity and performance, and provided technical support for intelligent ripening and harvesting inspection equipment for Hemerocallis citrina baroni.