基于改进YOLOv5的轻量级黄花成熟检测方法

Lightweight method for maturity detection of Hemerocallis citrina Baroni based on improved YOLOv5

黄花菜具有较短的采摘周期和相对严格的采摘要求,针对传统人工采摘效率低、主观性高的问题,提出一种基于深度学习的SSH-YOLOv5黄花成熟度检测算法。以YOLOv5模型为基础,结合轻量级网络ShuffleNet V2基本残差单元压缩网络模型大小,提升模型目标检测速度;引入SE Net通道注意力机制模块,增强模型对有用特征信息的敏感度,提高目标检测精度;将普通卷积替换为深度可分离卷积模块,进一步减少模型计算量。试验结果表明,改进后的SSH-YOLOv5模型参数量和浮点运算量分别减少61.6%和68.3%,网络层数减少18%,同时SSH-YOLOv5的检测精度由原算法的88.8%提高到91.2%,实时检测速度达到66.4 f/s,相比原算法提高18.1%,达到实时检测要求。改进后的算法不仅实现模型的轻量化,同时也使黄花成熟度检测更加准确和快速,可以较好地满足黄花检测需求。

Hemerocallis citrina Baroni has a short picking cycle and relatively strict picking requirements.Aiming at the problems of the low efficiency and high subjectivity of manual harvesting of Hemerocallis citrina Baroni,a deep learning-based SSH-YOLOv5 Hemerocallis citrina Baroni maturity detection algorithm was proposed.Based on the YOLOv5 model,combined with the lightweight network ShuffleNet V2 basic residual unit to compress the size of the network model,and improve the model target detection speed.The attention mechanism module of Squeeze-and-Excitation network was integrated into the model to enhance the sensitivity of the model to useful feature information,and improve target detection precision,and ordinary convolution was replaced with depth-separable convolution module to further reduce the model computation.The experimental results showed that the number of parameters and floating point operations of the improved SSH-YOLOv5 model were reduced by 61.6%and 68.3%respectively,and the number of network layers was reduced by 18%,while the detection precision of SSH-YOLOv5 was improved from 88.8%to 91.2%of the original algorithm.The real-time detection speed reached 66.4 f/s,which was 18.1%higher than the original YOLOv5 algorithm and met the real-time detection requirements.The improved algorithm not only makes the model lightweight,but also makes Hemerocallis citrina Baroni maturity detection more accurate and faster,which can better meet the demand of Hemerocallis citrina Baroni detection.