基于改进YOLOv7的草莓成熟度检测方法

Improved YOLOv7-Based Method for Strawberry Ripeness Detection

【目的】为实现自然环境下草莓及其成熟度的高效准确检测,设计了一款改进的YOLOv7的草莓成熟度检测模型。【方法】模型采用YOLOv7作为基础网络对草莓成熟度进行检测。首先,使用PConv卷积替换原Head部分的3×3卷积以提高检测速度。其次,在部分特征层添加CBAM注意力机制提高模型关注重要信息的能力;再将原本上采样中简单的双线性插值算子替换为CARAFE,增加模型对草莓果实细节的感知能力。最后,利用迁移学习实现草莓数据集的训练和微调。【结果】改进前的模型mAP50为85.6%,改进后的模型mAP50为87.7%,参数量从原来的9.14×10^(6)下降为7.32×10^(6),GFLOPs下降为19.8,提高了检测速度。【结论】改进后的模型在草莓及其成熟度的检测中具有更高的检测精度和速度,可以在实际应用场景中实现草莓果实的生长监测。

【Objective】To achieve efficient and accurate detection of strawberry maturity in natural envi⁃ronments,an improved YOLOv7 model for strawberry ripeness detection is designed.【Method】In this study,YOLOv7 is used as the base network for strawberry ripeness detection,based on which,firstly,the 3×3 convolution of the original head part is replaced by the PConv convolution to improve the detec⁃tion speed.Secondly,the CBAM attention mechanism is added to some feature layers to improve the model′s ability to focus on important information.Furthermore,the simple bilinear interpolation operator in the original upsampling is replaced by CARAFE,which can increase the model′s ability to perceive straw⁃berry fruit details.Finally,migration learning is used to achieve training and fine-tuning of the strawberry dataset.【Result】The experimental results showed that the mAP50 of the model before improvement was 85.6%,while the mAP50 of the enhanced model reached 87.7%.Additionally,the number of parameters decreased from 9.14×10^(6)to 7.32×10^(6),and the GFLOPs decreased to 19.8,resulting in an enhanced de⁃tection speed.【Conclusion】The improved model has higher detection accuracy and speed in the detection of strawberries and their ripeness,which can realize the growth monitoring of strawberry fruits in practical application scenarios.