基于改进 YOLOv8n和探地雷达图像的冬笋快速识别研究OA

Winter bamboo shoots usually grow at a depth of 20 cm underground,and it is generally difficult to determine their location through visual methods by bamboo farmers.Ground penetrating radar(GPR)technology is used to detect winter bamboo shoots.However,the features of winter bamboo shoots in GPR echo grayscale images are complex and variable,posing challenges to the efficiency and accuracy of on-site interpretation by staff.Therefore,proposed an im-proved YOLOv8n-based winter bamboo shoot echo image recognition method,ODE-YOLOv8n.In the ODE-YOLOv8n model,ODConv was used to construct the C2f-ODConv module,replacing all the original C2f modules.The use of four-dimensional convolution strategies allowed the model to better adapt to the irregular echo features of winter bamboo shoots,enhancing its feature extraction capabilities.The DAT mechanism was inserted at the end of the backbone net-work to improve the flexibility and efficiency of the self-attention module,capturing more winter bamboo shoot informa-tion features.The Efficient-Detect head shared Conv layer parameters and using SCConv to improve detection accuracy.A dataset of 1,346 ground-penetrating radar grayscale images was constructed for winter bamboo shoots,and ablation and comparative experiments were conducted on this dataset.The results showed that the ODE-YOLOv8n network model achieved an Precision of 94.6%,Recall of 84.1%,mAP50 of 92.0%,and an mAP50-90 of 56.1%.Additionally,compared to SSD,Faster R-CNN,YOLOv3,YOLOv5s,YOLOv5m,YOLOv7-tiny,YOLOv7 and YOLOv8n,the mAP50 increased by 6.6,9.1,4.8,9.1,12.3,7.0,6.2 and 4.2 percentage points,respectively.Finally,the ODE-YOLOv8n model was deployed on an NUC host using the OpenVINO inference framework,achieving a single-im-age inference time of 80 ms,which met the speed requirements for winter bamboo shoot detection.