基于YOLO v8-GSGF模型的葡萄病害识别方法研究

Grape Disease Identification Method Based on YOLO v8-GSGF Model

摘要:

为进一步提高葡萄病害识别精度及速度，本文对YOLO v8模型进行了改进。首先，引入GhostNetV2主干特征提取网络，提高模型特征提取能力和识别性能。其次，嵌入SPPFCSPC金字塔池化，在保持感受野不变的情况下取得速度上的提升。再次，添加GAM-Attention注意力机制，减小信息缩减并放大特征信息，加快识别速度。最后，使用Focal-EIoU作为损失函数，使检测模型边界框回归性能得到提升，最终形成葡萄叶片病害识别模型YOLO v8-GSGF(YOLO v8+GhostNetV2+SPPFCSPC+GAM-Attention+Focal-EIoU)。经识别试验验证，YOLO v8-GSGF模型识别精度可达97.1%，推理时间为45.3ms，对各葡萄病害都能做到高精度识别。消融试验结果表明，各项改进均对模型识别性能有提升效果，其中，GhostNetV2主干网络对模型提升效果最为明显。YOLO v8-GSGF模型在消融试验中识别精度可达98.2%及推理时间为43.7ms，与原YOLO v8模型相比提升8.6个百分点及20.4ms，改进效果明显，可视化图更加直观地证明YOLO v8-GSGF模型可靠以及性能优越。与目前主流识别模型相比，YOLO v8-GSGF模型有更好的表现，识别精度和速度都更优，曲线图也直观地表明YOLO v8-GSGF模型性能优越，改进效果显著，能够满足葡萄果园病害识别的需求。

Abstract:

In order to further improve the accuracy and speed of grape disease identification, the YOLO v8 model was improved. Firstly, the GhostNetV2 backbone feature extraction network was introduced to improve the feature extraction ability and recognition performance of the model. Secondly, the SPPFCSPC pyramid pooling was embedded to improve the speed while keeping the receptive field unchanged. Thirdly, the GAM-Attention mechanism was added to reduce the information reduction and amplify the feature information to speed up the recognition. Finally, Focal-EIoU was used as the loss function to improve the bounding box regression performance of the detection model, and finally the grape leaf disease identification model YOLO v8-GSGF was formed. The recognition test verified that the YOLO v8-GSGF model can achieve 97.1% recognition accuracy and 45.3ms inference time, and can achieve high-precision identification of various grape diseases. The results of the ablation test showed that all the improvements had an effect on the recognition performance of the model, and the GhostNetV2 backbone network had the most obvious effect on the model. The YOLO v8-GSGF model can achieve 98.2% recognition accuracy and 43.7ms inference time in the ablation test, which was 8.6 percentage point and 20.4ms higher than that of the original YOLO v8 model. Compared with the current mainstream recognition model, the YOLO v8-GSGF model had better performance, better recognition accuracy and speed, and the curve chart also intuitively showed that the performance of the YOLO v8-GSGF model was superior, and the improvement effect was remarkable, which can meet the needs of grape orchard disease identification and had the potential for practical application.