Mask R-CNN模型在茄花花期识别中的应用研究

将Mask R-CNN实例分割模型应用于茄子花期识别研究，针对Mask R-CNN模型对大目标物存在误检和漏检的情况，提出使用混合空洞卷积融合普通卷积的方法，在ResNet50的残差块中进行参数修改，通过堆叠残差块完成对整个特征提取网络的改进，扩大了特征图感受野，增强了全局信息关联性；针对出现的过拟合问题，运用迁移学习方法将预训练的ResNet50特征提取网络作为茄花识别模型的初始参数，提高了模型在测试集泛化能力的同时提升了模型训练速度。运用改进的模型在测试集上的mAP为0.962，mIOU为0.715。通过定性分析并与其他模型进行对比，证明改进的模型能有效提高大目标物分割能力，对茄子花期识别具有良好效果。该研究为茄花自动授粉与花期管理提供了技术支持，对保证授粉质量，提升经济效益具有重要意义。

Application of Mask R-CNN Model in Identification of Eggplant Flowering Period

The Mask R-CNN example segmentation model is applied to the identification of eggplant flowering stage. In order to solve the problems of misdetection and omission of large targets, a method of using hybrid dilated convolution to fuse ordinary convolution. The parameters are modified in the residual blocks of ResNet50. The whole feature extraction network is improved by stacking the residual blocks, the feature map perception field is expanded, and the global information correlation is enhanced. For the over-fitting problem, the pre-training ResNet50 feature extraction network is used as the initial parameter of the eggplant flower recognition model by using the transfer learning method, which improves the generalization ability of the test set and the training speed of the model. The mAP and mIOU of the improved model are 0.962 and 0.715 on the test set. By comparison with other models, the improved model can effectively improve the large target segmentation ability and has a good effect on eggplant florescence recognition. This study provides technical support for automatic pollination and florescence management of eggplant, which is important for ensuring pollination quality and improving economic benefits.