In agricultural grafting cultivation,it is usually necessary to ensure that the scion leaves and rootstock leaves are cross-shaped af-ter grafting.In order to enable the automatic grafting machine to accurately segment the scion leaves in real time and find out the characteristic parameters of the cotyledons,a lightweight segmentation method based on improved UNet is proposed,Using the MobileNetV2 backbone as the feature extraction backbone,the Ghost Module is used to implement double convolution operations in the enhanced feature extraction lay-er,which improves network accuracy while reducing network parameters and calculations.The experimental results show that compared with the original model,the Mobile-UNet model has increased by 5.69％,1.32％,4.73％and 3.12％in indicators such as MIoU,Precision,Re-call and Dice coefficients,and the calculation amount and parameter amount of the model have decreased by 27.4％and 35.3％.In addition,compared with SegNet and DeepLabV3+classic segmentation models,this model has higher segmentation accuracy and fewer parameters.It is deployed in the automatic grafting machine system to realize the segmentation of scion cotyledons on the clamping mechanism.
万方数据