Image Segmentation Method with Joint Loss Optimization for Apple Picking

In the process of automated apple picking, when depth estimation is performed based on the center position of the target detection candidate frame, errors may occur due to the influence of complex environments such as leaf occlusion and lighting changes. Therefore, this work proposes an image segmentation method with joint loss optimization for apple picking. This method automatically labels the collected apple data set to obtain accurate apple label files. Then, the BCEDice is used as the joint loss optimization function to balance the problem of sparse training gradients, and the processed apple data set is input into the autoencoding model, combining ResNet and UNet for training. Finally, experiments on the apple segmentation dataset showed that the model performed well on evaluation indicators, with the Dice coefficient and precision reaching 89.7% and 89.6%, respectively. This validation result shows that the model also exhibits good generalization ability and segmentation accuracy on the diverse apple dataset collected.