Method for Segmentation of Litchi Branches Based on the Improved DeepLabv3+

It is necessary to develop automatic picking technology improve the efficiency of litchi picking, and accurate segmentation branches key that allows robots complete task. To solve problem inaccurate under natural conditions, this paper proposes a method for based on improved DeepLabv3+, which replaced backbone network DeepLabv3+ used Dilated Residual Networks as enhance model’s feature extraction capability. During training process, combination Cross-Entropy loss dice coefficient was function cause model pay more attention branch area, could alleviate negative impact imbalance between background. In addition, Coordinate Attention module added atrous spatial pyramid pooling, channel location information multi-scale semantic features acquired by are simultaneously considered. The experimental results show mean intersection over union pixel accuracy 90.28% 94.95%, respectively, frames per second (FPS) 19.83. Compared with classical network, 13.57% 15.78%, respectively. This can accurately segment branches, provides powerful technical support help litchi-picking find branches.