Abstract: In order to realize the pruning recognition of dormant fruit trees, a network model based on semantic segmentation was studied to identify pruned branches and determine the coordinates of pruning points. A binocular camera was used to build a visual system to obtain the data set of fruit trees. VGG16 and RestNet-50, which were respectively integrated with pre-training weights and CBAM （attention mechanism）, were used as two deep learning models of U-Net backbone feature extraction network to segment pruned branches. At the same time, their effects were obtained and compared. Based on the obtained segmented image, two methods, skeleton extraction and pruning point clustering, were used to determine the coordinates of pruning points. The results showed that the U-Net model based on VGG16 feature extraction network had better recognition results. The mean intersection over union （MIOU）, mean pixel accuracy （MPA） and F scores during the training of the model were 84.80%, 91.83% and 92.679% respectively. By segmenting the model image of artificial simulated fruit trees and using the pruning point clustering method, the two-dimensional coordinates of pruning points could be determined quickly and in real time, which laid the foundation for pruning operations.

Key words: branches identification, coordinates of pruning points, extraction, pre-training weight, CBAM（attention mechanism）