关键词: 果树, 树冠分割, Deeplabv3+, 语义分割, 条状池化, 注意力机制

Abstract: In order to solve the problem that it was difficult to accurately identify and segment individual fruit trees in hilly and mountainous orchards due to environmental factors such as terrain, light and boundary interference, an improved highprecision Deeplabv3+ semantic segmentation network model was proposed. Firstly, features were extracted from ResNet50 main trunk network, and pyramid splitting attention (PSA) mechanism was introduced to obtain clearer fruit tree contour boundary information. Then, the stripe pooling (SP) module was connected to the decoding part in series, and the feature extraction was enhanced by SP to obtain rich context information along the horizontal and vertical dimensions respectively, which expanded the range of sensitivity field and ensures the integrity and continuity of information. Through semantic segmentation, it could be concluded that in the tree crown image data set of fruit trees in hilly and mountainous areas with autonomous image annotation using Labelme tool, the identification and segmentation accuracy of individual fruit trees was 98.91%, and the average intersection ratio of fruit tree segmentation was 74.94%. Compared with PSPNet, UNet, FCN and Deeplabv3+, PA was increased by 2.5%, 1.88%, 1.03% and 1.85% respectively, while MIoU was increased by 10.93%, 8.19%, 2.78% and 5.73% respectively, there was obvious improvement data. The research results could provide data support for intelligent agricultural equipment in fine operations such as target spraying and growth identification in orchards.

Key words: fruit tree, crown segmentation, Deeplabv3+, semantic segmentation, strip pooling, attention mechanism