Abstract: The accurate segmentation and recognition of plant target leaves is the premise and foundation for the plant recognition and growth state monitoring based on leaf images, but the complex background brings great challenges to the segmentation and recognition of leaves. In this study, an algorithm based on the Mask-RCNN deep learning network to segment and identify multi-target leaves in the complex background was proposed. A total of 7 357 images of common plant leaves in the nature were taken, and 3 000 images of four plant leaves were labeled as training databases. The 3 000 images contained four species of plants: Calathea makoyana, Viburnum odoratissinum, Hedera helix L. and Bauhinia tomentosa. 80 sample images of the four plant species were selected for the segmentation, identification and misclassification rate analysis. The results showed that the Mask-RCNN deep learning network had a good recognition effect for the four plants and there was no false recognition. The average segmentation error rate was 0.93%, with the maximum value not exceeding 2.49%, which meant the segmentation accuracy was 97.51%. At the same time, the algorithm had strong migration ability.

Key words: leaf recognition, Mask-RCNN, error rate, segmentation accuracy