Abstract: Aiming at the problems that the traditional leaf recognition is susceptible to the environmental interference and is difficult to realize the multi-leaf real-time recognition in complex background, a real-time leaf recognition method based on CNN network and multi-task loss function is proposed. The CNN network is used to the extract image feature maps and input them into RPN network to generate regional proposals. On the basis of the feature maps and region proposals, the feature map is proposaled, the leaf classification and bounding box regression are performed respectively, and the leaf classification and position of the leaf prediction box are predicted. The multi-task loss function is used to constrain the classification and regression to improve the accuracy and speed of the leaf image classification and regression. Experimental results show that the average real-time leaf recognition accuracy is 91.8%, and the average real-time leaf recognition speed is 25 fps.

Key words: leaf recognition, feature map, CNN network, multitask loss function, region proposal