Research Area:  Machine Learning
The identification of plant species mainly depends on the recognition of plant leaf characteristics. However, most recognition systems show the weak performance on detecting small objects like plant leaves in the complicated background. In order to improve the recognition ability of plant leaves in the complex environment, this paper proposes an improved deep convolutional neural network, which takes advantage of the Inception V2 with batch normalization (BN) instead of convolutional neural layers in the faster region convolutional neural network (Faster RCNN) offering multiscale image features to the region proposal network (RPN). In addition, the original images first are cut into the specified size according to the numerical order, and the segmented images are loaded into the proposed network sequentially. After the precise classification through softmax and bounding box regressor, the segmented images with identification labels are spliced together as final output images. The experimental results show that the proposed approach has higher recognition accuracy than Faster RCNN in recognizing leaf species in the complex background.
Keywords:  
Deep learning
Convolutional neural network
Leaf recognition
Complex background
batch normalization (BN)
Author(s) Name:  Xiaolong Zhu, Meng Zhu, Honge Ren
Journal name:  Cognitive Systems Research
Conferrence name:  
Publisher name:  Elsevier
DOI:  10.1016/j.cogsys.2018.06.008
Volume Information:  Volume 52, December 2018, Pages 223-233
Paper Link:   https://www.sciencedirect.com/science/article/abs/pii/S1389041717303480