PLANT LEAF SEGMENTATION USING A REFINED FULLY CONVOLUTIONAL NETWORK

Abstract

Traditional segmentation method has the problem that background noise disturbs to the accuracy of segmentation. To improve the segmentation accuracy, an improved fully convolutional network (FCN) for plant leaf segmentation is proposed in this paper. In order to realize the rapid and accurate segmentation of plant leaves, firstly, the original skip connection structure was improved by reducing the number of feature maps in convolution layer C6. Secondly, features of the shallower layers were further integrated. Thirdly, the direct-connected structure (Direct-FCNs) was adopted by removing some layers. Moreover, the feature maps reduced by the VGG16 model were up-sampled to the original image size by de-convolution. Therefore, several improved models were obtained. Finally, much images were divided into training set and test set. The proposed method is compared with other traditional methods. It can be observed from all the experimental results that the refined model has higher segmentation accuracy, less parameters and stronger robustness, which is able to be used for the automatic segmentation of plant leaves.