Research On Segmentation Of Plant Leaf Image

The function of plants is an indispensable part of human production activities.The value of plants is very important in agriculture,industry,medical field and ecological environment protection.In recent years,with the development of telecommunication technology,computer vision,image processing and other modern information technology,Plant Phenomics has become a hot research topic.In order to accurately obtain the information of various phenotypic parameters of plant leaves(as the number of blades,blade area and blade azimuth etc).In the real planting environment,the distribution of plant leaves is often dense and there are a large number of overlapping parts.Due to the influence of complex background,illumination and other external conditions,it makes the segmentation of plant leaves and the extraction of phenotypic parameters a challenge.The main research work is as follows:First,for segmented the leaves of cucumber plants planted under the glass greenhouse and plastic greenhouse,we uesd the Bilayer Convolutional Network and improved it.The improvement scheme is as follows: 1)Add a space to depth module at the network input and convert the input image into different depth modules;2)Introducing attention mechanism in order to make the network can pay more attention to the edge feature information of distinguishing each leaf;3)Replace Loss function in BCNet with Varifocal Loss.Our improved BCNet can accurately segment the leaves of cucumber plants under different growth stages and light conditions.The accuracy,recall and AP of target recognition were 97%,94% and 96.57%respectively and the accuracy,recall and AP of instance segmentation were 87%,83% and84.71% respectively.This is better than Faster R-CNN,Mask R-CNN,YOLOv4,PANet and other common target recognition and instance segmentation algorithmsSecond,we propose a set of algorithm of plant leaf detection and segmentation,to avoid some situations like labeling data will lead to excessive labor cost in the use of deep learning and make use of high requirements for the configuration of computer hardware in the process of training the model.This algorithm is suitable for oval plant leaf.We used four different kinds of plant leaves with elliptical shape and the following is the process of the algorithm: 1)Image preprocessing;2)Leaf center point extraction and statistics of plant leaves;3)Segmentation of plant leaves;4)Speculate the complete contour of plant leaves.Compared with the overlapping object segmentation algorithm based on deep learning,it saves labor cost and computer hardware cost on the premise of ensuring the segmentation effect and accuracy.