Research On Leaf Recognition Technology Combined With Image Segmentation

With the rapid development of computer technology,real-time classification and identification of plants through leaves is a hot and difficult issue in current research.Using computer technology to classify and identify plants in real time,it can be used in digital plant specimens,forestry informatization and other fields.However,current leaf recognition methods are susceptible to environmental factors,and the recognition rate and calculation efficiency are low.This thesis mainly studies the leaf recognition method combined with image segmentation,and designs and implements a real-time leaf detection and recognition system.The main research results of this article are as follows:(1)Aiming at the problems of over-segmentation and inaccurate segmentation boundary in current image segmentation,proposing and realizing a leaf image segmentation method based on target contour.Firstly,performing grayscale and noise reduction on the leaf image.Secondly,constructing a fully convolutional network structure,using VGG-16 as the initial training model of FCN.Thirdly,using FCN to segment the leaf contour,restoring the original image size by deconvolution upsampling,using jump structure fusion feature information.Finally,using porous convolution to improve contour resolution,and using CRF to optimize edges,outputting leaf target segmentation map.(2)Aiming at the problem that it is difficult to automatically extract leaf features in current supervised feature extraction,researching and designing a leaf feature extraction method based on CNN.Firstly,labeling and mapping leaf categories.Secondly,establishing Inception-V2 network model structure,using asymmetric convolution to reduce computing costs,using Batch Normalization assistant to improve computing efficiency.Thirdly,using Inception-V2 network model to extract leaf image features.Finally,using reverse learning mechanism adjusts the parametersand outputting the leaf image feature map.(3)Aiming at the problems that the current leaf identification is susceptible to environmental interference,it is difficult to realize real-time leaf identification in a complex background,and multi-leaf identification,a real-time leaf identification method based on regions is proposed and realized.Firstly,inputting the feature map to the RPN network to generate the regional proposal.Secondly,using the ROI pooling method to integrate the feature map and the proposal information to extract the feature map proposal.Thirdly,sending the feature map proposal to the fully connected layer for leaf classification and boundary box regression.Finally,using multi-task loss function to constrain classification and regression,improving accuracy and operation speed,and accurately predicting the leaf classification and positioning of the leaf prediction box.(4)Designing and implementing a leaf recognition system combined with image segmentation.The system is mainly divided into four modules: optional module,leaf feature extraction module,leaf identification module,and human-machine interaction module.The results of leaf classification and recognition are obtained.The system has been tested and operated in the laboratory,and the accuracy of real-time leaf identification is high,and the system is robust in operation.