Research On Faster-rCNN Based Root Scanning Image Location And Recognition

The root system plays a vital role in the growth and development of trees.The underground root system is intricate and intertwined,so the identification and positioning of the tree root system is a problem that needs to be solved urgently.The traditional root scanning image positioning and recognition methods are mostly based on digital image analysis methods,and there are various problems that limit their wide application.Faster Region CNN(Faster-rCNN)as a target recognition model in deep learning,not only can achieve the effect of classification,but also can quickly locate the region of interest through the Region Proposal Network(RPN),and automate the recognition and positioning of the root scan image.This paper is devoted to research on the method of localization and identification of root scanning images based on Faster-rCNN.The main work includes the following aspects:1.Construct a root recognition data set.The data set is composed of two parts,which are simulated images using GPRMAX V2.0 forward modeling and field images collected by the rooting method.2.Design Faster-rCNN network structure.This paper uses simulated image data sets to train the network,and the results show that the classification accuracy rate reaches 86%.It shows that the model has application value for the identification and positioning of hyperbola.Then the field image data set is used to train the network,and the accuracy of the model is 72%,which further illustrates the feasibility of the model for hyperbola identification and positioning.3.Explore the application of principal component analysis(PCA)in root recognition.First,PCA is applied to the field of gesture recognition to explore the best data dimension of PCA.The results show that for samples of the same size,there is an optimal dimensional range that makes the classification accuracy the best.After that,try to use PCA to perform feature extraction on the root scan image,and send the extracted image samples to the model training.The results show that when the sample compression rate is 4%,the recognition accuracy of the Faster-rCNN model can be improved by nearly 10%.4.Carry out the on-site experiment of the Summer Palace to test the recognition effect of the algorithm.The model in the article is used to identify and locate the hyperbola of the root scan images obtained in the Summer Palace,and compare the model recognition results with the manual recognition results,and horizontally compare with the Tree WIN software processing results.The results show that the model can basically meet the needs of identification and localization of root scanning images in the field.