Design And Implementation Of LiDAR Data Classification Algorithm Based On Residual Network

Light Detection and Ranging（Li DAR）,technology is an active remote sensing technology.Li DAR data can provide valuable information such as the height of the features in the research area,so as to facilitate computer processing to distinguish the types of features.Traditional Convolutional Neural Networks（CNN）have been widely used in remote sensing image classification in recent years.However,as the depth of the network increases,traditional CNN will have over-fitting problems.The Residual Network（Res Net）with jump connections in the deep neural network can alleviate the above problems.However,in the process of deepening the network,Res Net also has problems such as the gradual reduction in the size of the feature map and the loss of detailed information.Therefore,this paper uses Res Net as the core basic network and improves and optimizes it to achieve a balance between the classification accuracy of the model and the time complexity.The main contents of this paper include:First,the comparison of Li DAR data classification research based on traditional methods and deep learning methods at home and abroad in recent years provides a basis for subsequent model improvement.At the same time,in order to comprehensively evaluate the model proposed in this paper,a representative evaluation index in the field of remote sensing image classification is selected.Secondly,the Li DAR data classification algorithm based on the Dilated Residual Capsule Network（Dilated-Res Caps Net）is designed and implemented.The model is based on the Res Net-34 network and introduces the dilated convolution module into the Res Net,which expands the receptive field of the model without increasing the amount of parameters.The design of the dilation rate of the dilated convolution adopts the mixed expansion rate of odd and even,which avoids the grid effect of the dilated convolution,and uses the capsule network to further extract more detailed spatial feature information.This model not only overcomes the insensitivity of traditional CNN to spatial information,but also improves the ability to extract fine features of the target.Compared with some classic algorithms,Dilated-Res Caps Net has better classification results.Finally,a Transformer Li DAR data classification algorithm based on residual network is designed and implemented（Res Transformer）.Aiming at the small sample problem of Transformer capturing global feature representation and Li DAR data,the model combines residual network with Transformer to fuse local feature representation and global feature representation at different resolutions in an interactive way.At the same time,the label smoothing regularization method is introduced in this model to solve the problem of over-fitting in the model training process.Finally,the multi-layer perceptron is used to output the classification result.Compared with some classical classification algorithms,the model training time is shorter and the consumption of computing resources is reduced.