Research On Classification Algorithm Of Seabed Sediment Based On Metric Learning

With the development of high-resolution imaging sonar technology,seabed sediment classification technology based on sonar image has become one of the mainstream sediment recognition technologies.Many scholars use SVM,KNN,K-means,random forest,neural network and other algorithms to classify seabed sonar images,and have achieved good results.In particular,KNN,K-means and other methods that rely on similarity measurement between sediment features overcome the data skew caused by uneven distribution of seabed sediment under certain circumstances.However,the seabed situation is complex.Once the difference in the number of samples of each sediment in the data set is too large,the classification accuracy will be greatly affected.Therefore,finding an appropriate measure and re-planning the distribution of sediment samples is of great significance to make up for the shortcomings of existing algorithms in seabed sediment classification and improve the classification and recognition ability of seabed sediment.In view of the characteristics of seabed sediment data and the shortcomings of some traditional algorithms relying on similarity measurement in seabed sediment classification,this paper applies the metric learning method to seabed sediment classification,and improves the distribution of seabed sediment characteristics and classification accuracy through the metric learning algorithm.Firstly,the original side scan sonar data is processed to realize image distortion correction and gain compensation,so as to ensure the data reliability for subsequent research.Based on this,the data set is made.The statistical method,model method and signal processing method are used to extract the characteristics of seabed sediment.On this basis,the multi feature extraction method of seabed sediment is studied to extract the characteristics of seabed sediment to the greatest extent.The filtering method is used to filter the redundant and irrelevant features in the seabed sediment feature set,complete the sediment feature selection,and provide accurate and reliable seabed sediment feature data for subsequent experiments.Then,aiming at the shortcomings of some traditional algorithms relying on similarity measurement in seabed sediment classification,this paper applies the idea of linear metric learning to seabed sediment classification.From the perspective of statistics and actual distance information of seabed sediment characteristics,ITML,a metric learning algorithm with low training complexity,and LMNN,a metric learning algorithm with high performance,are used to optimize the spatial distribution of seabed sediment characteristics,increase seabed sediment identification and improve the classification effect.The experimental results show that the two linear metric learning algorithms are feasible in seabed sediment classification,and the seabed sediment classification algorithm based on LMNN has better discrimination and recognition ability.Finally,aiming at the problem that the linear metric learning algorithm LMNN can not distinguish the class boundary element points in the seabed sediment feature space,combined with the actual classification results,this paper analyzes the limitations of linear metric learning,and applies the idea of nonlinear metric learning to seabed sediment classification.The nonlinear metric learning algorithm GB-LMNN is used to optimize the seabed sediment classification algorithm.At the same time,the linear metric learning algorithm LMNN and the robust structural metric learning algorithm R-MLR ignoring redundant feature interference are used as the comparative model of GB-LMNN.The experimental results show that GB-LMNN algorithm has a better effect on the spatial distribution of seabed sediment features.