Research On Seabed Target Automatic Recognition Based On Deep Learning

In recent years,with the increasing attention of various countries to ocean exploitation,seabed target detection and recognition has gradually become a research hotspot,playing an important role in military confrontation,seabed engineering,seabed resource exploitation and seabed detection.At present,a great deal of research has been done on the recognition of seabed artificial targets,but there is a lack of a versatile method to realize the recognition of different types of targets,while the recognition of seabed natural targets is relatively rare at home and abroad.As a popular technology in the near future,deep learning can learn high-level features autonomously and realize end-to-end training and recognition.Applying deep learning to target recognition with large amount of ocean surveying and mapping data can avoid the problem of artificial features losing information,thus improving recognition efficiency.This paper mainly studies the automatic recognition of seabed artificial targets and natural targets.Taking three types of samples of artificial targets such as aircraft,wreck and drowning victim and pockmarks of natural target as example,the feasibility of deep learning in seabed target recognition is discussedFor the research of seabed artificial target recognition,this paper studies the recognition of sunken ship,aircraft and drowning victim based on side-scan sonar data set,expands the model training samples through data enhancement,and uses a semi-synthetic method to generate images of aircraft and drowning victim to improve the problem of sample category imbalance.The classification and recognition method based on deep neural network and the target detection and recognition method based on SSD(Single Shot Multibox Detector)model are studied respectively.Using transfer learning to load pre-training model parameters to fine-tune and accelerate the convergence of the model,the experimental results show that semi-synthetic data can improve the model recognition accuracy,and the classification recognition accuracy based on single target sonar image is higher than the target detection accuracy,and the accuracy and recall rate of SSD model exceed 90%,thus obtaining better recognition results.For the research of seabed natural target recognition,pockmarks landform is selected as the research object.Pockmarks data is the data of the southeastern South China Sea.After processing the water depth data,output grid and elevation rendering map,a semi-automatic identification method for seabed pockmarks is proposed.Through preliminary recognition based on hydrological analysis,pockmarks feature extraction and screening,elimination of false pockmarks and manual assistance discrimination,the classification,location and boundary information of 220 pockmarks in the study area are obtained,and the results are used as basic data for deep learning model training Finally,after dividing the pockmarks area into blocks,SSD model is used to train the marked pockmarks data to realize end-to-end recognition.Two typical blocks are selected for verification.The results show that deep learning can realize automatic recognition of seabed pockmarks,but the recognition results are poor for smaller pockmarks and poorer pockmarks.Comparing and analyzing the recognition methods,the semi-automatic recognition of pockmarks based on hydrological analysis is better than other method in this study area,while it is feasible to apply deep learning to the recognition of seabed pockmarks and there is a large space for improvement.