Pipeline Identification Method In Side-scan Sonar Image

With the continuous excavation of seabed resources,submarine pipelines play an increasingly important role in oil and gas transportation.The laying and daily maintenance of submarine pipelines inevitably need to be positioned first.Pipeline detection in shallow water mainly relies on divers to work deep underwater.It is difficult and dangerous.In deep sea,it mainly relies on the vision and distribution of underwater robots.Optical fiber sensing system.However,the above-mentioned operation can achieve local pipeline detection in close range,and can not complete macro-detection task in long range.Side-scan sonar system uses the return information of sound wave to form an image,which has a high rate of separation and can realize long-distance observation of pipelines.Therefore,this paper takes the pipelines in side-scan sonar image as the research object,and mainly completes the online automatic identification task of submarine pipelines.Traditional pipeline recognition methods need to extract pipeline features manually.When pipeline is buried by sediment,it can not be recognized.At present,there is no recognition for curved pipeline.Therefore,in view of the shortcomings of the above methods,this paper proposes a pipeline recognition method based on DenseNet and position fitting.Plenty of comparative experimental results show that the proposed method is accurate in the field of pipeline identification.The method with the highest and the strongest anti-interference ability.The main research work is as follows:Firstly,the class data set is established and expanded for the neural network.The recognition effect of neural network depends largely on the learning information provided by the data set.In order to improve the training accuracy,it is necessary to expand the data and make "duplicates" which have certain differences and are real and effective.Because the resolution of submarine pipeline in side-scan sonar image is higher and clearer,the training samples mainly come from side-scan sonar image.However,the acquisition of side-scan sonar images is difficult and costly,so it is proposed to transfer optical pipeline images to side-scan sonar images using neural style migration.In addition to the above expansion methods,this paper also uses affine transformation,elastic distortion and other traditional data expansion methods to provide more training samples for recognition network.Secondly,aiming at the problem of incomplete feature extraction by BP neural network,this paper proposes to input image pixels into the network and use BP neural network to automatically extract features,so as to improve the classification accuracy of the network.In order to further improve the accuracy of local pipeline recognition,DenseNet,which has better performance on common data sets,is used to classify data.Thirdly,the data need to be fitted further on the basis of neural network classification.Because of some errors in the classification results,the traditional straight line fitting method will greatly reduce the fitting accuracy due to the existence of outliers.Therefore,this paper uses robust fitting algorithm to fit straight lines and curved pipelines respectively.Improved Fish swarm algorithm and RANSAC algorithm are used for line fitting,and fast Fourier curve approximation method is used for curve fitting.Finally,the method of pipeline identification based on Hough transform and the method based on full convolution neural network are compared with the method proposed in this paper.The experimental results show that the method has the strongest anti-jamming ability and the highest recognition accuracy.