Research On Image Segmentation Of Sonar Image Based On Active Learning And Semi-supervised Learning

Sonar is the only equipment that people have at present that can carry out long-distance detection in the ocean,and it provides solid data support for the detection and recognition of underwater targets.However,due to the complex marine environment and various physical interferences in the propagation process of acoustic signals,the accuracy of traditional sonar image segmentation research is not high,and it depends on the noise characteristics of specific application environments.In addition,with the widespread application of sonar equipment,the increasingly large sonar image database has promoted the improvement of traditional algorithms to meet the new requirements of the new era in terms of processing efficiency,performance and intelligence.As a new type of high-efficiency and high-precision learning method,deep learning can effectively make up for the above technical limitations.Based on the above ideas,this article first designs a sonar image segmentation framework based on semi-supervised learning.Through data enhancement techniques,the training process of the existing semi-supervised learning model is improved,the performance of the model is enhanced,and the computational cost is reduced.In extreme cases A regression accuracy rate(m AP)of 50.91 can be reached using only 40% of labeled data.This essay designs a corresponding anti-noise module for the common ocean noise in sonar images.Finally,the training results on the noise samples show that the newly added sidelobe noise generation module makes the framework have good anti-noise robustness,but for The difficult cases in some data sets still cannot be effectively identified.At the same time,in order to solve the shortcomings of deep learning that the process of labeling data is expensive,tedious and time-consuming.In this essay,the active learning method is used to further improve the segmentation accuracy of the deep learning framework for sonar images,reduce the cost of sample labeling,and solve the problem that some difficult cases cannot be identified after the semi-supervised learning model converges.This essay uses five types of classification information or space The active learning method of information is integrated into Mask RCNN,and the active recognition framework of sonar image is designed.Through experiments on the collected sonar image data set,this essay verifies that the introduction of active learning technology can improve the recognition performance of the model.In addition,this article verifies that the framework can help experimenters use less data for training,and in the best case,only 54% of the labeled data can be used to train to close to convergence.Finally,this article combines the active learning framework with the semi-supervised learning framework mentioned above.The experimental results show that active learning can effectively help the model learn and master difficult cases when combined with the semisupervised learning model,so that the performance of the trained model is higher than Pure semi-supervised learning or active learning.