Font Size: a A A

Research On Sonar Image Recognition Based On Convolutional Neural Network And Ensemble Learning

Posted on:2024-05-31Degree:MasterType:Thesis
Country:ChinaCandidate:Y B DuFull Text:PDF
GTID:2530307157950999Subject:Information and Communication Engineering
Abstract/Summary:PDF Full Text Request
Accurate identification of underwater targets plays a vital role in the field of underwater unmanned exploration.It has a very wide range of applications in marine surveying and mapping and exploration,seabed rescue,marine operations,and exploration of oil and gas.Sonar images usually have the characteristics of limited available information and few samples.The recognition of sonar images using traditional machine learning methods is not only autonomous and inefficient,but also unsatisfactory for high-level feature extraction,which makes it difficult for traditional sonar target recognition methods to achieve better recognition results.At present,convolutional neural networks have been applied in the field of underwater target recognition.However,too deep convolutional neural networks have problems such as large models,high hardware requirements,and low recognition speed.These problems also restrict the application of deep learning network model in reality.Therefore,according to the characteristics of sonar image and its wide demand in military and civilian fields,this thesis studies the sonar image recognition method based on convolutional neural network and ensemble learning to improve the accuracy of underwater target recognition under the condition of ensuring reasonable recognition speed.The main research and work of this thesis are as follows :Firstly,the related theories of convolutional neural network and ensemble learning are studied.Then,an image preprocessing method combining YOLO-Fastest fast batch sonar image target location clipping and visual saliency detection is designed.This method can effectively reduce the interference of the input image background on the target feature extraction process,focus on the valuable areas in the image,and improve the data processing efficiency.The preprocessed image is used as the input of the subsequent feature extraction network.Secondly,aiming at the problems that the existing large-scale deep convolutional neural network model is too large and is not conducive to the deployment of embedded hardware platforms,this paper studies the current lightweight convolutional neural network Efficient Net-B0 baseline model.Aiming at the characteristics of less effective information in sonar images,spatial feature pyramid pooling SPP is introduced to enhance feature extraction and local high-level semantic multi-scale feature fusion,which improves the recognition accuracy.Aiming at the problem of low recognition accuracy in sonar image small sample training,a sonar image recognition model based on convolutional neural network and ensemble learning is studied and designed.The model strong classifier is integrated by logistic regression,support vector machine and decision tree.It can not only obtain better results than a single classifier,but also reduce the dependence of the training process on the required data.In the case of less training samples of sonar images,its superiority is more prominent,making the model more consistent with small sample sonar images.Based on this,this paper designs a sonar image recognition model based on improved Efficient Net and ensemble learning,a sonar image recognition model based on improved Mobile Net V3 and ensemble learning,and a sonar image recognition model based on joint dual model and ensemble learning..Finally,The experimental results show that the sonar image recognition method based on convolutional neural network and ensemble learning can effectively improve the recognition accuracy on the measured sonar image data set,and the highest recognition accuracy can reach 99.5 %.The trained sonar image recognition model based on improved Efficient Net and ensemble learning is transplanted to the embedded hardware platform Jetson Nano and accelerated by Tensor RT.The average recognition accuracy of the network model is 99.1 %,and the average recognition time of each image is about 168 ms,which meets the real-time requirements of embedded platform deployment.
Keywords/Search Tags:Convolutional neural network, Significance detection, EfficientNet, Ensemble learning, Sonar image recognition
PDF Full Text Request
Related items