Classification Of Seabed Sediment Sonar Image Based On Deep Learning Model Fusion

Seabed sediment classification is an important part of marine exploration.It is a prerequisite for the development of marine industry,the rational use of marine resources and the maintenance of marine ecological environment.Compared with the early mechanical drilling method,the classification of seabed sediment by sonar image is more in line with the concept of sustainable development in today's society,and it is more automatic and intelligent,which has a huge research space and application prospect.At present,most of the classification methods based on sonar images use the traditional machine learning method.This kind of method needs to extract features manually,which is greatly affected by human subjective factors,and the operation is complex,and the classification accuracy needs to be improved.Aiming at these problems,this paper puts forward a complete system from the beginning of sonar image preprocessing to the final completion of the classification of sand,reef and mud,and achieves the corresponding results,which has a certain technical reference value in the field of seabed exploration.The main contributions are summarized as follows:1.Due to the influence of environmental factors,seabed sonar image usually has a lot of interference information.In order to make it better used in subsequent experiments,Gaussian filter,Median filter and Wavelet Transform filter are used to denoise the seabed sonar image,then the SNR,PSNR and EPI are introduced to evaluate the denoising results.In the end,the Gaussian filter with the best effect is selected as the final preprocessing method.2.In order to solve the problem that the small sample data of seabed sediment sonar image is not conducive to the training of deep learning model,the transfer learning is applied to the training of seabed sediment sonar image.The Inception V3 model is used as a pre-training model.By freezing the parameters before the bottleneck layer,and accessing the new fully connected layer and softmax classifier to train the model,it effectively solves the model training problem caused by small samples,with an accuracy rate of 93.6 %,which provides a basis for follow-up experiments.3.The features extracted by a single model are relatively single.In order to further improve the classification accuracy,the three models of Inception V3,Mobile Net and Res Net50 are trained by the method of transfer learning,and their features are fused to obtain four fusion models.By comparing the classification accuracy and parameters of the seven models,it is determined that the fusion model of Inception V3 and Mobile Net as the optimal model,and the weighted fusion method is further optimized.The final overall classification accuracy rate is 97.26%,and the classification of reef,sand and mud is relatively stable.