Study On Sonar Images Segmentation Technology Based On Generative Adversarial Network

Sonar imaging is an important tool for underwater environment analysis and plays a crucial role in underwater target detection and recognition.However,current sonar image segmentation methods are mainly focused on traditional algorithms,which often suffer from limitations such as model constraints and poor robustness.With the continuous development of deep learning techniques,incorporating deep learning into sonar image segmentation as an intelligent form of technology can significantly improve the accuracy of sonar image segmentation.In this study,we focus on generative adversarial networks(GAN)and make reasonable improvements to apply them to sonar image segmentation.Firstly,we construct a sonar image segmentation experimental dataset.Due to the lack of publicly available sonar image datasets,it is necessary to create matching training and testing sets for deep learning models.We collected 1000 real measured original images of underwater targets with different types using synthetic aperture sonar devices at the bottom of a reservoir,and then manually annotated all the images using open-source software Label Me to obtain a dataset for subsequent model training and testing.This ensures the rationality and availability of the experimental dataset.Secondly,we improve the U-Net network as the generative model for sonar image segmentation.The improved network uses even-sized convolutional kernels instead of odd-sized kernels for feature extraction,which reduces the number of training parameters and improves the training efficiency of the model.At the same time,an attention mechanism is introduced into the encoding-decoding structure to make the model more focused on important features and reduce the interference of unimportant features on feature extraction.In addition,multi-scale convolution is also introduced to enlarge the receptive field,allowing the model to better capture features of different scales,thereby more effectively extracting features of the target region.Through these improvements,we optimize the U-Net segmentation generative model and significantly improve the segmentation performance.Finally,in this stage,we use the improved generative model based on the previous stage’s research as the foundation,and construct a generative adversarial network for sonar image segmentation experiments.In addition,residual networks are introduced to further enhance the performance of the generative adversarial network.The introduction of residual networks helps the model better learn the features of sonar images,thereby improving the accuracy and robustness of segmentation.At the same time,residual networks can effectively address the problem of network degradation and improve the overfitting issue caused by the increase in network depth.The experiments of this study are conducted using the constructed sonar image segmentation dataset,and qualitative and quantitative data analysis is performed on the same dataset with some classical image segmentation algorithm models.The final experimental results show that using the improved generative adversarial network can effectively improve the performance of sonar image segmentation.