Sonar Image Segmentation Based On Convolutional Neural Network

With the development of the ocean exploration,the demand for sonar image detection technology of our country has been improved continuously.How to acquire and utilize image information effectively has vital impact on the subsequent image recognition and analysis.The research on the target recognition in sonar image has become an important issue in the field of image processing.Image segmentation,as a key step in image recognition,has an important impact on the targets recognition.Since the sonar image is obtained by the sonar detection system,it has severe speckle noise.The purpose of sonar image segmentation is to extract the target bright and the dark region from the complex image reverberation noise effectively while preserving the original image information.At present,the improvement on the traditional segmentation method is the mainly method of sonar image segmentation,and the image segmentation accuracy needs to be improved.Through analysis,the existing sonar image segmentation method can not segment all kinds of sonar images effectively,but only can segment the specific types of sonar images,and the suppression on noise is not ideal.The most important thing is that it can not realize automatic image segmentation.With the continuous progress in the field of deep learning research,artificial neural networks are increasingly favored by scholars in the field of image processing.Convolutional Neural Networks(CNN)continue to make progress in the field of image classification and recognition.The emergence of Full Convolution Neural Network(FCN)has opened up a new direction for image segmentation.Therefore,the paper uses the Full Convolution Neural Network to study the sonar image segmentation deeply,and solves the defects of traditional sonar image segmentation.Firstly,the paper analyzes the characteristics of sonar images and obtains the statistical law of sonar image noise.An image filtering method combining dual-tree complex wavelet transform and adaptive median filtering is adopted.This filtering method can utilize the good performance of dual-tree complex wavelet transform and adaptive median filtering to remove the characteristics of impulse noise well.The method can filter the sonar image speckle noise effectively;secondly,the speckle noise statistical law of the sonar image obeys the Rayleigh distribution,and the Rayleigh noise can be represented by the average noise.Based on this characteristic,the sonar image training and testing dataset is established.Finally,Using Full Convolution Neural Network to accomplish sonar image segmentation.In order to achieve accurate segmentation of sonar image with speckle noise.Based on the FCN model,a Full Convolution Neural Network-Relative loss function(FCN-RLF)is proposed.The FCN has the advantage of accepting arbitrary size image input and retaining the original input image spatial information.Through the autonomous learning of FCN,the ability of the network to extract the features of sonar image can be improved.The Mean square error(FCN-MSE)learning rule is improved.The network training process takes both the target region and the non-target region pixel into accounts,and achieves the pixel level segmentation of the sonar image.Therefore,the big training error and the segmentation accurate problem are solved.The experiment results show that the improved algorithm improves sonar image segmentation accuracy,and the model is robust to the suppression of sonar image speckle noise.The improved method plays an important role in the special operation of marine target recognition and the accurate identification of submarine military targets.