Research On Underwater Target Classification Of Sonar Image

Underwater target classification of sonar image is one of the most challenging research topics of marine strategy.Because of the influence of lots of non-targets and shadows,underwater sonar image generally has low resolution and lots of noises.In addition,the low resolution would lead to poor classification accuracy of underwater sonar image.AdaBoost cascade algorithm is a commonly used machine learning classification algorithm to obtain the model by selecting the feature extraction algorithm and weak classifiers.However,this method is targeted to deal with the specific problem,but the generalization is poor,thus it is limited in practical application.Convolutional neural network is superior to massive data which doesn't need preset information.However,it has poor performance when it is applied to sonar image datasets with small amount of data.Above the problems,the research work in this paper is as follows:(1)Aiming at the question that sonar image classification has poor integrity and cannot be directly used for image classification.we make a sonar image classification dataset,which has category labels and consistent size.We use affine,rotate and zoom to simulate different imaging orientations of sonar.We use channel separation algorithms to simulate the influence of illumination on imaging.We add different noise types to simulate the noises of underwater environment.We use filters to simulate the lack of image features.The classifier performs well with the dataset made in this chapter.This shows that the dataset proposed in this chapter can be used to validate the effectively of models.(2)Aiming at solving the question about the low resolution,lots of noises and poor accuracy of underwater sonar image.We propose an improved AdaBoost cascade classifier framework.Firstly,histogram of oriented gradient is used to obtain the histogram features of the input images.Next,AdaBoost cascade framework is constructed with multiple support vector machines.These meta-classifiers are trained by using the extracted features and labels given by the dataset.Meanwhile,a new sample weight update function in cascade framework is proposed,which provide the meta-classifiers votes according to classification errors.Also,a new iteration rule is established,and when the meta-classifiers obtain the best classification accuracy,the train process will be finished.Finally,a random gradient descent algorithm is introduced to update the parameters of the model to ensure that the loss function converge gradually.The results on MNIST,Cifar-10 and sonar image dataset show that,the improved classifier framework can effectively distinguish images with low resolution and lots of noises.(3)Aiming at the question that the underwater sonar image dataset is not suitable for deep convolutional neural networks for its scant samples,we propose a feature-excitation convolutional neural network.On the one hand,depth-wise separable convolution is used to reduce the amount of network's parameters.Different from traditional convolution,depth-wise separable convolution performs depth-wise convolution and point-wise convolution in each input channel.On the other hand,the feature-excitation module is used to excavate the input image features.Feature channels which have great influence on classification are extracted from input images,in this manner,features from input images are used as maximized.The results on different classification datasets show that,the accuracy of feature-excitation convolutional neural network is higher than many neural networks used for classification,the feature-excitation convolutional neural network achieved effective classification of a small number of samples.