Research On Method Of Edge Feature Extraction For Side-scan Sonar Images Based On Deep Learning

Due to the particularity of side-scan sonar imaging principle and the interference of underwater detection environment,compared with optical images,sonar images have serious speckle noise pollution and low imaging resolution,which makes feature extraction of sonar images very difficult.The traditional sonar images feature extraction methods have poor anti-noise capability,and the selection of features generally depends on manual work,resulting in low efficiency and poor generalization capability.In order to overcome the shortcomings of traditional methods,this paper proposes a method of feature extraction for side-scan sonar images based on deep learning,and verifies the effectiveness of the deep learning algorithm and its improved algorithm through experiments.Firstly,sonar image data set required for feature extraction of deep learning system is constructed.According to the working principle of side-scan sonar,the original sonar data is analyzed to obtain the seabed topographic map of the real sonar image;Based on the analysis of noise model of sonar images,and the image tag data used for feature extraction network training and testing are produced;Secondly,based on the analysis of the basic structure and typical characteristics of convolutional neural network,the fully convolutional neural networks model(FCN)is constructed.Three network models of FCN are trained respectively by the mini-batch gradient descent method with momentum items,and the effects of different batch sizes on the stability of FCN network and the accuracy of feature extraction are analyzed through experiments;The three trained network models are used to extract the edge contour of submarine topography respectively,and the advantages and disadvantages of feature extraction results of different network models are compared and analyzed;Thirdly,aiming at the problem that the imbalanced classification of edge sample data of submarine topography will reduce the performance of the classification algorithm,a function named Weighted the Positive Samples with the Softmax Loss(WPSL)is proposed to balance the proportion of positive and negative samples in the loss function by adding weights to the positive samples,so as to improve the impact of the sample imbalance on the classification performance;Then,aiming at the shortcomings of the FCN network,the following improvement schemes are proposed: Firstly,aiming at the problem that the skip structure of FCN network easily leads to insufficient network training,an improved design of skip structure is realized by adding Batch Normlization layer to the skip structure,and a fully convolutional neural network model(FCNB)is constructed based on the improved skip structure.By adding batch normalization layer to the two skip structures of FCN network,the re-parameterization of the skip structure is realized to ensure that the gradient is more predictive and stable,and the analytic performance of "skip" network is significantly improved,so that the network can get more sufficient training;Secondly,aiming at the problem that the pooling layer contains less detailed information,a fully convolutional neural network with richer fused features(FCNR)is constructed by using the strategy of fusing all convolutional layer information in the "convolution group" instead of the strategy of fusing only pooling layer information in FCN.The convolution layer contains more edge detail information than the pooling layer,and the complementary edge information can be realized between different convolution layers in the combination,so the FCNR network can retain more detailed information of seabed topography,can more accurately divide non-edge pixels near the edge,has thinner edge lines and more accurate positioning capability;Thirdly,aiming at the problem of parameter redundancy and large dimension span between convolution layers in FCNR network,a lightweight fully convolutional neural networks with richer fused features(LFCNR)is constructed;The last,in order to further improve the performance of the network,the lightweight fully convolutional neural networks which combines batch normalization layer and richer fused features(LFCNBR)is constructed by combining the improved skip structure with the LFCNR network;Finally,all the improved networks are trained by using the mini-batch gradient descent method with momentum items.At the same time,several groups of comparative experiments are designed to verify the algorithm.Through the experimental comparison between FCN and the improved algorithm,it is proved that the improved scheme proposed in this paper has improved the classification performance of the network to a certain extent.Among them,the method combining WPSL loss function with LFCNBR network has the best performance,with the Mean Intersection Over Unit(MIU)reaching 87.33%.Among them,the WPSL loss function combined with LFCNBR network has the best performance,with the average area coincidence rate reaching 87.33%.It can still accurately extract the edge features of seabed topography under the conditions of fuzzy edge lines and complex background.The comparison between the deep learning method and the traditional method proves that the accuracy of the deep learning algorithm is much higher than that of the traditional algorithm,and it still has strong ability to resist speckle noise without denoising.