Research On Feature Extraction Of Underwater Target Method Based On CNN

With the advancement of science and technology,the area of human activitiesis continually expanding.The detection of underwater target in the ocean has become a key technology for exploring the secrets of the ocean.In the past detection technology of underwater target,mainly rely on features both in time and frequency domains,the extraction process of these features are too dependent on human experience and subjective consciousness,then which makes features extraction technologies consistency become the technical difficulty of detection of underwater target.In the paper,the research is aimed at the background of this research,using the method of LoFAR Spectrum to effectively fuse the domain signal and the frequency domain signal,to achieve feature extraction method that basing on signal fusion.And aiming at the shortcoming of Convolutional Networks will lose the information of feature map space in fully connected operation,by introducing the multi-dimensional feature weighting algorithm for weighting spatial information.This algorithm uses the maximum detection hypothesis theory and the two-dimensional image entropy theory as evaluation standard of weight of the feature layer,and weighted algorithms of spacel and channel these two dimensions.Then put forward the algorithm flow of detection of underwater target feature based the multi-dimensional feature weighting network.In order to further enhance the extraction ability of local feature of Convolutional Networks,the paper continues to analyze the limitation that linear convolution filter handles linear non separable features,and over complete linear filter banks can also produce redundancy of feature extraction,and then bring the burden to the network training.Thef combining with nonlinear processing ability of radial basis neural network,proposing a RICNN network model.In this model,through introducing local regularization layer to solve the influence of the increase in depth on the convergence speed of the neural network.The strategy of global average pooling replaces fully connected layer in traditional Convolutional Networks,further resolving the over-fitting phenomenon of the model.At the same time,the network model takes embedded Radial Basis Function Networks as universal nonlinear filter,which enhances extracting abilityof local feature of Convolutional Networks.And through specificexperimental simulation method to verify the effect of characteristic graphMultidimensional weighted convolutional neural network and RICNN model that are proposed in this paper in feature extraction.Through the research of this paper,not only achieving the efficient fusion of time domain and frequency domain signals,and at the same time effectively verifying the effect of in feature extraction of characteristic graph Multidimensional weighted convolutional neural network and RICNN model,making the research of this paper has very important research significance in the exploring feature extraction of underwater target.