Research On Underwater Target Detection Method Based On Deep Learning

In the modern environment,the advancement and development of science and technology are accelerating,and the pace of mankind has begun to spread throughout the three-dimensional space.At the same time,as the natural resources on the ground are drying up,a large number of resources in the ocean have not yet been exploited for various reasons.Under this environment,underwater target detection,as a very important part in the field of marine research,has attracted non-mathematicians to invest in this work.At the same time,deep learning has outstanding performance in many fields.Therefore,the research work of this thesis is mainly to apply deep learning algorithms to the field of underwater target detection.The main work is as follows:First of all,in the process of underwater target detection,due to the complex underwater environment,it is difficult to extract the characteristics of underwater acoustic signals.By studying the relevant theories of convolutional neural networks in deep learning algorithms,this thesis analyzes the convolutional and pooling method for feature extraction,an adaptive convolution method for underwater acoustic signal feature extraction is proposed.On the basis of conventional convolution,this method introduces additional convolutional layers to enable the convolution kernel to adaptively change the sampling position and range of its receptive field according to different features through learning,significantly enhanced the ability of the convolution kernel to extract multi-dimensional features.Based on the convolutional neural network ignores the relationship between features,and proposes a weighted filtering method based on the channels of the feature map.The effect of each channel on the result in the multi-channel feature map extracted by the convolution kernel is changed.Degree to further improve the feature extraction effect of the adaptive convolution method.Then,for the problem that the boundary of underwater target detection is difficult to define,it is necessary to arrange the features extracted by the above-mentioned method to form a feature sequence according to the time axis,and then combine the features from the perspective of time series to obtain the change rule of the underwater acoustic signal characteristics to finally obtain the detection result.By analyzing the changes of underwater acoustic signal characteristics,a feature joint method for underwater target detection is proposed.Aiming at the problem of long training time of the short-term memory network,this method optimizes the structure of the loop unit,and proposes a fast loop unit,which speeds up the training of the model by reducing the serial calculation of the loop unit.Combining adaptive convolution operation with fast recurrent unit,an adaptive fast recurrent neural network model for underwater target detection is proposed.Finally,the method and model proposed in this thesis are experimentally verified on multiple data sets.Firstly,the effect of adaptive convolution and feature selection methods on traditional convolutional neural networks is verified using target underwater radiated noise under different conditions.Experiments were performed on the Hekou Reservoir dataset and the Songhuajiang dataset using an adaptive fast recurrent neural network model to verify the feasibility and effectiveness of the model.