Research On Underwater Complex Target Detection Based On Line Spectrum

In the era of continuous development of science and technology,people's living areas have been continuously expanded and they have continued to develop from the land to the sea.However,the underwater environment is relatively complex,so effective target detection becomes an important research content.In the past,the target detection mainly depends on people's experience.The result to be obtained is subjective,and at the same time it has higher requirements on people's professional knowledge.In this regard,the main content that needs to be studied is how to automatically detect the target.Based on this background,this thesis conducted a research analysis.This thesis mainly introduces two parts.The first part is the extraction of the line spectrum.The second part is the detection of the extracted line spectrum.The properties of the line spectrum can well represent the type of the target.Therefore,the main purpose of the online spectrum extraction is to extract the line spectrum more completely.In this thesis,the method of line spectrum extraction is improved.When the LOFAR map is generated,adding a logarithmic function can display the spectrum of strong and weak lines within a limited dynamic range.When the line spectrum is extracted,an integral function is used to quantify the threshold.Reduce human intervention.After the line spectrum is extracted,how to detect it is also a very important step.In the research of this thesis,a convolutional neural network is used for detection.The traditional convolutional network has some deficiencies.Training requires a lot of The time and parameters are more dependent on the hardware.Therefore,how to reduce the calculation cost and parameter capacity is the focus of this thesis.In response to this situation,a method of Fast and Compact in Convolutional Neural Networks model is proposed in which the convolution kernel is decomposed using a low-order expansion approach.Decomposing a layer of network into form of Conv_de1 and Conv_de2 and reducing computational costs by reducing the number of Conv_de1.At the same time,the method of Adaptive Drop_weight was introduced to discard the redundant weight parameters.This reduces the number of parameters.In the convolutional neural network model,parameters are very large for the fully connected layer,and this is the main source of parameter capacity.Similarly,an adaptive Drop_weight algorithm is used to discard the weight less than a certain threshold,so that the number of parameters will be significantly reduced.Through the research of this thesis,the effective extraction of line spectrum is realized.In the process of model training,the cost of calculation was also reduced,and a relatively good result was also obtained in the accuracy of detection.Therefore,the research in this thesis has important significance for the target detection of underwater complex environment.