Small Target Detection In The Background Of Sea Clutter Using Deep Learning Method

In our country the area of waters is widely and the coastline is longer than others.It has a great significance for us to identify the target,detect the buoy,spilled oil and other small target signal and ensure the safety of navigation and so on.This paper first discusses the chaos under the background of weak signal detection,simulate the actual situation and add noise to Chaos observation sequence to analysis.This paper then put forward two kinds of signal to detect small targets under the background of sea clutter model.They are respectively Deep Belief Networks and Stacked Auto-encoder neural network method which can detect the small target under the background of sea clutter.We present a weak signal detection method based on Deep Belief Networks,because the noise signal which has a characteristics of chaotic will cause large error of prediction.First of all,the method used the unlabeled data to the Restricted Boltzmann Machine with the unsupervised greedy training layer by layer,in order to get the connection weights between the visual layer and the hidden layer.After the unsupervised training is completed,each layer of Restricted Boltzmann Machine can be adjusted to a suitable initial value.Then made a combination of multiple Restricted Boltzmann Machine through the bottom-up,Constructed a preliminary Deep Belief Networks model.Finally,using the error back-propagation algorithm(BP)optimized the parameters of the whole network based on Deep Belief Networks model,establish Deep Belief Networks prediction model and detect weak signals.The simulation experiment is carried out by using Lorenz system as chaotic background,this method can detect the weak signal of the lower amplitude the chaotic noise background and it has a better prediction accuracy.Deep Belief Networks model also can effectively restrain the noise signal,and has a strong anti-noise performance.Then the paper put forward a weak signal detection method by Stacked Auto-encoder neural network under the background of chaos.First of all to preliminary training of network,the unsupervised training since the encoder greed to obtain the optimal value.After the complement of unsupervised training,the network parameters are just the optimal value of each layer when it is trained independently.Then training for network tuning by using the BP algorithm training the overall network parameters in order to obtain the global optimal solution,build the stacked encoding neural network detection model.Then use the model to detect transient signal and periodic signal.Compared with Deep Belief Networks model,the results show that under the condition of gradually increasing implicit layers,in view of the weak signal detection,the mean square error which output by the stacked encoding neural network detection model were similar.That means Stacked Auto-encoder neural network has better stability,and the model can effectively suppress noise signal.Compared with traditional method,the two methods can detect weak signal in chaotic background.This paper puts forward the Deep Belief Networks and Stacked Auto-encoder methods of the small target detection according to the chaotic characteristic of sea clutter and related theories.Deep Belief Networks and Stacked Auto-encoder neural network methods to unsupervised training method to optimize the weights of network of initial value,and not as random initial value as the traditional neural network initialization of network weights,which is not easy to fall into local optimum.This paper first analyzes the chaotic characteristic of sea clutter and the related research situation and then uses the two methods to construct the sea clutter signal detection of small target model.Finally uses the IPIX radar data which had measured by Canadian McMaster in the experiment and evaluate the using performance by the root mean square error.The experimental results show that,in view of the 54#sea clutter data,and the existing selective root mean square error of 0.0264 support vector machine(SVM)integration method,k-means-effectively the root mean square error of extreme learning machine income compared to 0.0428,this paper puts forward a stack of adaptive coding predicted the root mean square error of 0.016,prediction accuracy improved.But both of the two methods need more time to train.Both of the two methods which studied by this research can realize the detection of small target under different sea condition.