Research On Sea Clutter Suppression And Small Target Detection Based On Deep Learning

The exploration of marine resources and the protection of marine boundaries will inevitably encounter the detection of small ships and obstacles,and there are often a large number of wave clutter in the sea surface radar echo,which makes it difficult to find the target in time and accurately.Therefore,how to suppress this kind of marine clutter to improve the detection performance effectively is very important.Convolutional neural network(CNN)has unique advantages in the face of such image processing,which can greatly improve the detection performance.This thesis mainly focuses on how to effectively reduce the clutter noise and identify the correct object when detecting small targets in the scene with poor sea conditions,and reduce the detection false alarm rate at the same time.The main work contents are as follows:(1)Due to the complex characteristics of sea clutter and the difficulty of data acquisition,the amount of data cannot meet the requirements of deep learning,so it is necessary to model and generate more data that conforms to the distribution.In this thesis,four clutter distribution models are constructed according to different sea conditions,and the simulation data are compared with the real signals to select the composite K distribution model with the best fitting effect.Finally,the generative adversarial network is trained based on real data and its performance is analyzed through comparative experiments.The results show that this method can better simulate clutter data than the composite K distribution model.(2)Study various physical properties of ice multi-parameter imaging X-band(IPIX)measured radar data and process it to obtain time-frequency domain datasets.Convolutional autoencoder(CAE)is constructed and trained by sea clutter data in fast and slow time domain and its performance is compared in various sea conditions.Finally,the real clutter data is changed to the time-frequency domain to construct a sample set training encoder to solve the problem that the time-domain data cannot suppress clutter under low sea conditions.The simulation results show that the suppression method has better performance in the time-frequency domain.(3)In this thesis,an algorithm framework combining CAE and visual geometry group16(VGG16)optimization network is constructed to realize clutter suppression and small target detection simultaneously.Firstly,the CAE is used to suppress most of the clutter in real time-frequency domain data,and then constructing the training sets of label and classification.Secondly,the shallow network of VGG16 is frozen and a new module is connected,which reduced 48% of training parameters and improved the convergence speed.Finally,the performance of VGG16 network is compared with that of original VGG16 network before and after clutter suppression and structure optimization under different sea conditions,and the performance of VGG16 network is compared with that of traditional detector under different polarization modes and average signal-to-clutter ratio conditions.The experimental results show that the algorithm framework can improve the detection performance and obtain stronger generalization while reducing the model size.