Research Of Ship Target Classification And Recognition Based On High Resolution SAR Images

Synthetic Aperture Rader(SAR)is an active microwave sensor that synthesizes an equivalent "large antenna" using the data-processing approach of a smaller real antenna based on the relative movement of the radar and the target,thereby obtaining a higher-resolution monitoring image.Compared with passive sensors such as infrared sensors and optical sensors,SAR imaging has been widely used in marine monitoring due to its high resolution,all-weather operation throughout the day and strong penetrating power.In this paper,aiming at the classification and recognition of ship targets based on high-resolution SAR images,the ship classification models based on ELM(Extreme Learning Machine)and KELM(Kernel Extreme Learning Machine)are constructed after the completion of high-resolution SAR image preprocessing,SAR image target feature extraction and feature selection.Considering the interference of human factors in the process of selecting classifier parameters,the continuous dragonfly algorithm is introduced to select the optimal kernel parameter and penalty factor for the KELM classifier.Further,in order to streamline the complicated feature selection work and avoid the influence of human factors and individual differences on the selection of optimal feature subsets,the discrete dragonfly algorithm is introduced to combine the automatically selection of optimal feature subset and automatically selection of optimal classifier parameters together.Finally,the effectiveness and advancement of the models are verified by comparison experiments based on high-resolution TerraSAR-X SAR image data sets and four multi-classification evaluation index.The main work and contribution of the dissertation are as follows:(1)In the process of high-resolution SAR images preprocessing,a two-dimensional OTSU threshold segmentation algorithm based on Radon transform is introduced.The Radon transform is used to determine the direction of the main axis of ship targets and a minimum bounding rectangle is determined around the main body of ship.Then,the two-dimensional OTSU threshold segmentation algorithm is used within the minimum bounding rectangle to effectively remove the sidelobe interference in the SAR image,thereby rapidly extracting the interested part,which is the main part of ship.(2)According to the definitions of commonly used identification features of SAR image ship targets,feature extraction is performed,including geometric features,gray statistical features,and local RCS density features.Further,in order to avoid the occurrence of "dimension disaster",a feature selection method based on Filter evaluation strategy and Wrapper evaluation strategy is used to perform feature selection,so that the optimal feature subset is selected.(3)Based on the sample feature vector containing tihe optimal feature subset,the ship classification models based on extreme learning machine and kernel extreme learning machine are constructed.Taking into account the impact of human factors interference in the parameter selection process on the model classification effect,a new type of swarm intelligence optimization algorithm—the dragonfly algorithm is introduced to find the optimal parameters for classifier.Finally,the validity of these models is verified by comparative experiments.(4)Taking into account the feature selection requires a lot of time and effort,and human factors and individual differences often have a great impact on the results of the optimal feature subset,discrete dragonfly algorithm is introduced to combine the automatically selection of optimal feature subset and automatically selection of optimal classifier parameters,and a ship classification model named BDA-KELM is proposed.The model simultaneously selects the optimal feature subsets and selects the optimal classifier parameters by automatic optimization process in discrete space of dragonfly individual The comparison experimental results show that the model can not only make the feature selection intelligent,but also can further improve the classification effect of the model.