Research On Ship Target Detection Method For SAR Images

Synthetic aperture radar(SAR)has become a widely used remote sensing geographic information acquisition technique due to its unique detection advantages of long distance,high resolution and strong meteorological adaptability,which has important value in both military and civilian applications.In view of the multi-scale ship target detection,the complex background in ports area,and the contradiction between the detection performance and the constraint of platforms’ resource,this dissertation conducts research on ship target detection for SAR images,which focuses on more robust ship detection methods based on clutter statistical characteristics,intelligent ship detection methods based on deep learning,and data generation methods based on semi-supervised adversarial networks.We expect to develop a more efficient and accurate SAR image ship interpretation system to achieve efficient and lightweight ship target detection in complex electromagnetic environment.The main work of this dissertation can be summarized as the following four points:1.With the improvement of synthetic aperture radar image resolution,the clutter characteristics of the sea surface become more and more complex,which brings new challenges to traditional ship target detection methods.In order to achieve fast and efficient detection of ship targets in complex backgrounds,a hierarchical detection method based on a "coarse-to-fine" detection mechanism is proposed.First,a visual attention mechanism combining image domain and frequency domain is constructed.The candidate ship target is obtained adaptively through the mean dichotomy method.On this basis,the accelerated regional kernel density estimation method is used to obtain the precise detection result of the target,so that the detection result has the characteristic of constant false alarm.The proposed method is simple,efficient and accurate.2.Traditional methods based on constant false alarm rate(CFAR)cannot work properly in complex conditions such as multi-scale situations or offshore ship detection.With the development of deep learning,methods based on convolutional neural network(CNN)have been widely used.However,the detection of multi-scale SAR ships is still a great challenge due to the diversity of target scales and the strong interference of near-shore backgrounds.Furthermore,most CNN-based ship object detectors focus on feature attention mechanisms,while ignoring the relationships between different feature levels.This dissertation proposes a deep feature fusion network for ship target detection(Deep Feature Fusion Detector,DFDet)in high-resolution SAR images.DFDet is a single-stage fully convolutional anchorfree detector designed to improve the performance of multi-scale ship detection under complex conditions by fusing feature maps from different levels and capturing the interplay between spatial and channel dimensions.Furthermore,a deformable residual feature refinement network is constructed to extract rotated non-local semantic information.3.With the development of deep learning,the SAR ship target detection method based on convolutional neural network has achieved good performance.However,SAR image data is relatively small,and the use of SAR images to achieve target detection and recognition faces the dilemma of small samples.To solve this problem,this dissertation proposes a new method of SAR sample augmentation based on semi-supervised adversarial network.The proposed method enhances scene-aware information through a two-port input form,then constructs a high-resolution attention U-Net generation network to generate directional new samples,and then uses a local extremum fully convolutional discriminator network to improve scene realism.The proposed method can achieve multi-type complex target scene through a novel style embedding form,and the data generation experiment verifies the effectiveness of the method in this dissertation.4.The current state-of-the-art convolutional neural network-based ship object detectors mainly focus on detection performance while ignoring computational complexity.To solve this problem,this dissertation proposes a lightweight densely connected sparsely activated ship object detection network(Densely Connected Sparsely Activated Detector,DSDet).First,a densely connected sparsely activated network is constructed.Then,a lightweight feature pyramid fusion network is constructed.On this basis,various loss functions are utilized to improve the detection performance.Finally,the experimental part verifies that the proposed method has excellent performance in both detection performance and model computational complexity.