Research On Multi-scale Ship Target Detection Method For SAR Image

Synthetic Aperture Radar（SAR）is capable of penetrating clouds to stably and continuously image the sea,enabling all-day,all-weather ocean remote sensing monitoring.With its technical advantages,SAR plays an important role in marine environmental protection,marine economic construction,and homeland military defense.In view of the problem of different scales of surface ships,due to the failure to fully consider the multi-scale characteristics of ships,traditional target detection methods are prone to a large number of interference and misjudgment and result in false alarms and missed detection of multi-scale targets,which are difficult to adapt to complex and changeable detection scenarios.In order to solve the above problems,this thesis focuses on the multi-scale ship target detection for SAR images,and studies the multi-scale ship target detection methods for multi-resolution SAR images and SAR images of complex scenes.The main work of this thesis includes:1.The target characteristics and regional distribution characteristics of ships in SAR images are analyzed.On this basis,the multi-scale ship target detection framework of SAR images is studied,and the scene analysis of multi-scale ship target detection in multiresolution and complex scene SAR images is given respectively.2.Aiming at the multi-scale ship target detection problem for multi-resolution SAR images,the multi-scale ship target detection method for SAR images based on superpixel is proposed.This method generates homogeneous superpixels by aggregating pixels with similar spatial distance,gray and texture structure,and accurately divides the target region and background without limitation of image resolution,so as to realize robust detection of multi-scale ship targets for multi-resolution SAR images.3.Aiming at the problem of multi-scale ship target detection in SAR images of complex scenes,a multi-scale ship target detection method based on improved YOLOv4 is studied by combining the spectral residual model of visual saliency detection and the YOLO series model of deep learning target detection.By combining the frequency domain characteristics of spectral residual aboriginal map and the spatial domain characteristics represented by deep learning convolutional neural network,the fusion feature map is constructed to enhance the key features of the target and reduce target misjudgment,so as to realize the robust detection of multi-scale ship targets in SAR images of complex scenes.The above work has been verified by the measured SAR image data.Experiments show that the methods proposed in this thesis are effective and robust,which can achieve robust detection of multi-scale ship targets for multi-resolution SAR images and complex scene SAR images.