Research On Application Technology Of Superpixel Segmentation And Target Detection In SAR Image Based On Deep Learning

Synthetic Aperture Radar(SAR)is a remote sensing technique used to generate high-resolution images by transmitting and receiving microwave pulses.It has significant application value in the field of environmental monitoring.In recent years,research based on SAR images has received widespread attention.However,the pixel-level algorithm has been difficult to meet the actual needs of massive SAR image fast processing.Superpixel segmentation can divide complex images into a small number of pixel sets,effectively reducing the computational complexity of image processing and interpretation,and improving the efficiency of algorithms.In addition,it can preserve information in the image,reducing the interference of noise on individual pixels in the image.However,the low dimensionality and complex structure of SAR image information make it difficult for existing superpixel segmentation algorithms to obtain accurate results on SAR images.Therefore,this thesis proposes an efficient and accurate superpixel segmentation algorithm for SAR images,which can effectively solve these problems.Superpixels can effectively reduce computational complexity and preserve image information,and therefore have been widely applied in the field of image interpretation.Based on the proposed superpixel segmentation algorithm for SAR images,This thesis explores the corresponding technical means and algorithms to apply superpixels of SAR images to ship detection.Currently,advanced SAR ship detection algorithms are mostly implemented using convolutional neural network models.However,the deep learning methods usually require a large number of annotated samples for training,resulting in high human labor costs.Therefore,this thesis proposes an unsupervised ship detection method without manual annotation of training atlas.In summary,this thesis conducts research in two aspects: SAR image superpixel segmentation and ship target detection based on superpixels of SAR images.The main research contents and innovations of this thesis are as follows:(1)A multi-task learning-based SAR image superpixel segmentation method is proposed.First,the deep neural network’s powerful feature extraction capability is used to extract deep features of the image,and the high-dimensional feature space is constructed by combining the SAR image intensity and spatial information to increase the pixel feature dimension in this way.Then,based on this high-dimensional feature space,a pixel distance measure that is robust to speckle noise in SAR images is defined to measure the distance between pixels.In addition,a differentiable soft assignment operation is designed to replace the nearest neighbor operation of simple linear iterative clustering(SLIC),enabling differentiable SLIC and feature extractors to be combined into an end-to-end superpixel generation network.Multi-task learning strategy is used to train the network and solve the problem of insufficient labeled data.(2)A ship detection method based on superpixels for SAR images is proposed,which combines superpixels and convolutional neural networks(CNN)for unsupervised ship detection in SAR images.Firstly,the SAR image is segmented into superpixels,and the ship and background superpixels are automatically selected based on the internal statistical properties of the superpixels,achieving the unsupervised property of the method.Then,shape-constrained training samples are generated based on the selected superpixels,and input into the CNN for training.Finally,the trained network is used to classify all superpixels and post-processing is performed to output pixel-level ship detection results.