Multi-scaled FCN-CRF And Reinforcement Learning For High Resolution SAR Image Semantic Segmentation

Synthetic aperture radar(SAR)has become an important tool for remote sensing based earth observation because of the ability to acquire images under all-time and all-weather conditions.With the development of SAR imaging techniques and improvement of the spatial resolution,segmentation and classification methods that are applicable for low and middle resolution SAR images are not able to cope with HR SAR images due to aforementioned characteristics.In view of the above problems,after surveying and analyzing existing methods,in this dissertation several segmentation and classification methods are proposed for HR SAR images.The main work and contributions are summarized as follows:1.Based on the FCN-CRF,the combination model would be used to deal with the task of HR SAR image segmentation and classification.Fully convolutional network(FCN)employs a deconvolution layer to up-sample the feature map of the last convolution layer,yielding a prediction for each pixel.Conditional random fields(CRF)is a graph model based on probability which could fuse and make use of various types of features and context information of SAR image,and split each step of CRF deductive reasoning and solving process into recurrent neural network(RNN)iterative process to a CRF-RNN structure.Finally,the FCN and CRF would be united into a module to form an end-to-end network.Experiments have proved that this method selects 5% of sampled data for training and completes the task of semantic segmentation of SAR images based on high-level semantic information.The accuracy of segmentation is higher than that of other comparison algorithms.2.Considering that HR SAR images contain rich multi-resolution,multi-directional and multi-scale information,a novel model of FCN-CRF which incorporates multiscale transform is raised.The random initializing filters in the convolution layer of FCN is replaced by the scale filter banks in nonsubsampled contourlet transform,and a multi-scale FCN network is obtained.Then as a feature extraction layer,the low-frequency and highfrequency coefficients would be extracted to the deep learning model.Experimental results reveal that the proposed approach can effectively utilize multi-scaled information contained in high-dimensional SAR image features,leading to enhanced segmentation accuracy for HR SAR images with complex scenes.3.Deep reinforcement learning can directly acquire images and receive feedback to make certain decisions.We use the FCN-CRF model as a semantic segmentation module for small areas,and deep reinforcement learning as a decision module in a larger area.Through the decision-making ability of deep reinforcement learning,the entire SAR image would be segmented and divided gradually with appropriate parameters.Finally,the neighborhood relationship between blocks that can be increased to some extent,so as to improve the classification accuracy.Meanwhile,the difficulty of classification caused by the noise and clutters would be tackled.