Research On SAR Image Denoising And Weakly Supervised Segmentation Based On Deep Learning

Segmentation of synthetic aperture radar(SAR)image is a basic step in SAR image interpretation.The traditional segmentation algorithm contains a lot of manual intervention and high complexity,which cannot meet the requirements of generalization and timeliness.In recent years,deep learning algorithm has developed rapidly in the field of image segmentation because of its powerful feature extraction ability.In the background of SAR big data,deep learning model can rely on massive annotation data to update the weight iteratively,so that the model output can approach the annotation results.It can be said that the generation of deep learning model is the result of the accumulation of expert knowledge and experience,and its inference accuracy is limited by the size and quality of the training dataset as well as the labor-intensive data annotation work.SAR image is disturbed by coherent speckle noise,and it also has imaging characteristics such as relief displacement and foreshortening.The special imaging mechanism makes the sample annotation of SAR image much more complex than that of photo or optical remote sensing image,and it is difficult to ensure the efficiency and accuracy of annotation.Therefore,relying on SAR big data for segmentation task,it needs to spend a lot of time and cost in exchange for accurate fully supervised pixel-level ground object label.In order to solve the above problems,it is necessary to reduce the completeness of fully supervised label,and switch to weakly supervised labeling methods,such as bounding box labeling,global category labeling,etc.,that have loss in the location and contour information of the object,which will inevitably affect the accuracy of feature segmentation.This dissertation not only hopes to use weakly supervised labeling to greatly reduce the cost of manual labeling,but also hopes that the accuracy of the weakly supervised segmentation model can approach the fully supervised segmentation model.For this reason,under the premise of analyzing the characteristics of SAR images,this dissertation carried out research on SAR image denoising and weakly supervised ground object segmentation based on deep learning.Among them,the former provides basic data support for weakly supervised segmentation,so that the weakly supervised segmentation algorithm is not affected by noise in boundary regression.The latter will be specifically developed for two types of man-made objects,buildings and ships,and will obtain accurate object segmentation masks with the aid of superpixel clustering algorithms,probability graph models and polarmask regression.The main research contents and innovations are as follows:1.Aiming at the problem of solidifying the performance of popular denoising methods,a deep learning SAR image denoising method based on texture level map is proposed.On the premise of fully analyzing the spatial coherence of SAR noise and the local texture characteristics of the image,the concept of texture level map is proposed and a novel two-component deep learning denoising network is designed.The network can automatically quantify texture features,and adaptively decide whether to smooth the noise or keep the details of the local area.Using the ultra-fine stripmap mode data of Gaofen-3,the proposed method obtains the equivalent number of looks of 29.23,the uniformity index of the noise image of 0.1183,and the structural index of the noise image of 0.0307.In addition,the denoising experiments of multi-source airborne and spaceborne SAR data show that the proposed method has good generalization performance.Compared with the current popular denoising methods,the proposed method has better performance in subjective visual evaluation and objective index evaluation.At the same time,it provides effective data support for the subsequent weakly supervised segmentation method,and improves the accuracy and smooth continuity of the segmentation boundary.2.Aiming at the time-consuming problem of fully supervised building area labeling,a fully polarized SAR building area weakly supervised extraction method based on superpixel segmentation and Convolutional Neural Network(CNN)is proposed.The method firstly uses an improved simple linear iterative clustering algorithm to adaptively determine the compactness factor to generate superpixels that fit the boundary of the ground object,and then performs multi-scale feature extraction and classification on the representative scene of each superpixel based on CNN.The proposed method converts the basic segmentation unit from pixels to superpixels.Meanwhile,it only relys on image global labels and fully considers polarization decomposition features and pixel spatial context features,which effectively improves the efficiency and accuracy of building area extraction.Experiments are conducted based on the fully polarized stripmap mode data of Gaofen-3,and the extraction result of the proposed method achieves an average overall accuracy of 93.25%,a detection rate of 91.55%,and a false alarm rate of 7.19%.3.Aiming at the problem of difficult ship positioning in complex scenarios,a weakly supervised segmentation algorithm for ships based on class activation maps and conditional random fields is proposed.Inspired by the attention mechanism of computer vision,the feasibility of extracting ship candidate areas based on the ship's global label is verified,and the multi-scale and weak supervision of the candidate areas are deeply analyzed.At the same time,a fully connected conditional random field is introduced to perform boundary regression on the ship candidate area to form a fine segmentation mask.Using the fine strip mode data of Gaofen-3 as the experimental data,the proposed method performs pixel-level segmentation of ship targets under multiple backgrounds,and obtains a ship detection rate of 88.54%,a ship false alarm of 8%,and an F1 score of 90.412.4.Aiming at the problem that densely arranged targets cannot be accurately located,a weakly supervised segmentation method for berthing ships based on pseudo-label and polarmask regression is proposed.First,pseudo-labels of berthing ships are made based on the method proposed in 3,combined with a small amount of fully-supervised labeled data to form a training dataset;then,the ship instance segmentation problem is decomposed into two sub-problems of center regression and polarmask ray regression;finally,the intersection over union loss under polar coordinates and the focal loss are combined to comprehensively train the network.The proposed method is mainly based on pseudo-labels for network training,which effectively solves the problems of timeconsuming and labor-intensive training sample preparation.At the same time,through the ship center regression and mask ray regression mechanism,the segmentation ability of densely arranged ships is significantly improved.Using the ultra-fine strip pattern data of Gaofen-3 to conduct experiments,the proposed method obtains a 90.75% ship detection rate and a 9.24% false alarm,which has the highest accuracy among the comparison methods.At the same time,the control variable ablation experiment of the training set augmentation method shows that after using the proposed training set augmentation method,the network detection rate will be significantly improved,and the false alarm will be significantly reduced.