【摘要】：合成孔徑雷達(dá)(Synthetic aperture radar,SAR)是一種高分辨率的微波成像雷達(dá),具有全天時(shí)、全天候工作,有效地識(shí)別偽裝等優(yōu)勢(shì),已廣泛應(yīng)用于農(nóng)業(yè)、軍事和海洋等領(lǐng)域,有著廣闊的應(yīng)用前景和發(fā)展?jié)摿ΑAR圖像中通常包含有多種地物目標(biāo)的信息,如何有效的對(duì)圖像中各類目標(biāo)進(jìn)行準(zhǔn)確分割,對(duì)SAR圖像的解譯具有重要意義。SAR圖像分割是SAR圖像解譯的重要組成部分,也是當(dāng)前SAR遙感領(lǐng)域研究的熱點(diǎn)與難點(diǎn)。由于SAR的成像機(jī)理決定了SAR圖像中不可避免的引入大量乘性斑點(diǎn)噪聲,基于光學(xué)圖像的分割方法在SAR圖像上很難取得良好結(jié)果。近年來(lái),隨機(jī)場(chǎng)模型理論的發(fā)展,為SAR圖像的分割開辟了一條新的路徑。本文就如何獲得有效和高效的SAR分割結(jié)果做了研究,提出了基于超像素級(jí)條件三場(chǎng)(Superpixel-level conditional triplet Markov field,SL-CTMF)的SAR圖像快速分割方法,主要的工作和貢獻(xiàn)如下:1.條件隨機(jī)場(chǎng)(Conditional random field,CRF)可以直接對(duì)圖像后驗(yàn)進(jìn)行建模,但對(duì)SAR圖像的建模缺少有效的訓(xùn)練數(shù)據(jù)和訓(xùn)練機(jī)制,所以CRF在SAR圖像分割上的應(yīng)用受到限制。三重馬爾可夫隨機(jī)場(chǎng)(Triplet Markov random field,TMF)引入了輔助U場(chǎng)來(lái)有效描述SAR圖像的非平穩(wěn)性,較好抑制了乘性斑點(diǎn)噪聲對(duì)SAR圖像分割所帶來(lái)的影響,取得了良好的分割結(jié)果,但TMF建模復(fù)雜,并且不能充分利用觀測(cè)數(shù)據(jù)的相關(guān)性。2.CRF直接對(duì)后驗(yàn)概率進(jìn)行建模的思想,正好解決了TMF模型存在的缺點(diǎn),并因此產(chǎn)生了像素級(jí)條件三場(chǎng)(Pixel-level conditional triplet Markov field,PL-CTMF)模型。該模型充分結(jié)合了CRF和TMF的優(yōu)勢(shì):直接對(duì)X場(chǎng)的后驗(yàn)概率進(jìn)行建模、并通過(guò)U場(chǎng)的引入來(lái)描述圖像的非平穩(wěn)性,簡(jiǎn)化了SAR圖像的建模方法,提高了SAR圖像的分割效果。3.在有效和高效的SAR分割上,PL-CTMF模型中不管像素的特征與其周圍鄰域點(diǎn)的特征有多么相似,它依然需要計(jì)算每一個(gè)點(diǎn)的分類概率,低效率和高冗余是不可避免的,所以本文就提出了SL-CTMF模型用于SAR圖像的快速分割。首先,針對(duì)SAR圖像,我們對(duì)TurboPixels算法進(jìn)行了改進(jìn),使它能夠獲取一個(gè)邊緣定位準(zhǔn)確的超像素SAR圖像;在超像素級(jí)的SAR圖像上,重新構(gòu)建了輔助場(chǎng)U來(lái)描述SAR圖像的非平穩(wěn)性,SL-CTMF的一元和二元?jiǎng)菽芡ㄟ^(guò)超像素級(jí)的特征和紋理信息得以重建。由于SL-CTMF正是對(duì)超像素進(jìn)行標(biāo)記的,且每個(gè)超像素的特征都是超像素內(nèi)所有像素點(diǎn)特征的綜合特征,所以算法的效率和分割結(jié)果的區(qū)域一致性都會(huì)得到有效提升。最后,結(jié)合最大后驗(yàn)邊緣(Maximum posterior marginal,MPM)方法將SL-CTMF應(yīng)用于無(wú)監(jiān)督SAR圖像的快速分割,在分割效果與PL-CTMF類似或略微好一些的前提下,SL-CTMF極大的縮短了算法的運(yùn)行時(shí)間,達(dá)到了PL-CTMF的1/4到1/6。
[Abstract]:Synthetic Aperture Radar (Synthetic aperture radar,SAR) is a kind of high resolution microwave imaging radar, which has the advantages of all-day, all-weather work, effective identification of camouflage and so on. It has been widely used in agriculture, military and marine fields. SAR images usually contain a variety of ground object information, how to effectively segment all kinds of objects in the image. Sar image segmentation is an important part of SAR image interpretation, and it is also a hot and difficult point in the field of SAR remote sensing. Because the imaging mechanism of SAR determines that a large number of multiplicative speckle noise is inevitably introduced into SAR images, it is difficult for optical image segmentation methods to achieve good results on SAR images. In recent years, the development of random field model theory has opened up a new path for SAR image segmentation. In this paper, how to obtain effective and efficient SAR segmentation results is studied, and a fast SAR image segmentation method based on hyperpixel level conditional three fields (Superpixel-level conditional triplet Markov field,SL-CTMF) is proposed. The main work and contributions are as follows: 1. Conditional random field (Conditional random field,CRF) can directly model the posterior image, but the modeling of SAR image lacks effective training data and training mechanism, so the application of CRF in SAR image segmentation is limited. Triple Markov random field (Triplet Markov random field,TMF) introduces auxiliary U field to effectively describe the nonstationarity of SAR images, which can effectively suppress the influence of multiplicative speckle noise on SAR image segmentation, and good segmentation results are obtained. However, TMF modeling is complex and can not make full use of the correlation of observation data. 2. The idea of TMF modeling posterior probability directly solves the shortcomings of CRF model. Therefore, the pixel-level conditional three-field (Pixel-level conditional triplet Markov field,PL-CTMF) model is generated. The model fully combines the advantages of CRF and TMF: modeling the posterior probability of X field directly, and describing the nonstationarity of the image through the introduction of U field, simplifying the modeling method of SAR image and improving the segmentation effect of SAR image. 3. In the efficient and efficient SAR segmentation, no matter how similar the pixel features are to the adjacent points in the PL-CTMF model, it still needs to calculate the classification probability of each point, and low efficiency and high redundancy are inevitable. Therefore, this paper proposes a SL-CTMF model for fast segmentation of SAR images. Firstly, for SAR images, we improve the TurboPixels algorithm so that it can obtain a super-pixel SAR image with accurate edge location. On the hyperpixel level SAR image, the auxiliary field U is rebuilt to describe the nonstationarity of the SAR image. The univariate and binary potentials of SL-CTMF can be reconstructed by the hyperpixel feature and texture information. Because SL-CTMF marks the super pixel, and the feature of each super pixel is the comprehensive feature of all the pixel features in the super pixel, the efficiency of the algorithm and the regional consistency of the segmentation results will be effectively improved. Finally, combined with the maximum posterior edge (Maximum posterior marginal,MPM) method, the SL-CTMF is applied to the fast segmentation of unsupervised SAR images, and the segmentation effect is similar to or slightly better than that of PL-CTMF. SL-CTMF greatly shortens the running time of the algorithm, reaching 1 鈮，

本文編號(hào)：2493507