【摘要】：SAR圖像的應用需求與日俱增,SAR圖像目標識別技術(shù)也在發(fā)展。由于硬件性能提升和有效訓練算法的提出,近年深度學習重獲關(guān)注,并在圖像識別領(lǐng)域取得成功。本文應用深度學習的理論和方法,結(jié)合SAR圖像的特點,研究了基于深度學習的SAR特征提取與目標識別方法。主要研究內(nèi)容如下:根據(jù)SAR圖像的特性指出了SAR圖像目標識別中的難點。SAR圖像目標具有多變不確定的特點,傳統(tǒng)識別方法需要大量的專業(yè)知識,需要對圖像預處理,不能自動提取有效的特征。深度學習具有盲學習和無監(jiān)督學習的能力,本文使用深度學習解決該問題,將普通神經(jīng)網(wǎng)絡(luò)、深度置信網(wǎng)絡(luò)和卷積神經(jīng)網(wǎng)絡(luò)三種深度結(jié)構(gòu)分別用于三類和十類的SAR目標識別。通過對比發(fā)現(xiàn)在帶標簽樣本足夠的情況下,深度置信網(wǎng)絡(luò)的預訓練對普通神經(jīng)網(wǎng)絡(luò)提升不大,二者的識別性能幾乎相同。深度學習對參數(shù)和結(jié)構(gòu)十分敏感,實驗發(fā)現(xiàn)卷積神經(jīng)網(wǎng)絡(luò)在不同激活函數(shù)下對SAR目標識別結(jié)果差異巨大,其中ReLu函數(shù)最適合作為卷積神經(jīng)網(wǎng)絡(luò)的激活函數(shù)。接下來分析了卷積神經(jīng)網(wǎng)絡(luò)中間結(jié)構(gòu)對SAR目標識別的影響。池化層分別選擇mean-pooling和max-pooling,對比識別結(jié)果,并利用池化后的特征重構(gòu)圖像,對比與原圖像的相似度,結(jié)果表明mean-pooling更適合作為SAR目標識別時的池化層特征選擇方法。改變卷積核大小發(fā)現(xiàn)最適合的卷積核大小和目標圖像尺寸是相關(guān)的。本文還考慮了SAR圖像目標在有遮擋情況下的識別問題。卷積神經(jīng)網(wǎng)絡(luò)在目標區(qū)域50%遮擋率的情況下識別率有所下降。Dropout方法的思想在于每次訓練時隨機丟棄部分神經(jīng)元,這樣訓練出來的模型具備了只使用部分信息進行推斷預測的能力。實驗結(jié)果證明了在SAR目標遮擋的情況下使用了Dropout的神經(jīng)網(wǎng)絡(luò)識別率有所提高。三種深度模型的對比顯示卷積神經(jīng)網(wǎng)絡(luò)對SAR目標特征提取的可分性好于其他二者,對三類和十類目標的識別率分別達到了99.8%和96.3%,明顯高于前兩種模型。對卷積神經(jīng)網(wǎng)絡(luò)提取的特征可視化,發(fā)現(xiàn)卷積神經(jīng)網(wǎng)絡(luò)在特征提取上能很好地抓住SAR圖像的局部相關(guān)性,保持圖像的空間結(jié)構(gòu)。
[Abstract]:The application demand of SAR image is increasing day by day, and the target recognition technology of SAR image is also developing. Due to the improvement of hardware performance and the development of effective training algorithm, the depth learning has been paid more attention in recent years, and it has been successful in the field of image recognition. Based on the theory and method of depth learning and the characteristics of SAR image, this paper studies the method of SAR feature extraction and target recognition based on depth learning. The main research contents are as follows: according to the characteristics of SAR image, the difficulties in SAR image target recognition are pointed out. SAR image target has the characteristics of changeable uncertainty, traditional recognition methods need a lot of professional knowledge, and need to preprocess the image. Can not automatically extract valid features. Depth learning has the ability of blind learning and unsupervised learning. In this paper, we use deep learning to solve this problem, and apply three depth structures of general neural network, depth confidence network and convolutional neural network to SAR target recognition of three and ten categories, respectively. It is found by comparison that the pre-training of the depth confidence neural network has little effect on the ordinary neural networks and their recognition performance is almost the same when the labeled samples are sufficient. Deep learning is very sensitive to parameters and structures. It is found that the results of SAR target recognition by convolutional neural networks vary greatly under different activation functions, and ReLu function is the most suitable for the activation functions of convolutional neural networks. Then the effect of convolution neural network intermediate structure on SAR target recognition is analyzed. The results of mean-pooling and max-pooling, contrast recognition are selected in the pool layer, and the reconstructed images are reconstructed by using the pooled features. The results show that mean-pooling is more suitable to be used as the feature selection method for SAR target recognition. Changing the size of the convolutional kernel, the most suitable size of the convolution kernel is found to be related to the size of the target image. The problem of target recognition in SAR images with occlusion is also considered in this paper. The recognition rate of convolutional neural network decreases with 50% occlusion rate in the target region. The idea of Dropout method is to discard some neurons at random during each training. This trained model has the ability to infer and predict only a portion of the information. The experimental results show that the recognition rate of neural network using Dropout is improved when SAR targets are occluded. The comparison of the three depth models shows that the convolution neural network has better separability than the other two models, and the recognition rates of the three and ten kinds of targets are 99.8% and 96.3%, respectively, which are significantly higher than those of the former two models. By visualizing the feature extracted by convolution neural network, it is found that convolution neural network can grasp the local correlation of SAR image and keep the spatial structure of the image.
【學位授予單位】：電子科技大學
【學位級別】：碩士
【學位授予年份】：2017
【分類號】：TN957.52

【參考文獻】

相關(guān)期刊論文 前5條

1 張新征;黃培康;;基于貝葉斯壓縮感知的SAR目標識別[J];系統(tǒng)工程與電子技術(shù);2013年01期

2 王世f^;賀志國;;基于PCA特征的快速SAR圖像目標識別方法[J];國防科技大學學報;2008年03期

3 黃世奇,王善成;微波遙感SAR軍事探測技術(shù)研究[J];飛航導彈;2005年04期

4 譚衢霖,邵蕓;成像雷達(SAR)遙感地質(zhì)應用綜述[J];地質(zhì)找礦論叢;2003年01期

5 吳中鼎,李占橋;星載SAR遙感在海洋水文觀測中的應用探討[J];海洋測繪;2001年02期

，

本文編號：2383136