【摘要】：側(cè)掃聲吶是海洋活動中常用的探測裝置,它通過成像的方式展示信息。側(cè)掃聲吶圖像具有分辨率低、噪聲大、灰度分布范圍窄的特點(diǎn),也因此給側(cè)掃聲吶圖像的目標(biāo)分割帶來很大困難。在分析當(dāng)前主流的圖像分割技術(shù)與發(fā)展趨勢后,本文對通用性強(qiáng)的聚類算法和基于非監(jiān)督特征學(xué)習(xí)的方法做了研究。首先,針對側(cè)掃聲吶噪聲強(qiáng)的特點(diǎn),分析了聲吶噪聲產(chǎn)生的原因,聲吶噪聲的分類并對噪聲進(jìn)行建模。在此基礎(chǔ)上分析了聲吶去噪的常用方法,將應(yīng)用在光學(xué)圖像上的若干效果突出的算法在聲吶圖像上進(jìn)行了嘗試。對具有不同場景特點(diǎn)的聲吶圖像做了去噪實(shí)驗(yàn)。實(shí)驗(yàn)結(jié)果表明,上述方法對于聲吶圖像也是有效的。在此基礎(chǔ)上,進(jìn)一步分析了算法之間去噪效果差異和不同噪聲去噪效果之間的優(yōu)劣原因,并描述了這些去噪算法在聲吶去噪應(yīng)用上的一些可能的改進(jìn)措施。接著,結(jié)合側(cè)掃聲吶圖像的特點(diǎn),遴選紋理特征中的兩種目前應(yīng)用廣泛的重要特征,局部二值模式和類哈爾特征,對它們的原理做了詳細(xì)的描述,并利用目前在諸多領(lǐng)域取得了突破性進(jìn)展的深度學(xué)習(xí)算法中的一種,稀疏自編碼器,來對側(cè)掃聲吶圖像進(jìn)行特征學(xué)習(xí),成功地構(gòu)建了專門針對側(cè)掃聲吶圖像的特征。通過對比分析認(rèn)為使用特征學(xué)習(xí)得到的特征對基于聚類的側(cè)掃聲吶圖像分割具有明顯的優(yōu)勢。然后,詳細(xì)地介紹了 K均值聚類、層次聚類、模糊聚類和譜聚類這四種能夠在側(cè)掃聲吶圖像分割中應(yīng)用的常用的基礎(chǔ)聚類算法,討論了它們的原理,并使用本文通過非監(jiān)督學(xué)習(xí)得到的特征,用圖像的灰度信息作為對照組,在側(cè)掃聲吶圖像的樣本中進(jìn)行了試驗(yàn),并對結(jié)果做了比較分析。最后,在基于前述K均值算法取得的良好聚類分割效果的基礎(chǔ)上,針對數(shù)據(jù)量大算法運(yùn)行耗時的問題,使用基于OpenMP和CUDA的兩種并行化算法對K均值聚類算法進(jìn)行加速,并通過實(shí)驗(yàn)對比說明側(cè)掃聲吶圖像的聚類分割速度得到提高。
[Abstract]:Side scan sonar is a common detection device in ocean activities. It displays information by imaging. Side-scan sonar image has the characteristics of low resolution, large noise and narrow gray-scale distribution. Therefore, it is very difficult to segment the target of side-scan sonar image. After analyzing the current mainstream image segmentation technology and the development trend, this paper studies the universal clustering algorithm and the unsupervised feature learning method. Firstly, according to the strong noise characteristics of side-scan sonar, the causes of sonar noise are analyzed, and the classification of sonar noise and the modeling of sonar noise are carried out. On this basis, the common methods of sonar de-noising are analyzed, and some prominent algorithms applied to optical images are tried on sonar images. The denoising experiments of sonar images with different scene characteristics are carried out. The experimental results show that the proposed method is also effective for sonar images. On this basis, the difference of denoising effect among algorithms and the advantages and disadvantages of different noise denoising effects are further analyzed, and some possible improvement measures of these denoising algorithms in sonar de-noising applications are described. Then, combining the characteristics of side-scan sonar images, two important features, local binary pattern and quasi-Hal feature, which are widely used in texture features, are selected, and their principles are described in detail. One of the deep learning algorithms, sparse self-encoder, which has made a breakthrough in many fields at present, is used to study the features of side-scan sonar images, and the features of side-scan sonar images are constructed successfully. Through comparative analysis, it is concluded that the features obtained from feature learning have obvious advantages in the segmentation of side-scan sonar images based on clustering. Then, K-means clustering, hierarchical clustering, fuzzy clustering and spectral clustering are introduced in detail, which can be used in the segmentation of side-scan sonar images, and their principles are discussed. Using the features obtained by unsupervised learning and using the gray-scale information of the image as the control group, the experiments were carried out in the samples of side-scan sonar images, and the results were compared and analyzed. Finally, two parallel algorithms based on OpenMP and CUDA are used to accelerate the K-means clustering algorithm, which is based on the good segmentation results obtained by the K-means algorithm and the time-consuming running of the algorithm with large amount of data, which is based on the above-mentioned K-means algorithm. The experimental results show that the clustering segmentation speed of side scan sonar images is improved.
【學(xué)位授予單位】：哈爾濱工程大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2016
【分類號】：TP391.41

【參考文獻(xiàn)】

相關(guān)期刊論文 前4條

1 馬文萍;黃媛媛;李豪;李曉婷;焦李成;;基于粗糙集與差分免疫模糊聚類算法的圖像分割[J];軟件學(xué)報;2014年11期

2 范習(xí)健;李慶武;黃河;王敏;;側(cè)掃聲吶圖像的3維塊匹配降斑方法[J];中國圖象圖形學(xué)報;2012年01期

3 馬秀麗;焦李成;;基于分水嶺-譜聚類的SAR圖像分割[J];紅外與毫米波學(xué)報;2008年06期

4 焦李成,杜海峰;人工免疫系統(tǒng)進(jìn)展與展望[J];電子學(xué)報;2003年10期

相關(guān)博士學(xué)位論文 前1條

1 張小峰;基于模糊聚類算法的醫(yī)學(xué)圖像分割技術(shù)研究[D];山東大學(xué);2014年

相關(guān)碩士學(xué)位論文 前1條

1 耿長磊;基于多進(jìn)制小波變換的聲吶圖像去噪[D];大連理工大學(xué);2014年

，

本文編號：2448899