發(fā)布時間：2018-11-24 11:16

【摘要】：近年來由于移動設(shè)備技術(shù)及互聯(lián)網(wǎng)技術(shù)的不斷發(fā)展,極大的方便人們隨時隨地進行圖片拍攝,這就使得以圖像為出發(fā)點的社交媒體如Flickr、Instagram等開始大量的涌現(xiàn)。如何有效地管理組織這些海量的圖像數(shù)據(jù),并對社會化媒體中圖像進行挖掘分析以促進個體在線交流,提升用戶體驗,輔助企業(yè)做出營銷決策成為了研究的熱點問題。而深度學(xué)習(xí)正是能夠從海量的數(shù)據(jù)中進行學(xué)習(xí)、挖掘的一種機器學(xué)習(xí)方法。其深度分層結(jié)構(gòu)與人類視覺系統(tǒng)具有深度分層的特點一致,所以深度學(xué)習(xí)人類符合人類生物學(xué)上對圖像認知的過程。自2006年深度學(xué)習(xí)被Hinton提出后就引發(fā)了學(xué)術(shù)界、工業(yè)界的研究熱潮,已經(jīng)涌現(xiàn)大量的研究和應(yīng)用。深度學(xué)習(xí)強調(diào)可學(xué)習(xí)性的特點,因此它適合于學(xué)習(xí)具有良好表達力的圖像特征,進而滿足社會化媒體圖像語義學(xué)習(xí)分類、圖像美學(xué)質(zhì)量評價以及以此為基礎(chǔ)的社會化媒體分析研究。本文針對以上問題提出了基于深度學(xué)習(xí)模型的社會化媒體圖像語義分類算法和社會化媒體圖像美學(xué)質(zhì)量評價算法并進行了應(yīng)用。本文首先詳細介紹了深度學(xué)習(xí)算法的基本思想、訓(xùn)練方法以及其與傳統(tǒng)神經(jīng)網(wǎng)絡(luò)的主要異同,并且對幾種得到廣泛研究應(yīng)用的深度學(xué)習(xí)模型作了闡述。其次,描述了深度學(xué)習(xí)在圖像語義分類應(yīng)用的問題定義及現(xiàn)有的方法,并提出了基于棧式去噪編碼器(Stacked denoising Auto-Encoder, SdEA)和基于卷積深度玻爾茲曼機(Convolution Deep Boltzmann Machine, CDBM)的圖像語義分類模型。并且用實驗證明這兩種模型在社會化圖像語義分類問題中的有效性。再次,介紹了深度學(xué)習(xí)在圖像美學(xué)質(zhì)量評價中的應(yīng)用的問題定義,并提出了深度卷積神經(jīng)網(wǎng)絡(luò)+sVM分類器的圖像美學(xué)質(zhì)量評價模型,用實驗驗證了其有效性和可行性。最后,總結(jié)分析了文章的不足之處,為后續(xù)的研究提供方向。
[Abstract]:In recent years, with the continuous development of mobile device technology and Internet technology, it is greatly convenient for people to take pictures at any time and anywhere, which makes social media such as Flickr,Instagram as the starting point to emerge in large numbers. How to effectively manage and organize these massive image data, and how to mine and analyze images in social media to promote individual online communication, improve user experience and assist enterprises to make marketing decisions has become a hot issue. Deep learning is a machine learning method that can learn from massive data. The structure of depth stratification is consistent with that of human visual system, so the deep learning of human is in accordance with the process of image cognition in human biology. Since the deep learning was put forward by Hinton in 2006, it has triggered a research boom in academia and industry, and a large number of research and applications have emerged. Deep learning emphasizes the characteristics of learnability, so it is suitable for learning image features with good expressiveness, thus satisfying the classification of image semantic learning in social media. Image aesthetic quality evaluation and social media analysis based on it. In order to solve the above problems, this paper proposes a social media image semantic classification algorithm based on the in-depth learning model and an algorithm for evaluating the aesthetic quality of social media images. In this paper, the basic idea of depth learning algorithm, training method and its main similarities and differences with traditional neural network are introduced in detail, and several kinds of depth learning models which have been widely studied and applied are described. Secondly, the problem definition and existing methods of depth learning in image semantic classification are described, and a stack denoising encoder (Stacked denoising Auto-Encoder, SdEA) and convolution depth Boltzmann machine (Convolution Deep Boltzmann Machine, are proposed. CDBM) image semantic classification model. Experiments show that the two models are effective in the problem of socialized image semantic classification. Thirdly, the problem definition of the application of depth learning in image aesthetic quality evaluation is introduced, and the evaluation model of image aesthetic quality of sVM classifier based on deep convolution neural network is proposed. The validity and feasibility of the model are verified by experiments. Finally, the paper summarizes and analyzes the shortcomings of the article, and provides the direction for further research.
【學(xué)位授予單位】：華北電力大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2015
【分類號】：TP391.41

【參考文獻】

相關(guān)期刊論文 前1條

1 高雋;謝昭;張駿;吳克偉;;圖像語義分析與理解綜述[J];模式識別與人工智能;2010年02期

，

本文編號：2353487