發(fā)布時間：2018-10-21 11:58

【摘要】：近些年來隨著社交網(wǎng)絡(luò)的快速發(fā)展以及普及，人們將越來越多的時間放到了社交網(wǎng)絡(luò)上，這使得社交網(wǎng)絡(luò)成為最有潛力的廣告以及商業(yè)平臺。品牌跟蹤是近些年出現(xiàn)的一種服務(wù)，通過分析品牌在媒體上曝光的頻繁程度以及用戶的評價來評估品牌的成長。由于社交網(wǎng)絡(luò)的發(fā)展，品牌跟蹤逐漸將重心轉(zhuǎn)移到社交網(wǎng)絡(luò)上。 對于品牌跟蹤，在當前社交網(wǎng)絡(luò)平臺上僅僅提供通用的關(guān)鍵詞搜索功能，這帶來兩個弊端：第一搜索結(jié)果中含有大量噪聲，盡管含有關(guān)鍵詞，，但經(jīng)常與該品牌并不相關(guān)；第二大量含有品牌圖片的信息無法被檢索到。為了解決問題二，本文提出一種新的LOGO(商標)檢測方法，將社交網(wǎng)絡(luò)中用戶上傳的包含品牌LOGO的圖像檢測出來。這既可以作為一個獨立的應(yīng)用，直接作為品牌跟蹤功能；也可以作為一個品牌分析系統(tǒng)的一部分。 社交網(wǎng)絡(luò)上的圖像有較大的比例為用戶自己拍攝上傳，圖像質(zhì)量往往較低，包括光線條件差、圖像模糊、拍攝角度差，這使得圖像中的LOGO發(fā)生光照不均勻、傾斜旋轉(zhuǎn)、彈性變形、部分被遮擋等問題。此外，為了增加辨識度，LOGO往往被設(shè)計成簡單的圖形，這使得其與自然圖像中的物體外形相似。這些都增加了LOGO檢測的難度。為了解決社交網(wǎng)絡(luò)圖像的LOGO檢測問題，本文研究一種基于機器學(xué)習的LOGO檢測方法并評估其在社交網(wǎng)絡(luò)上的應(yīng)用。 本文主要貢獻如下，一方面，本文建立了一個包含100個品牌LOGO的圖像訓(xùn)練集以及測試集。其中訓(xùn)練集給出LOGO的位置、大小以及其旋轉(zhuǎn)角度。測試圖像包括100萬張圖像，每張圖像已經(jīng)標注好是否含有LOGO，以及LOGO的位置和大小。訓(xùn)練集中每個LOGO的樣本數(shù)量平均超過300張。該數(shù)據(jù)集涵蓋了LOGO在不同光照、面內(nèi)旋轉(zhuǎn)、模糊、拍攝角度的情況，對后續(xù)科研人員進行使用并測試具有很大的價值。另一方面，本文使用了一種新的LOGO檢測算法。由于本課題采用機器學(xué)習的方法進行LOGO檢測，這是一個正負樣本嚴重不均衡的問題。而訓(xùn)練的過程中指定正負樣本比例，因此本課題提出將每一級AdaBoost的節(jié)點選擇出來的特征作為輸入，得到一個線性分類器，克服正負樣本不均衡的情況。最后本文給出一種基于LOGO檢測算法新的品牌跟蹤的方法，通過判斷社交網(wǎng)絡(luò)圖像中是否含有LOGO來給出品牌的關(guān)注程度，給出階段性的品牌關(guān)注度分析，從而補充了現(xiàn)有基于文本關(guān)鍵詞的缺陷。
[Abstract]:With the rapid development and popularity of social networks in recent years, people will spend more and more time on social networks, which makes social networks the most potential advertising and business platform. Brand tracking is a service emerging in recent years. It evaluates brand growth by analyzing the frequency of brand exposure in the media and the evaluation of users. Due to the development of social network, brand tracking is gradually shifting its focus to social network. For brand tracking, only general keyword search function is provided on the current social network platform, which brings two disadvantages: the first search result contains a lot of noise, although it contains keywords, it is often not related to the brand; The second large amount of information containing brand images cannot be retrieved. In order to solve the second problem, this paper proposes a new LOGO (trademark) detection method, which detects the images uploaded by users in social networks including brand LOGO. This can be used as an independent application, directly as a brand tracking function, or as part of a brand analysis system. There is a large proportion of images taken and uploaded by users on social networks, and the quality of the images is often low, including poor light conditions, blurred images, poor shooting angles, which results in uneven illumination and skewed rotation of the LOGO in the images. Elastic deformation, partial occlusion and so on. In addition, in order to increase the degree of identification, LOGO is often designed as a simple figure, which makes it similar to the shape of objects in natural images. All of these increase the difficulty of LOGO detection. In order to solve the problem of LOGO detection of social network images, this paper studies a LOGO detection method based on machine learning and evaluates its application in social network. The main contributions of this paper are as follows: on the one hand, we build an image training set including 100 brand LOGO and test set. The training set gives the position, size and rotation angle of LOGO. The test image consists of 1 million images, each of which has been marked with the location and size of LOGO, and LOGO. The average number of samples per LOGO in the training set is more than 300. This data set covers the situation of LOGO in different illumination, in-plane rotation, blur and shooting angle, so it is of great value for the subsequent researchers to use and test. On the other hand, this paper uses a new LOGO detection algorithm. In this paper, machine learning is used to detect LOGO, which is a serious imbalance of positive and negative samples. The proportion of positive and negative samples is specified in the process of training, so this paper presents a linear classifier to overcome the imbalance of positive and negative samples by using the selected features of each level of AdaBoost as input. Finally, this paper presents a new brand tracking method based on LOGO detection algorithm. By judging whether there is LOGO in the image of social network, this paper gives the attention degree of the brand, and gives the stage analysis of brand concern. Thus, it complements the existing defects based on text keywords.
【學(xué)位授予單位】：哈爾濱工業(yè)大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2013
【分類號】：TP391.41

【共引文獻】

相關(guān)期刊論文 前10條

1 郭平;劉波;沈岳;;農(nóng)業(yè)云大數(shù)據(jù)自組織推送關(guān)鍵技術(shù)綜述[J];軟件;2013年03期

2 劉潤龍;;云計算及關(guān)鍵技術(shù)研究[J];數(shù)字化用戶;2013年06期

3 容春琳;;公共圖書館應(yīng)用大數(shù)據(jù)的策略研究[J];圖書館建設(shè);2013年07期

4 李伶;;網(wǎng)絡(luò)環(huán)境下高校圖書館信息服務(wù)模式探析[J];情報理論與實踐;2013年08期

5 莫榮強;艾萍;吳禮福;岳兆新;馮鵬;;一種支持大數(shù)據(jù)的水利數(shù)據(jù)中心基礎(chǔ)框架[J];水利信息化;2013年03期

6 馮鈞;許瀟;唐志賢;徐黎明;;水利大數(shù)據(jù)及其資源化關(guān)鍵技術(shù)研究[J];水利信息化;2013年04期

7 陳建昌;;大數(shù)據(jù)環(huán)境下的網(wǎng)絡(luò)安全分析[J];中國新通信;2013年17期

8 陸t

本文編號：2285026