發(fā)布時間：2018-03-04 10:35

  本文選題：人臉檢測　切入點：候選區(qū)域-快速卷積神經(jīng)網(wǎng)絡(luò)　出處：《中國科學(xué)技術(shù)大學(xué)》2017年碩士論文　論文類型：學(xué)位論文

【摘要】：隨著媒體和社交網(wǎng)絡(luò)的發(fā)展,人臉年齡與性別識別在現(xiàn)實生活中的應(yīng)用越來越多,吸引了廣泛的研究興趣。由于人臉圖像的生物特征識別是非接觸的,比較簡單快速,還具有一定的娛樂性,在社交網(wǎng)絡(luò)、視頻監(jiān)控、人機交互等領(lǐng)域具有廣闊的應(yīng)用前景。本文主要研究了人臉檢測方法,以及人臉年齡與性別識別方法,并分別提出兩種解決方案,以適應(yīng)不同的應(yīng)用場景。第一種方案,使用Faster R-CNN算法進行人臉檢測,提取人臉的CNN特征進行訓(xùn)練和測試。第二種方案,使用基于比例特征和Adaboost算法進行人臉檢測,提取圖像的LBP特征作為人臉特征。上述兩種方案提取特征之后,均使用隨機森林進行訓(xùn)練和測試,具體內(nèi)容如下:(1)第一種方案,由于Faster R-CNN算法在各個目標檢測數(shù)據(jù)集上取得驚人的成績,因此本文在WIDER大規(guī)模人臉數(shù)據(jù)集上,訓(xùn)練一個Faster R-CNN模型進行人臉檢測,并在FDDB數(shù)據(jù)庫上對該模型進行評估,結(jié)果表明該算法有較高的人臉檢測率。為了提高在非限制性環(huán)境下對人臉年齡與性別的識別準確率,本文提出一種基于深度卷積神經(jīng)網(wǎng)絡(luò)的人臉特征提取方法,使用"一般到特殊"的微調(diào)方案。首先采用在大規(guī)模數(shù)據(jù)集上進行人臉識別預(yù)訓(xùn)練得到的VGG-Face模型;接著使用該模型在CelebA人臉屬性數(shù)據(jù)集上,對選取的5個特定的屬性進行微調(diào)訓(xùn)練,得到人臉屬性模型,這幾個屬性分別是:①是否留胡子,②是否年輕,③是否戴眼鏡,④性別是否為男,⑤是否戴帽子。將所有全連接層的輸出值連接起來,構(gòu)成一個向量,作為人臉特征;最后使用隨機森林分類器,在Adience數(shù)據(jù)集上訓(xùn)練和測試。實驗結(jié)果表明,該方法的分類準確率較高,提取的人臉CNN特征具有魯棒性。(2)第二種方案,提出基于比例特征和Adaboost的人臉檢測算法,然后提取圖像LBP直方圖作為人臉特征向量。具體的,本文提出的比例特征,描述的是圖像中任意兩個點的比例關(guān)系,它具有尺度不變性,有界性等特點。本文使用深度二次樹去學(xué)習(xí)比例特征及其組合的最優(yōu)子集,使得人臉不同部位可以通過學(xué)習(xí)的規(guī)則被分割,再使用一個soft-cascade級聯(lián)結(jié)構(gòu)的分類器對滑動窗口進行分類,檢測人臉位置。接著,本文使用圖像分塊的方法,分別提取各級人臉圖像的LBP直方圖特征,并使用隨機森林算法進行訓(xùn)練和測試。該方法的實驗結(jié)果跟上述基于人臉CNN特征的分類方法相比,準確率要低一些。
[Abstract]:With the development of media and social network, face age and gender recognition is more and more widely used in real life, attracting wide research interest. Because the biometric recognition of face image is non-contact, it is relatively simple and fast. It has a wide application prospect in social network, video surveillance, human-computer interaction and so on. This paper mainly studies the face detection method, as well as the face age and gender recognition method. Two solutions are proposed to adapt to different application scenarios. First, face detection using Faster R-CNN algorithm, CNN feature extraction for training and testing. Face detection is based on scale feature and Adaboost algorithm, and LBP feature of image is extracted as face feature. After these two schemes are extracted, they are trained and tested by random forest. The concrete contents are as follows: 1) the first scheme. Because the Faster R-CNN algorithm has achieved remarkable results in each target detection data set, this paper trains a Faster R-CNN model on the WIDER large-scale face data set to detect the face, and evaluates the model on the FDDB database. The results show that the algorithm has high face detection rate. In order to improve the accuracy of face age and gender recognition in the unrestricted environment, this paper proposes a face feature extraction method based on deep convolution neural network. Using the "general-to-special" fine-tuning scheme. Firstly, the VGG-Face model, which is pre-trained on large-scale data sets for face recognition, is used, and then the model is used on the CelebA face attribute data set. The five selected attributes are trained to fine tune, and the face attribute model is obtained. These attributes are: 1, whether he has a beard, whether he is young, whether he is wearing glasses, whether he is a man, and whether he wears a hat. All the output values of the full connection layer are connected together to form a vector as a feature of a face; Finally, a random forest classifier is used to train and test the Adience dataset. The experimental results show that the classification accuracy of this method is high, and the extracted face CNN features are robust. A face detection algorithm based on scale feature and Adaboost is proposed, and then the LBP histogram of the image is extracted as the face feature vector. It has the characteristics of scale invariance and boundedness. In this paper, we use the deep quadratic tree to learn the optimal subset of the scale feature and its combination, so that different parts of the face can be segmented by learning rules. Then a classifier with soft-cascade cascade structure is used to classify the sliding window to detect the position of the face. Then, the LBP histogram features of the face images at all levels are extracted by using the method of image segmentation. The experimental results of this method are lower than the classification method based on face CNN feature.
【學(xué)位授予單位】：中國科學(xué)技術(shù)大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2017
【分類號】：TP391.41

【參考文獻】

相關(guān)期刊論文 前2條

1 劉江華,陳佳品,程君實;基于Gabor小波特征抽取和支持向量機的人臉識別[J];計算機工程與應(yīng)用;2003年08期

2 劉穎超,張紀元;梯度下降法[J];南京理工大學(xué)學(xué)報(自然科學(xué)版);1993年02期

，

本文編號：1565336