發(fā)布時(shí)間：2018-09-01 17:54

【摘要】：人臉識(shí)別一直是機(jī)器學(xué)習(xí)領(lǐng)域的熱門問(wèn)題,針對(duì)不同場(chǎng)景和不同目標(biāo),人們提出了各種解決算法。隨著壓縮感知理論的發(fā)展和成熟,其在人臉識(shí)別領(lǐng)域的一項(xiàng)應(yīng)用就是稀疏表示識(shí)別算法,該算法具有對(duì)特征提取不敏感和對(duì)遮擋物魯棒性很好的優(yōu)點(diǎn)。本論文的核心是一種協(xié)同表示識(shí)別算法,其改進(jìn)于稀疏表示識(shí)別算法,繼承了對(duì)遮擋物較好的魯棒性,并大幅提升了算法的速度。針對(duì)協(xié)同表示識(shí)別算法的準(zhǔn)確度和實(shí)用性,本論文對(duì)其進(jìn)行了三個(gè)方面的改進(jìn)工作:(1)使用多元特征集作為算法模型的輸入,并行地訓(xùn)練多個(gè)基于不同特征的模型,然后加權(quán)求和同一分類下不同特征模型的殘差,以此作為識(shí)別依據(jù)。這種改進(jìn)方式可以綜合利用不同特征從不同角度提取的有效信息,從而提高算法的準(zhǔn)確度。(2)給出了一種加權(quán)相對(duì)距離的指標(biāo)作為對(duì)Outlier情形的判決依據(jù)。協(xié)同表示識(shí)別算法得到的編碼稀疏性變?nèi)?因而稀疏表示識(shí)別算法中的稀疏集中因子不再適合本算法。加權(quán)相對(duì)距離指標(biāo)繞過(guò)了對(duì)編碼稀疏性的依賴,綜合考慮了最優(yōu)解和次優(yōu)解之間的距離和相似度進(jìn)行判決,更加適合當(dāng)前場(chǎng)景。(3)針對(duì)實(shí)際應(yīng)用中常見(jiàn)的樣本不足的難題,本論文給出了一種基于變換字典的解決方案。通過(guò)從標(biāo)準(zhǔn)人臉庫(kù)中提取不同光照、不同姿態(tài)、各種遮擋物等情形下的變換基,生成一個(gè)變換字典,擴(kuò)展和補(bǔ)充當(dāng)前訓(xùn)練集的不完備字典,從而能夠使用少量訓(xùn)練樣本就能表示不同場(chǎng)景下的各類人臉。
[Abstract]:Face recognition has always been a hot problem in the field of machine learning. With the development and maturity of compression perception theory, one of its applications in the field of face recognition is sparse representation recognition algorithm, which is insensitive to feature extraction and robust to occlusion. The core of this paper is a collaborative representation recognition algorithm, which is improved in sparse representation recognition algorithm, inheriting better robustness to occlusion, and greatly improving the speed of the algorithm. Aiming at the accuracy and practicability of cooperative representation recognition algorithm, this paper improves it in three aspects: (1) using multivariate feature set as the input of algorithm model, training several models based on different features in parallel. Then weighted sum the residuals of different feature models under the same classification, which is used as the basis of recognition. The improved method can make use of the effective information extracted from different points of view by using different features to improve the accuracy of the algorithm. (2) A weighted relative distance index is given as the basis for judging the Outlier case. The coding sparsity obtained by the cooperative representation recognition algorithm is weak, so the sparse set factor in the sparse representation recognition algorithm is no longer suitable for this algorithm. The weighted relative distance index bypasses the dependence on coding sparsity and synthetically considers the distance and similarity between the optimal solution and the sub-optimal solution, which is more suitable for the current situation. (3) aiming at the problem of lack of samples, which is common in practical application, the weighted relative distance index is more suitable for the current situation. This paper presents a solution based on the transformation dictionary. By extracting the transform bases from the standard face database under different illumination, different posture and various occlusion objects, a transformation dictionary is generated to extend and supplement the incomplete dictionary of the current training set. Thus, a small number of training samples can be used to represent different kinds of faces in different scenes.
【學(xué)位授予單位】：北京郵電大學(xué)
【學(xué)位級(jí)別】：碩士
【學(xué)位授予年份】：2016
【分類號(hào)】：TP391.41

【參考文獻(xiàn)】

相關(guān)期刊論文 前1條

1 陳晶;黃曙光;;分布式并行矩陣乘算法分析[J];兵工自動(dòng)化;2005年05期

，

本文編號(hào)：2217927