基于卷積神經(jīng)網(wǎng)絡(luò)人臉識(shí)別方法研究

摘 要：人臉識(shí)別可以靠很多技術(shù)手段來(lái)實(shí)現(xiàn)，而本文則主要探討了通過(guò)深度機(jī)器學(xué)習(xí)卷積神經(jīng)網(wǎng)絡(luò)來(lái)實(shí)現(xiàn)人臉識(shí)別，人臉是圖像識(shí)別中相對(duì)復(fù)雜的識(shí)別對(duì)象，提高識(shí)別精度相對(duì)困難，通過(guò)卷積神經(jīng)網(wǎng)絡(luò)可以有效地提高人臉識(shí)別精度，使其達(dá)到一個(gè)比較滿意的程度。本文重點(diǎn)論述卷積神經(jīng)網(wǎng)絡(luò)進(jìn)行人臉識(shí)別的過(guò)程與方法，介紹了如何通過(guò)改進(jìn)卷積神經(jīng)網(wǎng)絡(luò)來(lái)提高識(shí)別精度。為從事圖像識(shí)別的研究者提供了一些可借鑒的研究思路。

關(guān)鍵詞：深度機(jī)器學(xué)習(xí)；卷積神經(jīng)網(wǎng)絡(luò)；人臉識(shí)別

中圖分類號(hào)：TP391.41；TP183 文獻(xiàn)標(biāo)識(shí)碼：A 文章編號(hào)：2096-4706（2018）10-0102-03

Abstract：Face recognition can be realized by many technical means. In this paper，we mainly discuss the realization of face recognition by deep machine learning convolution neural network. Face is a relatively complex recognition object in image recognition. It is difficult to improve recognition accuracy. The face recognition can be effectively improved by convolution neural network. Accuracy is achieved to a satisfactory degree. This paper focuses on the process and method of face recognition based on convolution neural network，and introduces how to improve the recognition accuracy by improving convolution neural network. It provides some useful research ideas for researchers engaged in image recognition.

Keyword：deep machine leaning；convolution neural network；face recognition

0 引 言

卷積神經(jīng)網(wǎng)絡(luò)進(jìn)行人臉識(shí)別需要經(jīng)過(guò)以下步驟，首先需要準(zhǔn)備訓(xùn)練集圖像、驗(yàn)證集圖像、測(cè)試集圖像；然后通過(guò)程序標(biāo)注圖像標(biāo)簽，形成標(biāo)簽文件；圖像準(zhǔn)備好后，我們需要將圖像格式轉(zhuǎn)化為L(zhǎng)MDB格式，便于提高訓(xùn)練模型的速度；進(jìn)而建立卷積神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)；配置模型訓(xùn)練參數(shù)；開(kāi)始通過(guò)訓(xùn)練集圖像、驗(yàn)證集圖像訓(xùn)練圖像識(shí)別模型；模型訓(xùn)練達(dá)到所希望的精度后，通過(guò)調(diào)用模型進(jìn)行圖像識(shí)別。下面則介紹了如何通過(guò)卷積神經(jīng)網(wǎng)絡(luò)實(shí)現(xiàn)人臉識(shí)別。

1 圖像標(biāo)注

編寫(xiě)訓(xùn)練參數(shù)備置程序，卷積神經(jīng)網(wǎng)絡(luò)采用train_val.prototxt編寫(xiě)的網(wǎng)絡(luò)結(jié)構(gòu)，驗(yàn)證集迭代3次，驗(yàn)證圖像每批次10張，學(xué)習(xí)率為0.0001，學(xué)習(xí)策略為“step”，衰減速率0.5，步長(zhǎng)100步進(jìn)行一次學(xué)習(xí)率衰減。每10次迭代顯示一次識(shí)別精度，模型總共迭代600次，動(dòng)量0.9，權(quán)重0.005，迭代300次快照一次，保存階段性訓(xùn)練模型。……