謝鑫
摘? 要:該文采用卷積神經(jīng)網(wǎng)絡(luò)(Convolutional Neural Networks, CNN)進(jìn)行人臉識(shí)別模型開發(fā),現(xiàn)有深度卷積神經(jīng)網(wǎng)絡(luò)的模型,提出了改進(jìn)的Faster R-CNN目標(biāo)檢測模型與ResNet50結(jié)合的方法進(jìn)行識(shí)別。ResNet50網(wǎng)絡(luò)大小適中,對算力要求不高,故可以采用訓(xùn)練耗時(shí)較長的杜鵑算法(Cuckoo Search,CS)對原有的網(wǎng)絡(luò)訓(xùn)練方法進(jìn)行一定變更,在CelebA上進(jìn)行訓(xùn)練與預(yù)測。與未改進(jìn)的ResNet50相比,其訓(xùn)練時(shí)間有所增加但預(yù)測準(zhǔn)確率上升1.7%。
關(guān)鍵詞:ResNet50? Faster r?cnn? Cuckoo Search? 人臉識(shí)別
中圖分類號(hào):TP391.41? ? ? ? ? ? ? ? ? ? ? ?文獻(xiàn)標(biāo)識(shí)碼:A? ? ? ? ? ? ? ? ? ?文章編號(hào):1672-3791(2020)12(b)-0009-03
Abstract: This paper uses Convolutional Neural Networks (CNN) for face recognition model development. For existing deep convolutional neural network models, an improved Faster R-CNN target detection model combined with ResNet50 is proposed for recognition. The ResNet50 network is of moderate size and does not require high computing power. Therefore, the Cuckoo Search (CS) algorithm, which takes a long time to train, can be used to make certain changes to the original network training method, and train and predict on CelebA. Compared with the unimproved ResNet50, its training time has increased but the prediction accuracy has increased by 1.7%.
Key Words: ResNet50; Faster r?cnn; Cuckoo Search; Face recognition
Yann LeCun于1989年將CNN用于手寫數(shù)字識(shí)別, 近年來CNN也被廣泛用于人臉識(shí)別領(lǐng)域[1]。而在人臉識(shí)別領(lǐng)域除CNN外目前還有SVM、小波變換、圖像二值化等檢測方法[2]。
1? 原有網(wǎng)絡(luò)與初始化方法
HE Kaiming等人在2015年提出ResNet[3]系列網(wǎng)絡(luò),結(jié)構(gòu)圖見圖1。ResNet提出了一種能有效降低訓(xùn)練過程中梯度消失和梯度爆炸問題的殘差模塊,其結(jié)構(gòu)圖見圖2。
針對非線性激活函數(shù)ReLU、HE Kaiming等人提出了一種權(quán)值初始化方式和另一種非線性激活函數(shù) PReLU[4],其權(quán)值計(jì)算方式為,,式中為PReLU函數(shù)第三象限的斜率;n為通道數(shù)。具體見圖3。
2? 本文變化
用CS[5]代替常用的隨機(jī)梯度下降法,采用式xi(t+1)=xi(t)+y(λ)迭代計(jì)算新解,式中,為可根據(jù)實(shí)際需要調(diào)整的步長參數(shù),常取為1;xi為每只杜鵑(即權(quán)值)的值,t為迭代次數(shù),⊕為逐項(xiàng)相乘。其迭代步驟具體如下。
(1)輸入目標(biāo)函數(shù),并初始化n個(gè)鳥巢xi(i=1,2,3…,n)。
(2)當(dāng)t<最大迭代次數(shù)時(shí),通過Lévy figthts隨機(jī)選擇一個(gè)杜鵑鳥,計(jì)算其適應(yīng)度Fi,并在它周圍任意選取一個(gè)鳥巢j。
(3)計(jì)算Fj,如果Fi>f Fj就將j作為新解。
(4)以預(yù)設(shè)的比例pa丟棄較壞的鳥巢并生成相同數(shù)目的新的鳥巢。
(5)t+1并重復(fù)第二步。
在CelebA[6]上進(jìn)行驗(yàn)證和訓(xùn)練,結(jié)果見表1。
3? 結(jié)語
CS收斂速度小于原算法,但可以取得更好的識(shí)別率。結(jié)合多次試驗(yàn),發(fā)現(xiàn)CS較不容易出現(xiàn)梯度爆炸問題,原生算法在調(diào)參時(shí)容易陷入梯度消失或爆炸問題中。
參考文獻(xiàn)
[1] ZHAO Zhongqiu, ZHENG Peng, XU Shoutao, et al. Object detection with deep learning: A review[J].IEEE transactions on neural networks and learning systems,2019,30(11):3212-3232.
[2] KAUR P, KRISHAN K, SHARMA SK, et al. Facial?recognition algorithms: A literature review[J]. Medicine, Science and the Law,2020,60(2):131-139.
[3] HE Kaiming, ZHANG Xiangyu, REN Shaoping, et al. Deep residual learning for image recognition[C]//Proceedings of the IEEE conference on computer vision and pattern recognition.2016:770-778.
[4] HE Kaiming, ZHANG Xiangyu, REN Shaoping, et al. Delving deep into rectifiers: Surpassing human?level perfor? mance on imagenet classification[C]//Proceedings of the IEEE international conference on com? puter vision.2015:1026?-1034.
[5] ZHANG Maoqing, WANG Hui, CUI Zhihua, et al. Hybrid multi?objective cuckoo search with dynamical local search[J].Memetic Computing, 2018,10(4):199-208.
[6] LIU Ziwei, LUO Ping, WANG Xiaogang, et al. Deep learning face attributes in the wild[C]//Proceedings of International Conference on Computer Vision (ICCV).2015.