传感器与微系统
傳感器與微繫統
전감기여미계통
TRANSDUCER AND MICROSYSTEM TECHNOLOGY
2014年
8期
30-33
,共4页
卷积神经网络%递归神经网络%k均值聚类
捲積神經網絡%遞歸神經網絡%k均值聚類
권적신경망락%체귀신경망락%k균치취류
convolutional neural networks%recursive neural networks%k-means clustering
提出了一种联合卷积和递归神经网络的深层网络结构,在卷积神经网络中引入了递归神经网络能学到的组合特征:原始图片先通过一级由k均值聚类学得滤波器的卷积神经网络,得到的结果再同时通过一级卷积和一级递归神经网络,最后得到的特征向量由Softmax分类器进行分类。实验结果表明:在第二级卷积和递归神经网络权重随机的情况下,该网络的识别率已经能够达到98.28%,跟其他网络结构相比,大大减少了训练时间,而且无需复杂的工程技巧。
提齣瞭一種聯閤捲積和遞歸神經網絡的深層網絡結構,在捲積神經網絡中引入瞭遞歸神經網絡能學到的組閤特徵:原始圖片先通過一級由k均值聚類學得濾波器的捲積神經網絡,得到的結果再同時通過一級捲積和一級遞歸神經網絡,最後得到的特徵嚮量由Softmax分類器進行分類。實驗結果錶明:在第二級捲積和遞歸神經網絡權重隨機的情況下,該網絡的識彆率已經能夠達到98.28%,跟其他網絡結構相比,大大減少瞭訓練時間,而且無需複雜的工程技巧。
제출료일충연합권적화체귀신경망락적심층망락결구,재권적신경망락중인입료체귀신경망락능학도적조합특정:원시도편선통과일급유k균치취류학득려파기적권적신경망락,득도적결과재동시통과일급권적화일급체귀신경망락,최후득도적특정향량유Softmax분류기진행분류。실험결과표명:재제이급권적화체귀신경망락권중수궤적정황하,해망락적식별솔이경능구체도98.28%,근기타망락결구상비,대대감소료훈련시간,이차무수복잡적공정기교。
Propose a joint convolutional and recursive neural network structure,bring the combinational feature that recursive neural networks can learn into convolutional neural networks,that is,the raw image is first passed through a convolutional neural network stage with filters trained by k-means clustering,the result is then passed through a convolutional and a recursive neural network stage simultaneously,at last,the obtained feature vector is classified by softmax classifier. Experimental result shows that even with weights randomly set for the second convolutional and recursive neural network,the network reaches a recognition rate of 98. 28%,compared to other network structures,it greatly reduces training time and requires no complex engineering tricks.