现代计算机(普及版)
現代計算機(普及版)
현대계산궤(보급판)
MODERN COMPUTER
2015年
7期
61-66
,共6页
卷积神经网络%深度学习%图像识别%机器学习%神经网络
捲積神經網絡%深度學習%圖像識彆%機器學習%神經網絡
권적신경망락%심도학습%도상식별%궤기학습%신경망락
Convolutional Neural Networks%Deep Learning%Image Recognition%Machine Learning%Neural Network
卷积神经网络在图像识别领域取得很好的效果,但其网络结构对图像识别的效果和效率有较大的影响,为改善识别性能,通过重复使用较小卷积核,设计并实现一种新的卷积神经网络结构,有效地减少训练参数的数量,并能够提高识别的准确率。与图像识别领域当前具有世界先进水平的ILSVRC挑战赛中取得较好成绩的算法对比实验,验证这种结构的有效性。
捲積神經網絡在圖像識彆領域取得很好的效果,但其網絡結構對圖像識彆的效果和效率有較大的影響,為改善識彆性能,通過重複使用較小捲積覈,設計併實現一種新的捲積神經網絡結構,有效地減少訓練參數的數量,併能夠提高識彆的準確率。與圖像識彆領域噹前具有世界先進水平的ILSVRC挑戰賽中取得較好成績的算法對比實驗,驗證這種結構的有效性。
권적신경망락재도상식별영역취득흔호적효과,단기망락결구대도상식별적효과화효솔유교대적영향,위개선식별성능,통과중복사용교소권적핵,설계병실현일충신적권적신경망락결구,유효지감소훈련삼수적수량,병능구제고식별적준학솔。여도상식별영역당전구유세계선진수평적ILSVRC도전새중취득교호성적적산법대비실험,험증저충결구적유효성。
Convolutional neural networks has achieved a great success in image recognition. The structure of the network has a great impact on the performance and accuracy in image recognition. To improve the performance of this algorithm, designs and implements a new architecture of the convolutional neural network by using convolutional layers with small kernel size repeatedly, which will reduce the number of training parameters effectively and increase the recognition accuracy. Compared with the state-of-art results in ILSVRC, experiments demonstrate the effectiveness of the new network architecture.