VGG论文阅读笔记

VERY DEEP CONVOLUTIONAL NETWORKS FOR LARGE-SCALE IMAGE RECOGNITION——VGG

论文作者 KarenSimonyan & Andrew Zisserman

原文 点击打开链接

I.  Features网络新特点

1.    Very deep networks, large scale image recognition更深，数据集更大

2.    Using small receptive field 更小的卷积核

a stack of two 3×3 conv.layers (without spatial pooling in between) has an effective receptive field of5×5; three such layers have a 7 × 7 effective receptivefield.

decreases the number of parameters :  e.g.  7²C² = 49C²-->3(3²C²)= 27C²

2.  The incorporation of 1 × 1 conv. layers is away to increase the nonlinearity of the decision

function without affecting thereceptivefields of the conv. layers发现用1 × 1卷积层可以提升效果

 

II. Architecture网络架构

● input224×224，subtracting the mean RGB value 输入图像224×224，只有一个channel, RGB取均值

● smallreceptive field: 3 × 3;

● 1 ×1 convolution filters are used;

● convstride:1 (padding=1)

● five max-pooling layers: 2 × 2 pixel window,with stride 2.池化层2 × 2，不重叠

● Theconfiguration of the fully connected layers is the same in all networks 每个网络后都跟3个全连接层

● Allhidden layers are equipped with ReLU, none (except one) contain Local Response Normalization (LRN) 都用ReLU,都不用LRN（测试结果表明加LRN效果没有明显提升）

与AlexNet的对比：模仿AlexNet的5conv+3fc结构，但减小卷积核的大小，增加层数,有更多的Channel数（下图AlexNet）

III. Training

1.    the training is carried out by optimizing the multinomial logistic regression objective using mini-batch gradient descent (based on back-propagation (LeCunet al., 1989)) with momentum. （这里还不太理解）多项式回归，梯度下降，BP

batch size: 256  momentum:0.9  weight decay: L2=5·10−4 

dropout regularization for thefirst two fully-connected layers前两层有Dropout

learning rate was initially set to 10⁻², and then decreased by a factorof 10 when the validation set accuracy stopped improving

2.    initialization权值初始化

began with training the configuration A , shallow enough to betrained with random initialization. When training deeper architectures, weinitialized the first four convolutional layers and the last three fully connected layerswith the layers of net A

先随机初始化A（因为A相对浅，可以随机赋初值），用A训练后的值初始化更深的网络

it is possible to initialize the weights without pre-training by usingthe random initialization procedure

不用预训练参数赋初值也是可能的（论文中没有展开叙述）

3.    Trainingimage size (S --- the smallest side of an isotopically-rescaled training image)

输入图像先经过预处理缩放，有single-scale和multi-scale两种

1） single-scaletraining S=256/384

2） multi-scaletraining S∈[S_min,S_max]

每次数据输入的时候，每张图片被重新缩放，缩放的短边S随机从[256,512]中选择一个。

IV.  Testing

input image rescaled---network applied---classscore map spatially averaged

multi-crop evaluation每张图像可以剪裁出多个224×224图像

V.  Evaluation测试结果

A与A-LRN差别不大----加LRN没有明显作用

层越深效果越好，到VGG19在现有数据集训练效果达到饱和

C优于B----加conv1可以提升效果

D优于C----conv3比conv1效果好

multi-scale优于single scale

multi-crop & dense 效果最好