Deep learning classification network -- vggnet

Deep learning classification network Summary

1. AlexNet
2. VGGNet

---

---

Preface

VGGNet yes 2014 year ILSVRC The second place , The first name is GoogLeNet. Put forward VGG Our motivation is mainly to study the influence of network depth on its accuracy , The main idea is to use only 3×3 Convolution builds a deep network . It turns out that , When the depth of the network reaches 16~19 Can achieve better performance .

One 、 Network structure

A total of 6 It's a network structure , This design has a certain purpose ：

• contrast A and A-LRN Performance on the test set LRN Whether it works ;
• contrast A and B To determine whether adding a convolution layer in the shallow layer of the network is effective ;
• contrast B and C To determine the increase in the depth of the network 1×1 Whether the convolution layer is effective ;
• contrast C and D To make sure 1×1 Convolution kernel 3×3 Convolution is better or worse ;
• contrast D and E To determine the increase in the depth of the network 3×3 Whether convolution is effective ;
  

Two 、 The main points of

1. Abandoning LRN

The Internet A-LRN Of top1 and top5 The error rate is higher than that of the network A, So in the following B、C、D、E None of them are used LRN layer .

2. Small size convolution

• Use stacked 3×3 Convolution instead of 5×5 or 7×7 Convolution , There are two advantages ：
  ① Add more nonlinearity on the basis of keeping the receptive field unchanged , At the same time, the network is also deeper . Suppose you use 3 individual 3×3 Convolution instead of a 7×7 Convolution , After each convolution, there is a ReLU, It is equivalent to adding two additional ReLU;
  ② Reduce the number of parameters . Assume that the input and output channels are C, Then use 7×7 The parameter of convolution is $7\ast 7\ast C\ast C=49C^2$, Use 3 individual 3×3 The parameter of convolution is $3\ast 3\ast C\ast C\ast 3=27C^2$, Reduced the appointment 45%.
• Use 1×1 Convolution increases nonlinearity without affecting receptive fields .（1×1 Convolution in Network In Network Used in architecture ）

3. Weight initialization

• Random initialization A Online weights and bias, Yes A Network training .
• In training deeper BCDE When the network , Use A Network parameters to initialize their pre 4 A convolution and post 3 All connection layers , The middle layer is initialized randomly , Do not change the learning rate of the pre initialization layer .
• Random initialization ：$w\sim N(0,0.01)$,$bias=0$

notes ： Mentioned in the text Xavier(glorot) Initialization can initialize weights without pre training .

4. Multiscale training

Make S Is the minimum side length of the training image after rescaling , Crop in rescaled image 224*224 As input to the network , therefore S≥224. There are two ways to set S：

• Fix S, Corresponding to single scale training ： Use S=256 and S=384 Two scales , First use S=256 Training network , To speed up S=384 Network training , Use S=256 Parameter initialization of S=384 Network weight , And use a smaller learning rate 0.001.
• from [256,512] Random sampling in range S, Corresponding to multi-scale training ： For speed reasons , adopt fine-tune Single scale with the same configuration （S=384） Pre train all layers of the model to train the multiscale model .

Why multi-scale training ？

• The same kind of objects can have different scales in different images , Therefore, it is beneficial for network learning to consider multi-scale in training ;
• It can be seen as a kind of data enhancement , To train a single model that can recognize objects of various scales

5. Multiscale & Multi crop test

Multi scale testing （ This method is called ‘dense’） Divided into the following steps ：

• Rescale the training image to the scale Q（ That is, the minimum side length ）, also Q It doesn't need to be equal to S, Because the article puts forward Converting a fully connected layer to a convoluted layer , So it can adapt to different sizes of input ;
• Q There are two choices ： For single scale training ,Q={S-32, S, S+32}; For multiscale training ,Q={256, 384, 512} ;
• Zoom to scale Q The image is directly input into the full convolution network , The feature map obtained at each scale is averaged in the spatial dimension , Get the category score vector , Then average the category score vectors of all scales to obtain the final result .

About “ Converting a fully connected layer to a convoluted layer ”： Because the input during training is 224×224, Output characteristic diagram after maximum pooling size Fixed for 7×7×512, The input channel of the first full connection layer is 7×7×512=25088, The output channel is 4096, Then the conversion to convolution is 4096 individual 7×7×512 Convolution kernel . Empathy , The second and third fully connected layers are converted to include 4096 A dimension for 1×1×4096 The convolution layer of the convolution kernel .

Multi crop test （multi-crop）： For each scale , Cut from 50 picture 224*224 Image , The three scales are 150 Images , Average the category score vector of all images as the final result .

3、 ... and 、 experimental result & Conclusion

1. Single scale testing

• Using local response normalization (A-LRN The Internet ) It cannot improve the model without any normalization layer A, Therefore, there is no deeper architecture (B-E) Use in LRN;
• A~E Decreasing error rate , It shows that increasing the network depth within a certain range can improve the classification accuracy ;
• Although the additional nonlinearity (1×1 Convolution ) It can really improve the accuracy (C Than B Better ), But by using 3×3 Convolution to capture spatial context is also important (D Than C good );
• Multi scale training is better than single scale training , It is confirmed that the training set enhanced by scale jitter is indeed helpful to capture multi-scale image statistical data .

2. Multiscale & Multi crop test

• Both multi-scale and multi tailoring tests can reduce the error rate , The combination of the two can further reduce the error rate .

Reference material

Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition[J]. arXiv preprint arXiv:1409.1556, 2014.