ACNet: Strengthening the Kernel Skeletons for Powerful CNN via Asymmetric Convolution Blocks

This article proposes ACNet, It uses a method that can replace the original CNN Normal square in model $d\times d$ Asymmetric convolution block of convolution kernel （ Asymmetric Convolution Block,ACB）. Asymmetric convolution block is formed by convolution kernel into square ($d\times d$)、 level ($1\times d$) And vertical ($d\times 1$) Three parallel layers of , The addition of horizontal and vertical convolution kernels and square convolution kernels can enhance the skeleton of square convolution kernels , Improve accuracy ：

explain ：

Top-5 Accuracy rate ： To an image , If the first five predicted probabilities contain the correct answer , That is to say, it is believed that .

Top-1 Accuracy rate ： To an image , If the most probable prediction is the correct answer , That is to say, it is believed that .

Use the enhanced convolution kernel parameters to initialize and deploy the original model , The accuracy of the original model can be improved without additional calculation costs .

for instance . about $3\times 3$ Convolution kernel , We can use convolution kernel as $3\times 3$、$1\times 3$ and $3\times 1$ Three parallel layers of ACB Replace it .ACB The output is the sum of the outputs of the three parallel layers , Pictured ：

that , Given the ready-made convolution network （ The original network ）, We use it ACB Replace each square convolution kernel layer to construct an asymmetric convolution network （ACNet）, And train it until it converges .

After training , We will ACB The three convolution kernels in are added at the corresponding positions , Get the skeleton (skeleton, The central intersection of the square convolution kernel ) Enhanced convolution kernel parameters ：

Finally, the weight parameters of the skeleton enhanced convolution kernel are initialized and deployed to the original network .

By the additivity of convolution （ Input $I$ For two of the right size 2D Convolution kernel ${\mathbf{K}}^{(1)}$ and ${\mathbf{K}}^{(2)}$ Yes ：

,$\oplus$ Add element by element for the corresponding position ）, The original model can be generated and trained ACNet Same output ：

That means only one $3\times 3$ The computation of convolution kernel can reach $3\times 3$、$1\times 3$ and $3\times 1$ The effect of three convolution kernels .

This way,

• Do not introduce other super parameters , It can be easily combined with some networks .
• Compared with the original network obtained by direct training , Only the weights of convolution kernel are different , Therefore, no additional calculation cost is required during deployment .

Recommended reading ：

Alchemy God trick,nb Of BN Algorithm is summarized

Convolution networks are often used to BN layer . The experiment of this article shows , First BN The post fusion effect is the best ：

If you use BN Then fuse , After training ACNet after , Convolution layer of the original model and BN Layer pass BN fusion、branch fusion Simplify to get expression substitution ：

among

$$b=-\frac{\mu \gamma }{\sigma }-\frac{\bar{\mu }\bar{\gamma }}{\bar{\sigma }}-\frac{\hat{\mu }\hat{\gamma }}{\hat{\sigma }}+\beta +\bar{\beta }+\hat{\beta }$$

The mean and standard deviation are the moving average during training .

ACNet Why it works ？

This paper observed through experiments , The weight on the skeleton of the nucleus is usually large ：

And at a certain sparsity rate (sparsity ratio) Put the convolution kernel in the corner (corner)、 skeleton (skeleton)、 overall situation (global) The weight of is randomly set to zero ：

skeleton (skeleton) The impact on accuracy is the most obvious , This shows the importance of the skeleton .

The skeleton is so important , therefore ACB Adding horizontal and vertical kernels to the skeleton of square kernels to enhance them can promote network learning .

in addition , square 、 The weights on the horizontal and vertical nuclei may be negative , So adding them together may make the skeleton weaker . However , According to the experiment, the author observed that the skeleton of each layer is always enhanced during model learning .