Dilated convolutions--Expansion convolution

Article Author: Tyan
Blog: noahsnail.com |  CSDN | Pinterest, 1. Expansion convolution

Dilated convolutions, translated as expanded convolution or void convolution. The expansion convolution, in addition to the size of the convolution nucleus, has a dilation rate parameter, which is mainly used to indicate the size of the expansion, compared to the normal convolution. The same point in the expansion convolution and the ordinary convolution is that the size of the convolution nucleus is the same, in the neural network, that is, the number of parameters is the same, the difference is that the expansion convolution has a greater sense of the wild. The sensing field is the size of the convolution nucleus seen on the image, such as a 3x3 3 \times 3 convolution nucleus with a size of 9. 2. Schematic diagram

The image below is a schematic of the expanded convolution.

(a) ordinary convolution, 1-dilated convolution, convolutional nuclei of the field for 3x3=9 3 \times 3 = 9.
(b) Expansion of convolution, 2-dilated convolution, convolutional nuclei of the field for 7x7=49 7 \times 7 = 49.
(c) Expansion of convolution, 4-dilated convolution, convolutional nuclei of the 15x15=225 \times 15 = 225.

As can be seen from the above figure, the number of parameters of the convolution kernel remains constant, and the size of the field of the dilation rate increases exponentially with the increase of the parameter of "the". 3. Advantages

The extended convolution increases the sensing field of the convolution nucleus without changing the number of parameters, and it ensures that the output feature map (feature map) remains the same size. A 3x3 convolution core with an expansion rate of 2, the same as the 5x5 convolution core, but with only 9 parameters, is 36% 36\% of the number of 5x5 convolution parameters. 4. Application

Expanded convolution has been applied in image segmentation, speech synthesis, machine translation and target detection. 5. References

Multi-scale Context Aggregation by dilated Convolutions