We can clearly see that the 256-dimensional channel has been reduced to 64 dimensions. Then, 1 Â 1 convolution is adopted for recovery, which can directly reduce
the number of dimensions and parameters through these two methods. When the
number of channels F(x) and x are different, the calculation methods for different
channels can be divided into the following two types (Fig. 4).
The connection part of the solid line adopts the convolution of 3 Â 3 Â 64, and the
number of channels is the same. The calculation method is as follows:
y ¼ Fx þ x
ð2Þ
Fig. 3. Two different residual models
Fig. 4. Different stacking modes
Combined with DCT, SPIHT and ResNet to Identify Ancient …
317