Unlike the first two networks, VGG Net uses only 3 Ã 3 filters. The authors believe
that two 3 Ã 3 filters can be combined to generate a 5 Ã 5 sensing zone [4]. In this way,
it can maintain the same function as the large size and has the advantage of small size.
For example, the small size can reduce the parameter, and the multi-purpose linear
rectification layer can reduce the performance of the system. Another feature of the
network is that it emphasizes depth, that is, the layer with the weights is superimposed
on the 16th or 19th layer, and the network structure deep enough to reflect the hierarchy
of processing visual data. VGG Net maintains depth and keeps it simple. Its structure is
shown in Fig. 4.
Fig. 3. Basic working principle of deconvolution
Fig. 4. Architecture of VGG Net use only 3 Ã 3 filter size
518
C. Chen and Y. Tong
that two 3 Ã 3 filters can be combined to generate a 5 Ã 5 sensing zone [4]. In this way,
it can maintain the same function as the large size and has the advantage of small size.
For example, the small size can reduce the parameter, and the multi-purpose linear
rectification layer can reduce the performance of the system. Another feature of the
network is that it emphasizes depth, that is, the layer with the weights is superimposed
on the 16th or 19th layer, and the network structure deep enough to reflect the hierarchy
of processing visual data. VGG Net maintains depth and keeps it simple. Its structure is
shown in Fig. 4.
Fig. 3. Basic working principle of deconvolution
Fig. 4. Architecture of VGG Net use only 3 Ã 3 filter size
518
C. Chen and Y. Tong