Res2net:A new multi-scale backbone architecture（Res2net: A new multi-scale backbone architecture ）

 

·                         ​​​​​​​        ​​​​​​​        ​​​​​​​        ​​​​​​​        ​​​​​​​         Figure 1

The module on the right adds a multi branch structure to the traditional bottleneck block on the left , This branch can observe the input image with different pixel sizes , It expands the perception field of the system to the image , Can increase the amount of information read .

                                                                    Figure 2

                                                                    Figure 3

As shown in Figure 1 , The first branch uses the original field of view , So do not use change focus processing ; The second branch changes the focal length of the original field of view ; The third branch is further changed on the basis of the second branch , Reduce the scope of observation again , Expand the observation accuracy of some areas ; The fourth branch performs the same operation based on the third branch . Because the four branches come from different channels , So finally use conv_cat() Connect the four branches .

Feature Aggregating via Parallel Partial Decoder（ Feature aggregation through parallel partial decoder ）

                                                                     Figure 4

                                                                     Figure 5

                                                                     Figure 6

                                                                       Figure 7

The function of parallel decoder is to reduce the huge amount of low-level feature data , The proportion of useful data is low . Extract only high-level features , And multiple high-level features are extracted and connected to improve the content of target data .

The purpose of multiple up sampling in Figure 7 , Bloggers think it is a relatively low-level high-level （ Third floor to fourth floor ） Come on , The bottom layer has more data , The target data content is low , Therefore, multiple sampling is adopted to reduce non target data .

 Reverse Attention Module（ Reverse attention module ）

                                                                         Figure 8

                                                                         Figure nine

                                                                         Figure ten

By removing the currently predicted significant region from the side output feature , The network can eventually detect missing object parts and details , So as to obtain high resolution and accuracy .

As shown in Figure 8 , Connect the global graph to the fifth layer in turn （f5）、 The fourth level （f4） And the third floor （f3） Aggregation uses three-tier features to make up for the missing features of the global graph .

Figure 10 shows the aggregation operation of the reverse attention module in the third layer and the global diagram aggregated through the fourth and fifth layers . Layer 5 and global graph aggregation , The fourth layer is similar to the global graph aggregation and figure 10 after the aggregation of the fifth layer .