SSD improvements cfenet

https://arxiv.org/abs/1806.09790v1 

chart 3.CFENet And its new modules CFE.（a） Enter the dimension as 300×300 Of CFENet Topology of .（b） CFE Module layer settings , Each box represents a conv+bn+relu Group .

Pictured 3.a Shown ,CFENet Will four Integrated features enhance （CFE） modular and Two feature fusion blocks （FFB） Assemble to original SSD in . These add ons are simple , It can be easily assembled into a traditional detection network .CFE The internal structure of is shown in the figure 3.b Shown , Consisting of two similar branches . for example , In the left branch , We use k×k Conv Heel 1×1conv[9] To learn more about nonlinear relationships , And broaden the acceptance field . At the same time k×k Conv Decompose into 1×k and k×1conv layer , It not only keeps the receiving field, but also saves CFENet Reasoning time of . The other branch is different in that it reverses 1×k and k×1conv Combination of layers .CFE The module is designed to enhance SSD Used to detect shallow features of small objects , This is actually made up of several existing modules （ Such as starting module （20）、XCEPT modular [3 ]、 Large separable modules [8 ] and ResNeXt  block [23 ] Driven . 

be based on CFE modular , We propose a new primary detector CFENet, It can detect tiny objects more effectively . More specifically , First, let's go to Conv4 3 and Fc 7 Of Conv Layer and the Fc 7 and Conv6 2 Of Conv Two... Are assembled between the layers cfe. Besides , We will also have two other independent cfe Connect to respectively Conv4 3 and Fc 7 Detection branch . Because these two layers are relatively shallow , The learned features are still not conducive to the latter recognition process , We use CFE Module to enhance Conv4 3 and Fc 7 Characteristics of . Step forward , Feature fusion strategies always help to learn better features , Combine the advantages of the original features [14,17]. We are also CFENet This method is applied in . In two ffb With the help of the , Generate new by feature fusion Conv4 3 and Fc 7. In the experimental part , We set it up CFE Modular k=7.