Semantic segmentation ｜ learning record (5) FCN network structure officially implemented by pytoch

---

Preface

Pytorch Officially realized FCN It is slightly different from the structure diagram of the paper in that year , Because now there are more backbone The choice of , And the expansion convolution technology is applied .

FCN Network structure

The picture below is Pytorch Officially realized FCN Network structure ：

The picture below is Resnet Network structure ：

Compare the two networks for analysis ：FCN Online backbone The choice is Resnet50 backbone, In this part of the blue dotted box in front of the structure and Resnet It's exactly the same .

Focus on layer3 and layer4 Part of .FCN Of layer3 and layer4 Corresponding to Resnet Of conv4_x and conv5_x. The difference lies in ：

• Resnet Of conv4_x Medium 6 The residual structures correspond to FCN Of layer3 Medium 1 individual Bottleneck1 and 5 individual Bottleneck2.
• Resnet Of conv5_x Medium 3 The residual structures correspond to FCN Of layer4 Medium 1 individual Bottleneck1 and 2 individual Bottleneck2.

Let's focus on that Bottleneck1 and Bottleneck2 Structure ：

• Bottleneck1 Corresponding to the dotted line structure in the residual structure , Here and resnet The difference is that it will be a shortcut to branch on $1\ast 1$ Convolution The step length is changed to 1. Therefore, there is no down sampling operation here , If the down sampling ratio in semantic segmentation is too large, the original image effect will be affected . In addition to $3\ast 3$ At the convolution kernel The step length is also changed to 1, And the expansion convolution is introduced .
• Bottleneck2 It also introduces Expansion convolution .

Let's take a look at FCN Head part ：

• Through a 3*3 The convolution of layer , Enter the... Of the characteristic graph channel The number is adjusted to the original 1/4 Turn into 512.
• Through one Dropout layer
• Through one more 1*1 Convolution layer , Adjust the feature layer channel For the number of categories .
• Finally, through a bilinear interpolation method , Return to the original size .

There is another one on the right FCN Head, The official reasons ： Preventing error gradients cannot be transferred to the shallow layer of the network , So the auxiliary classifier is introduced here .
You can enable this auxiliary classifier during training , You can try it . When actually deploying to the real environment , We only use output.

Reference material

up Lord thunderbolt Wz Original video