When reading the paper, you should see two operations, one is coupling and the other is decoupling. The two words that often appear together are coupled head (Coupled head) and Decoupled head (decoupled head). So why do we need to Coupling, why should we decouple.
The concept of coupling
The following is a word encyclopedia of self-coupling
Coupled plowing (two people work together to farm. It is a way of plowing in ancient times); Coupling plowing (a farming method ). Extended to a group of two.
It can be seen that the basic meaning of coupling is to work together for a certain goal at the same time. Looking at the words Coupled and Decoupled, do you think the translations of coupling and decoupling are very subtle.
At the same time, in a Zhihu article, the difference between coupling and decoupling is also explained in a very easy-to-understand manner:Understand the two concepts of coupling and decoupling in a popular way - Zhihu Hu
The role of coupling and decoupling in deep learning
In the paper "An Integrated Model for On-Site Teaching Quality Evaluation Based on Deep Learning", there is a concise and clear diagram showing the difference between coupling and decoupling in deep learning;
In target detection, the conflict between classification tasks and regression tasks is inevitable. Therefore, decoupling heads for classification and localization are widely used in most single-stage detectors. The backbone network and feature pyramid network of the YOLO series are constantly developing, but the detection heads are still coupled. For each level of FPN features, feature integration is first performed through a 1×1 convolutional normalized activation function to reduce the number of feature channels to 256. Then add two parallel branches for classification and regression tasks respectively (as shown in Figure 2). Each branch has two 3×3 convolutional normalized activation functions for feature extraction. The Cls branch determines the type of input feature points through 1×1 convolution. Another branch also divides into two parallel branches. Among them, the Reg branch obtains the regression coefficient of the feature point through 1×1 convolution to adjust the predicted frame; and the Obj branch determines whether the feature point has a corresponding object through 1×1 convolution. Experiments show that replacing the coupling head with a decoupling head greatly improves the convergence speed, and the decoupling head is crucial for the end-to-end version of YOLO.