Pan details of deep learning

Introduction to deep learning, rookie , I hope it's like taking notes and recording what I've learned , Also hope to help the same entry-level people , I hope the big guys can help correct it ~ Tort made delete .

One 、 Put forward the reason

Speaking of PAN, Can always think of similar FPN. of FPN The introduction and analysis of Kangkang can be found in the previous articles of Kangkang blogger （PAN Recommendations and FPN Study together ）

In depth learning FPN Detailed explanation _tt Ya's blog -CSDN Blog

FPN It is a feature pyramid from top to bottom , Pass down the high-level strong semantic features , Enhance the entire pyramid , But it only enhances the semantic information , But the location information is not transmitted （ Or because the upward path is too long , The transmission effect is not good ）.PAN This is the point , stay FPN Add a bottom-up pyramid after , Yes FPN To supplement , Transfer the low-level positioning features , The pyramid formed in this way combines semantic information and positioning information ,“ double play, double kill ”.

---

Two 、PAN structural analysis

PAN The structure diagram is shown in the figure below ：

stay FPN Based on the characteristic pyramid formed from top to bottom , Come to the following wave of operations PAN La

（1） First copy the bottom layer of the feature pyramid （①）, Become the bottom of the new feature pyramid .

（2） Take a down sampling operation at the bottom of the new feature pyramid , Then the penultimate layer of the original feature pyramid is a 3 * 3 Convolution , The stride is 2; Then make a horizontal connection with the bottom layer after down sampling , Add the two together . Last one more 3 * 3 Convolution to fuse their features .

（3） The operation of other layers of the new feature pyramid is similar to （2） Agreement .

---

  You are welcome to criticize and correct in the comment area , Thank you. ~