当前位置：网站首页>Yolov5 network structure + code + application details | CSDN creation punch in

Yolov5 network structure + code + application details | CSDN creation punch in

2022-07-03 05:07:00 【TT ya】

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 .

This article is right YOLOv5 A summary of the series .

Catalog

One 、 Input analysis

Two 、 Overall network structure diagram and general introduction

3、 ... and 、Backbone

Four 、Neck

5、 ... and 、Head

6、 ... and 、 The code realizes the specific construction of the network

7、 ... and 、 utilize YOLOv5 Train your own dataset

One 、 Input analysis

YOLOv5 Input end （ One ）—— Mosaic Data to enhance |CSDN Creative punch in _tt Ya's blog -CSDN Blog

YOLOv5 Input end （ Two ）|CSDN Creative punch in _tt Ya's blog -CSDN Blog （ Adaptive anchor frame calculation + Adaptive image scaling ）

Two 、 Overall network structure diagram and general introduction

The network structure diagram is shown in the figure below ：

（ among CBL The annotations of are all about output ）

Thank you here for your blog friends 352（ Brother, what is the symbol in front of you? I can't show it here , sorry ） Reminder

v5.0 edition

Current update to v6.0 edition

（ The big change here is ： use CBL Instead of Focus layer , It is mainly to facilitate model export ）

YOLOv5 The network structure is divided into 3 Parts of ,Backbone（ The main part ）,Neck and Head.

Backbone effect ： feature extraction

Neck effect ： A wave of mixing and combination of features , And pass these characteristics to the prediction layer

Head effect ： Make the final prediction output

3、 ... and 、Backbone

Backbone Mainly by Focus,CBL,BottleneckCSP（/C3） as well as SPP Other components .

Focus effect ： Perform a down sampling , And reduce the amount of computation and speed up the network .（v5.0）

CBL effect ：（ It's what we often call convolution , Although the convolution here is not in the direct sense nn.Conv2d.） Get features .

of Focus and CBL The analysis can be found in the blog written by Kangkang blogger

YOLOv5 Medium Focus Layer details _tt Ya's blog -CSDN Blog

BottleneckCSP（/C3） effect ： Residual structure , Learn more features .

of BottleneckCSP（/C3） The analysis can be found in the blog written by Kangkang blogger

YOLOv5 Medium CSP structure _tt Ya's blog -CSDN Blog

SPP/SPPF effect ： Space Pyramid pooling , Transform the arbitrary size feature map obtained above into a fixed size feature vector .

（v5.0 yes SPP,v6.0 yes SPPF）

of SPP/SPPF The analysis can be found in the blog written by Kangkang blogger

YOLOv5 Medium SPP/SPPF The structure, _tt Ya's blog -CSDN Blog _spp yolov5

Four 、Neck

Neck To put it bluntly FPN+PAN The design of the structure .

About FPN and PAN You can read the previous articles written by bloggers

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

Deep learning PAN Detailed explanation _tt Ya's blog -CSDN Blog

（ here v5.0 and v6.0 Please refer to the above figure , Too lazy to change, hee hee ）

5、 ... and 、Head

I understand. Head It's the output . If only from the network structure , He is right Neck Coming out 3 The outputs are convoluted separately . However Head The main part of is its output reasoning and processing . See the previous articles of bloggers for details , There is a loss function ,NMS And source code analysis .

YOLOv5 Detailed explanation of the output end of _tt Ya's blog -CSDN Blog