Flag-Detection

The project is based on the PyTorchand uses part of the open source yolo v3 code. (I did not make too many modifications and changes to the original yolo v3 part of the code, and the original author's comments have been fully retained.)

This project provides a dataset of national flags and a detection model that I trained on Titan XP.

(When I recall which coder's code I adopted, I will indicate the source of reference here and express my gratitude)

1 Introduction

1.1 National flags categories

This project provides a dataset and a model trained based on this dataset. The project can complete the classification of 64 national flags, covering the major countries in the world, and at the same time select some countries from the beginning based on the lexicographical list of countries.

(The project has a certain degree of scalability, and more developers are welcome to improve the data set)

The following is a list of countries covered in this project:

• 'Afghanistan': 0, 'Albania': 1, 'Andorra': 2, 'Angola': 3,
• 'Argentina': 4, 'Armenia': 5, 'Australia': 6, 'Austria': 7,
• 'Azerbaijan': 8, 'Bahamas': 9, 'Bahrain': 10, 'Bangladesh': 11,
• 'Barbados': 12, 'Belarus': 13, 'Bhutan': 14, 'Bolivia': 15,
• 'Brazil': 16, 'Bulgaria': 17, 'Cambodia': 18, 'Canada': 19,
• 'CapeVerde': 20, 'Chile': 21, 'China': 22, 'Christmas Island': 23,
• 'Colombia': 24, 'Congo': 25, 'Cook Island': 26, 'Costa rica': 27,
• 'Cuba': 28, 'Cyprus': 29, 'Czech': 30, 'Democratic Republic of the Congo': 31,
• 'France': 32, 'German': 33, 'India': 34, 'Iran': 35,
• 'Iraq': 36, 'Israel': 37, 'Japan': 38, 'Jordan': 39,
• 'Laos': 40, 'Lebanon': 41, 'Malaysia': 42, 'Myanmar': 43,
• 'NewZealand': 44, 'NorthKoera': 45, 'Palestine': 46, 'Philippines': 47,
• 'Russia': 48, 'Saudi Arabia': 49, 'Singapore': 50, 'Slovakia': 51,
• 'SouthKoera': 52, 'Spain': 53, 'Sverige': 54, 'Switzerland': 55,
• 'Syrian': 56, 'Tailand': 57, 'The Central African Republic': 58, 'Turkey': 59,
• 'UK': 60, 'US': 61, 'Ukraine': 62, 'Vietnam': 63}

1.2 Dataset

The data set used in this project is in VOC format. The data set is divided into two parts:

• Manually annotated photos from the Internet:

There are about 2400 sheets in total. These photos cover 18 of the countries mentioned above (it seems to be, but not deep in my memory). It exists in the form of a file name of [XXX.jpg] in the data set.

• Generated in batches based on national flag images and random base plates: There are about 1240 sheets in total. It exists in the data set as the file name of [qwq_XXX.jpg].

The compressed package of the data set has been uploaded to the Baidu network disk:

Link: https://pan.baidu.com/s/1_xyZJAbi1sLhT1bzW97dPg

password: 6xqg

2 How to start

2.1 Data set preparation

The data set should be placed under the path'./data/dataset', and each picture should correspond to a VOC format xml tag file with the same file name.

2.2 Data set processing

First run "1.pre_process.py" to complete the preprocessing of the data set. The significance of this step is to generate two files:

(1) Extract the information in the data set in the VOC format and store it in the'./data/dataset.txt' text file to facilitate the statistics of all data samples and facilitate the statistics, scrambling and division of the data set.

(2) At the same time, the program will number all the labels that have appeared in the VOC format data set to form a mapping relationship {number: country name}, and store it in the file'./model/dict.npy'. This is done to facilitate the model to perform multiple classifications and to map to actual country names based on the classification results.

This operation will automatically check the integrity of the data set, ignoring files with missing pictures and or missing tags.

2.3 Training

Run "2.train.py" to start the training of the model. Limited by my own code development level, this project does not support running in an environment where the Cuda framework is not installed. (During the calculation process, a large number of vectors will be sent to the video memory, and designing CUDA and non-CUDA codes at the same time will significantly increase the amount of code).

• Please change the batch_size, learning_rate, epoch, ratio of training set to validation set, and the name of the model saved in the training process in ".\utils\train_details.py"
• Please change the number of categories, input size and other model data in ".\utils\config.py"

The trained model will be stored in the "model/[save_name_specified_by_yourself].pth" location

PS: The model I trained has been uploaded to Baidu Netdisk, or you can download it directly and put it in the model folder

Link: https://pan.baidu.com/s/1QO80UkUktFRfS1pTJnAeDQ

Extraction code: n9rx

2.4 Use the trained model

To use the trained model to make predictions, you can refer to a small demo I wrote "3.predict.py". This demo provides an example: how to load the model, how to open a picture and make the model complete the prediction.

• Please change the [model name] on line 26 and the [mapping between predicted serial number and country name] to be loaded on line 27 in file "yolo.py".

国旗检测

该项目基于PyTorch框架，在一定程度上使用了开源的yolo v3部分代码。(本人并未对原有的yolo v3部分的代码做过多的修饰与更改，原作者的注释得到了充分的保留。)

本项目提供了数据集和本人在Titan XP上训练完成的检测模型。

(当我回忆起我采用了哪位dalao的代码以后，我会在此注明引用来源，并表示感谢)

1 简介

1.1 支持的国旗类别

本项目提供了数据集，并提供了基于本数据集训练完成的模型。该项目可以完成对64种国旗的分类，涵盖了世界上主要国家，同时根据基于字典序的国家列表从头选择了一部分国家。

（该项目具有一定的可扩展性，欢迎更多的开发者完善数据集）

以下是本项目中所涵盖的国家列表：

• 'Afghanistan': 0, 'Albania': 1, 'Andorra': 2, 'Angola': 3,
• 'Argentina': 4, 'Armenia': 5, 'Australia': 6, 'Austria': 7,
• 'Azerbaijan': 8, 'Bahamas': 9, 'Bahrain': 10, 'Bangladesh': 11,
• 'Barbados': 12, 'Belarus': 13, 'Bhutan': 14, 'Bolivia': 15,
• 'Brazil': 16, 'Bulgaria': 17, 'Cambodia': 18, 'Canada': 19,
• 'CapeVerde': 20, 'Chile': 21, 'China': 22, 'Christmas Island': 23,
• 'Colombia': 24, 'Congo': 25, 'Cook Island': 26, 'Costa rica': 27,
• 'Cuba': 28, 'Cyprus': 29, 'Czech': 30, 'Democratic Republic of the Congo': 31,
• 'France': 32, 'German': 33, 'India': 34, 'Iran': 35,
• 'Iraq': 36, 'Israel': 37, 'Japan': 38, 'Jordan': 39,
• 'Laos': 40, 'Lebanon': 41, 'Malaysia': 42, 'Myanmar': 43,
• 'NewZealand': 44, 'NorthKoera': 45, 'Palestine': 46, 'Philippines': 47,
• 'Russia': 48, 'Saudi Arabia': 49, 'Singapore': 50, 'Slovakia': 51,
• 'SouthKoera': 52, 'Spain': 53, 'Sverige': 54, 'Switzerland': 55,
• 'Syrian': 56, 'Tailand': 57, 'The Central African Republic': 58, 'Turkey': 59,
• 'UK': 60, 'US': 61, 'Ukraine': 62, 'Vietnam': 63}

1.2 数据集

本项目使用的数据集是VOC格式的。数据集分为两部分:

• 手工标注的来自网络的照片：

共2400张左右。这些照片涵盖了上文提到的国家中的18种国家~~(似乎是的,记忆不深)~~。 在数据集中以[XXX.jpg]的文件名形式存在。

• 根据国旗图像与随机的底版批量生成的： 共1240张左右。在数据集中以[qwq_XXX.jpg]的文件名形式存在。

数据集的压缩包已经上传到百度网盘当中了：

链接：https://pan.baidu.com/s/1_xyZJAbi1sLhT1bzW97dPg

提取码：6xqg

2 如何启动

2.1 数据集的准备

数据集应当放置在'./data/dataset'路径下，每一个图片都应该对应一个文件名相同的VOC格式的xml标签文件。

2.2 数据集的处理

首先运行“1.pre_process.py”，完成对数据集的预处理，这一步操作的意义在于产生两个文件：

(1)将VOC格式的数据集中的信息提取出来，存放到'./data/dataset.txt'文本文档中，方便统计所有的数据样本，方便进行数据集的统计、打乱、划分。

(2)同时程序会将VOC格式的数据集中的所有出现过的Label进行编号，形成映射关系{编号：国家名}，存放到'./model/dict.npy'文件中。这样做是为了方便模型进行多分类，方便根据分类结果映射到实际的国家名。

该操作会自动进行数据集的完整性检查，忽略掉缺失了图片和或缺失了标签的文件。

2.3 训练

运行“2.train.py”，开始模型的训练。受限于本人的代码开发水平，本项目并不支持在未安装Cuda框架的环境中运行。（在运算过程中，多如牛毛的向量会被发送到显存中，而同时设计CUDA和非CUDA的代码会显著的增加代码量）。

• 请在utils\train_details.py 更改训练过程中的batch_size, learning_rate, epoch, 训练集与验证集的比例, 模型保存的名称
• 请在utils\config.py 更改类别数量、输入大小等其他模型数据

训练完成的模型将会存储在＂model/[你制定的保存名称].pth＂位置

PS：本人训练完成的模型已经上传到了百度网盘，也可以直接下载下来放到model文件夹

链接：https://pan.baidu.com/s/1QO80UkUktFRfS1pTJnAeDQ

提取码：n9rx

2.4 使用训练完成的模型

要使用训练完成的模型进行预测，可以参考本人写的一个小demo “3.predict.py”，这个demo提供了一种实例：如何载入模型，如何打开一个图片并使模型完成预测。

• 请在yolo.py里更改要加载的[模型的名称]、[预测的序号与国家名之间的映射]