北大选课网2021年春季验证码识别

Overview

北大选课网验证码识别 2021 年春季学期

Powered by Elector Quartet (@Rabbit, @xmcp, @SpiritedAwayCN, @gzz)

数据集描述

最初的数据集为 5130 张人工标记的验证码,之后利用早期训练好的模型在选课网上进行自动验证 (自举),又收集到 102741 张验证码,并对其中 2363 张识别错误的验证码进行人工标注。以上三部分合并为 GIF 验证码数据集

对 GIF 验证码数据集进行预处理,得到实际用于模型训练的数据集,它由裁剪好的单字母灰度图组成,图片尺寸 52x52,8 位色深,文件格式 PNG

训练集/测试集按照 8:2 进行划分,对训练集的数据进行数据增强,方式为旋转特定的几个角度,测试集不做特殊处理

在 const.py 中,DATA_RAW_DIR 对应最初的 5130 张人工标记的验证码,DATA_BOOTSTRAP_DIR 对应所有通过自举得到的验证码 (包含验证正确/错误的验证码),DATA_MANUALLY_LABELED_DIR 对应所有人工标注的验证码

数据集下载

百度网盘下载链接:https://pan.baidu.com/s/12GcfXYaLbHFC1WkzWhre2w (提取码:xsl3)

  1. 最初的 5130 张人工标注的验证码 xmcp_5130.zip,由 xmcp 提供,详见 xmcp/elective-dataset-2021spring
  2. 自举得到的验证码以 10000 张为单位,分为 11 个 bootstrap_all.x.zip
  3. 人工标注的验证码为 manually_labeled.zip,其中 manually_labeled_20210309/ 对应第一批加入到 CNN 模型训练的人工标注验证码,manually_labeled_20210310/manually_labeled_20210311/ 对应第二批加入到 CNN 模型训练的人工标注验证码
  4. 直接用于模型 210311_1 训练/测试的单字母灰度图 crop.20210311.1.tar.gz,共包含 354008 张单字母灰度图,对应于毫秒时间戳不超过 1615449512408 的 88502 张 GIF 验证码

图像预处理

基本采用 SpiritedAwayCN/ElectiveCaptCha 的方法进行预处理,相关算法的介绍参看 SpiritedAwayCN/ElectiveCaptCha 的说明

本项目对 SpiritedAwayCN/ElectiveCaptCha 的算法做了以下两点改进:

  1. 验证码的字母可能会变色,此时隔 4 帧作差,会有上一个字母的残留,因此加入 M_mask 的相关代码,每提取出一个字母,就将其加入到 M_mask 中,下一个字母提取的时候,会通过借助 M_mask 将之前已经提取出的字母对应的像素从当前字母的提取结果中抹去,以消除字母变色对后 3 个字母的提取所产生的影响
  2. 验证码的第一个字母可能在前 4 帧里颜色较浅,取第 4 帧转为灰度图后,第一个字母对应像素的灰度值大于全局平均灰度值,利用 OSTU 二值化时会被抹去,填充成白色,最后得到一张只剩下少量噪音残留的空白图像。改进的方法是,首先将 4 个关键帧的灰度图叠加,保留颜色最深的灰度值 (最小值),利用第 5-16 帧中第一个字母的颜色逐渐加深的特点,获得灰度值较深的包含第一个字母轮廓的 M_merge,此时再利用 OSTU 二值化,就基本上能留住第一个字母的信息,然后将 M_merge 的右三个字母抹去,仅保留包含第一个字母的轮廓,再根据此时的 M_merge,将第 1 个关键帧的灰度图中第一个字母对应像素的灰度值降低,此时再调用之前的处理函数,就可以更加准确地提取出第一个字母

模型训练与测试

以测试集的测试结果作为模型训练结果的参考,以模型在选课网上在线验证的结果作为模型的实际准确率

模型 测试集准确率 (单字母) 在线测试准确率 (四字母) / % 备注
210308_3 0.9968 (15818 / 15869) 0.9675 (17253 / 17832) 扩大数据集,进行数据增强
210309_2 0.9952 (33811 / 33973) 0.9799 (11659 / 11898) 扩大数据集,加入第一批人工标注的验证码
210309_3 0.9951 (33807 / 33973) 0.9788 (11847 / 12104) 同 210309_2,但数据增强的旋转角度范围缩小至 [-15, 15]
210309_4 0.9928 (33730 / 33973) 0.9727 (9527 / 9794) 同 210309_2,但不进行数据增强
210311_1 0.9977 (70467 / 70630) 0.9916 (10219 / 10306) 扩大数据集,加入第二批人工标注的验证码,优化第一个字母的图像处理算法
210311_2 0.9961 (70643 / 70918) 0.9838 (11230 / 11415) 同 210311_1,但使用未经优化的图像处理算法

目前准确率最高的模型为 210311_1,该模型的训练结果、训练时的终端输出、训练/测试的相关代码已在 dist/ 中提供

环境依赖

开发环境 Python 版本为 3.6.8,相关依赖包版本参考 requirements.txt,其中 PyTorch 版本应大于 1.4.x,否则无法读取 dist/ 中提供的模型

模型训练在阿里云服务器上进行,CUDA 版本 11.0,使用的 GPU 为 Tesla V100-SXM2-32GB,使用的 CPU 为 Intel(R) Xeon(R) Platinum 8163 CPU @ 2.50GHz,训练集一次性读入内存,以 210311_1 的训练/测试为例,数据量为 88502 张 GIF 验证码,数据增强的旋转角度范围为 [-30, 30],数据集扩增至 9 倍规模,占用的内存量为 32.9 GB,每个 Epoch 的耗时约为 309 秒

文件描述

preprocess.py

图像预处理的相关算法,实际有用的函数是 extract_c0_v1, extract_c0_v2, extract_c123, crop,处理流程参考 main 函数

cnn.py

与 CNN 模型相关的代码,包括数据集定义、数据增强、CNN 模型定义、训练集/测试集划分、模型训练、模型测试等

bootstrap.py

与模型自举相关的代码,在选课网上在线测试现有模型的准确率,并且将预测正确/错误的模型分类收集,用于扩增数据集

label_server.py

用于人工标注验证码的网站后端,前端代码在 web/ 中。提供图像预处理、输入检查等功能,为人工标注验证码提供方便,所使用的数据源为模型自举过程中收集到的识别错误的验证码

config.ini

与选课网账户相关的信息,包括学号、密码、是否为双学位账号,主要用于 bootstrap.py

致谢

Owner
Rabbit
CCME, Peking University
Rabbit
Python PID Controller and Process Simulator (FOPDT) with GUI.

PythonPID_Simulator Python PID Controller and Process Simulator (FOPDT) with GUI. Run the File. Then select Model Values and Tune PID.. Hit Refresh to

19 Oct 14, 2022
OWASP Foundation Web Respository

WWWGrep OWASP Foundation Web Respository Author: Mark Deen & Aditi Mohan Introduction WWWGrep is a rapid search “grepping” mechanism that examines HTM

OWASP 34 Jun 15, 2022
*考研学习利器,玩电脑控制不住自己时,可以使用该程序定日期锁屏,同时有精美壁纸锁屏显示,也不会枯燥。

LockscreenbyTime_win10 A python program in win10. You can set the time to lock the computer(by setting year, month, day), Fullscreen pictures will sho

PixianDouban 4 Jul 10, 2022
OLDBot (Online Lessons Discord Bot)

This program is designed to facilitate online lessons. With this you don't need to get up early. Just config and watch the program resolve itself. It automatically enters to the lesson at the specifi

Da4ndo 1 Nov 21, 2021
Hera is a Python framework for constructing and submitting Argo Workflows.

Hera is an Argo Workflows Python SDK. Hera aims to make workflow construction and submission easy and accessible to everyone! Hera abstracts away workflow setup details while still maintaining a cons

argoproj-labs 241 Jan 02, 2023
Скрипт позволяет заводить задачи в Панель мониторинга YouTrack на основе парсинга сайта safe-surf.ru

Скрипт позволяет заводить задачи в Панель мониторинга YouTrack на основе парсинга сайта safe-surf.ru

Bad_karma 3 Feb 12, 2022
Abilian Core: an enterprise application development platform based on the Flask micro-framework, the SQLAlchemy ORM

About Abilian Core is an enterprise application development platform based on the Flask micro-framework, the SQLAlchemy ORM, good intentions and best

Abilian open source projects 47 Apr 14, 2022
Let's pretend you want to create a AWS Lambda project called "sns-processor".

Usage Let's pretend you want to create a AWS Lambda project called "sns-processor". Rather than using lambda and then editing the results to include y

1 Dec 31, 2021
Reference python implementation of Chia pool operations for pool operators

This repository provides a sample server written in python, which is meant to server as a basis for a Chia Pool. While this is a fully functional implementation, it requires some work in scalability

Chia Network 451 Dec 13, 2022
Simple script to match riders with drivers.

theBestPooler Simple script to match riders with drivers. It's a greedy, unoptimised search, so no guarantees that it works. It just seems to work (ve

Devansh 1 Nov 22, 2021
Beginner Projects A couple of beginner projects here

Beginner Projects A couple of beginner projects here, listed from easiest to hardest :) selector.py: simply a random selector to tell me who to faceti

Kylie 272 Jan 07, 2023
A performant state estimator for power system

A state estimator for power system. Turbocharged with sparse matrix support, JIT, SIMD and improved ordering.

9 Dec 12, 2022
A demo of a data science project using Kedro

iris Overview This is your new Kedro project, which was generated using Kedro 0.17.4. Take a look at the Kedro documentation to get started. Rules and

Khuyen Tran 14 Oct 14, 2022
token vesting escrow with cliff and clawback

Yearn Vesting Escrow A modified version of Curve Vesting Escrow contracts with added functionality: An escrow can have a start_date in the past.

62 Dec 08, 2022
Google Fit Sensor Component

Google Fit Sensor Component

Ivan Vojtko 21 Dec 20, 2022
Pdraw - Generate Deterministic, Procedural Artwork from Arbitrary Text

pdraw.py: Generate Deterministic, Procedural Artwork from Arbitrary Text pdraw a

Brian Schrader 2 Sep 12, 2022
Thinky nature dots with python

Thinky Nature Welcome to my rice dotfiles of my ThinkPad X230 You surely will need to change some configs to fit your files and stuff and maybe tweak

Daniel Mironow 26 Dec 02, 2022
A faster copy of nell's comet nuker

Astro a faster copy of nell's comet nuker also nell uses external libraries like it's cocaine man never learned to use ansi color codes (ily nell) (On

horrid 8 Aug 15, 2022
A Python script made for the Python Discord Pixels event.

Python Discord Pixels A Python script made for the Python Discord Pixels event. Usage Create an image.png RGBA image with your pattern. Transparent pi

Stanisław Jelnicki 4 Mar 23, 2022
Architectural Patterns implementation by using notification handler module prototype

This repository covers singleton, indirection, factory, adaptor, mediator patterns in python language by using university hypothetical notification module prototype. The code is just for demonstratin

Muhammad Umair 2 Jan 08, 2022