COCR

COCR is designed to convert an image of handwriting chemical structure to graph of that molecule.

COCR, Optical Character Recognition for Chemical Structures, was once a demo for my undergraduate graduation thesis in 2021.6. It brings OCSR(optical chemical structure recognition) capability into handwriting cases.

Symbol	String	Ring	Solid-	Hash-	Wavy-	Single-	Double-	Triple-
Appearance	(CH2)2COOEt	⏣	▲	△	~~	/	//	///
Status	✔️	✔️	✔️	✔️	✔️	✔️	✔️	✔️

COCR is developed under Qt framework. It handles images with YOLO and CRNN models using opencv or ncnn backend.

Online Demo

• try on chromium-based web browser like Chrome and Edge: https://xuguodong1999.github.io/COCR.html

Supported platforms:

• Windows
• Android
• WebAssembly
• Mac OS
• Linux

News

• [2021/06] COCR v1.1 released with support for strings and wavy-bonds.

v1.1.on.Ubuntu.mp4

• [2021/02] COCR v1.0 released for simple cases.

Input	Detection	Render

1. Support single element symbols: C、H、O、N、P、B、S、F、Cl、Br、I.
2. Support bond types: single, double, triple, hash wedge, solid wedge, circle.

Build

System Requirements

1. CMake (At least 3.22), Ninja
2. Git (For submodule clone)
3. Qt 6 (Test with 6.5.2)
4. Latest C++ Compiler (Test with GCC-11.x, MSVC-v17.x, Clang-16.x)

Qt Beginner's Guide

• for windows developer: have 7zip, wget and git-bash ready, or simply use WSL:

# download mingw, msvc binary
wget -r -np -nH -e robots=off http://mirrors.nju.edu.cn/qt/online/qtsdkrepository/windows_x86/desktop/qt6_652/
# extract all 7z files to a folder called "result"
find . -name '*.7z' -exec 7z x {} -aos -o./result \;

• for linux developer:

# download gcc_64 binary
wget -r -np -nH -e robots=off http://mirrors.nju.edu.cn/qt/online/qtsdkrepository/linux_x64/desktop/qt6_652/
# extract all 7z files to a folder called "result"
find . -name '*.7z' -exec 7z x {} -aos -o./result \;

• for macos developer:

# download clang_64 binary
wget -r -np -nH -e robots=off http://mirrors.nju.edu.cn/qt/online/qtsdkrepository/mac_x64/desktop/qt6_652/
# extract all 7z files to a folder called "result"
find . -name '*.7z' -exec 7z x {} -aos -o./result \;

after downloading and extracting, the folder structure looks like:

result/6.5.2
└── gcc_64
OR
result/6.5.2
└── msvc2019_64
OR
result/6.5.2
└── clang_64

Then copy gcc_64 or msvc2019_64 or clang_64 to your ideal library path and it's OK.

---

Get the Code

git clone https://github.com/xuguodong1999/COCR.git --branch main --single-branch --recursive

or using ssh

git clone git@github.com:xuguodong1999/COCR.git --branch main --single-branch --recursive

All third-party libraries except Qt are in third_party directory, including boost, openbabel, rdkit, opencv, etc.

Only a minimal set of source codes is kept, and you can find custom changes in patch files under third_party directory.

Nearly all build scripts for third-party libraries have been rewritten, to make cross-build, bugfix and feature hack easier.

Compile

Building COCR project is the same as common Qt 6 projects, the basic step is:

mkdir build && cd build
cmake .. -G Ninja -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON -DCMAKE_PREFIX_PATH=path/to/Qt6/binary/dir
cmake --build . --parallel --target COCR

for example, on linux desktop,

cmake .. -G Ninja -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON -DCMAKE_PREFIX_PATH=~/shared/Qt/6.5.2/gcc_64

More custom flags to control SIMD, Release Info, etc.

FLAG	VALUE	USAGE
XGD_USE_OPENMP	ON/OFF	enable openmp, default auto-detect
XGD_USE_VK	ON/OFF	enable vulkan, default auto-detect
XGD_USE_CCACHE	ON/OFF	enable ccache, default auto-detect
XGD_FLAG_MARCH_NATIVE	ON/OFF	add -march=native for gcc compiler, default ON
XGD_FLAG_WASM_SIMD128	ON/OFF	add -msimd128 for wasm, default ON
XGD_BUILD_WITH_GRADLE	ON/OFF	disable custom cmake directory layout for android studio cmake plugin, default OFF
XGD_OPT_RC	ON/OFF	add release info for products, default OFF
XGD_NO_DEBUG_CONSOLE	ON/OFF	hide debug console, default OFF
XGD_OPT_ARCH_X86	ON/OFF	enable x86 specified code, default auto-detect
XGD_OPT_ARCH_ARM	ON/OFF	enable ARM specified code, default auto-detect
XGD_OPT_ARCH_MIPS	ON/OFF	enable mips specified code, default auto-detect
XGD_OPT_ARCH_POWER	ON/OFF	enable power pc specified code, default auto-detect
XGD_OPT_ARCH_32	ON/OFF	enable 32-bit specified code, default auto-detect
XGD_OPT_ARCH_64	ON/OFF	enable 64-bit specified code, default auto-detect
XGD_FLAG_NEON	ON/OFF	enable simd neon, default auto-detect
XGD_FLAG_FMA	ON/OFF	enable simd fma, default auto-detect
XGD_FLAG_F16C	ON/OFF	enable simd f16c, default auto-detect
XGD_FLAG_XOP	ON/OFF	enable simd xop, default auto-detect
XGD_FLAG_SSE_ALL	ON/OFF	enable simd sse/sse2/sse3/ssse3/sse41/sse42, default auto-detect
XGD_FLAG_AVX	ON/OFF	enable simd avx, default auto-detect
XGD_FLAG_AVX2	ON/OFF	enable simd avx2, default auto-detect

---

Cross-Compile

There is a typescript based cmake wrapper under www/tools/cmake-builder, to manage complex environment variables, toolchains, system libraries and cmake options for many cross-compile cases, including mxe, emscripten, msvc-wine, etc.

Data

COCR uses SCUT-COUCH2009 as meta handwriting data, and uses QtGui::QTextDocument as rich text renderer.

A chemical structure generator for handwriting cases is written to provide training data for YOLO and CRNN models, which composes meta-character into random chemical structure formulas.

COCRDataGen is a gui/cli tool to display or generate training samples.

1. For GUI, directly running COCRDataGen will display samples through cv::imshow

./COCRDataGen

2. For CLI, add -yolo [number of samples] [an empty, existing directory path], for example,

# generate 10 object detection samples under ./yolo_data/
./COCRDataGen -yolo 10 ./

3. For CLI, add -crnn [number of samples] [an empty, existing directory path], for example,

# generate 10 text recognition samples under ./crnn_data/
./COCRDataGen -crnn 10 ./

4. For CLI, add -isomer [number of samples] [an empty, existing directory path], for example,

# generate all alkane isomers for C-num ≤ 16 namely {CARBON_NUM}.dat under ./
# dont play with number over 20 without taking a look at src/data_gen/isomers.cpp.
# it may comsume a lot of memory and cpus.
./COCRDataGen -isomer 16 ./

License

GPLv3 Clause

Citation

@misc{COCR2021,
  title = {COCR: Optical Character Recognition for Handwriting Chemical Structures},
  author = {xuguodong1999},
  year = {2021},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/xuguodong1999/COCR}}
}