Photo ID Creator - A tool to prepare photo IDs
This app allows users to create passport photos with automatic picture size and rotation cropped to comply with the standards accepted in most countries. Users get a tiled photo in your favorite print format (e.g. 4"x6" or 5"x7") with the appropriate resolution ready for printing.
The main motivation for this app is how expensive passport photo services can be in many countries. For instance, OfficeWorks in Australia currently charges $16.95 for four passport photos.
Walgreens near my home charged me $14.99 for only two a few weeks ago. Obviously, not all photos taken at home with a smartphone camera will have the quality to meet photo ID requirements, but if you are on the budget and/or have the time to ensure you capture a picture of yourself with good resolution, proper focus and illumination, then this app will do the maths to prepare an output tiled photo ready to be printed for less than a dollar at your closest photo center.
In order to crop and scale the face of the person to a particular passport requirement, the following approach was investigated. Given the set of detected face landmarks A, B, C and D, we would like to estimate P and Q with an accuracy that is sufficient to ensure that the face in the output photo fall within the limits of the size requirements. In other words, the estimated location of the crown (P’) and chin point (Q’) should be such that the distance P’Q’ scaled by the distance between the ideal location of the crown (P) and chin point (Q) falls within the tolerance range allowed in the passport specifications. For instance, in the case of the Australian passport, the allowed scale range is ±5.88% for a face height between 32 and 36mm: [(36mm - 32mm)/(36mm + 32mm)]
To develop and validate the proposed approach, facial landmarks from the SCFace database were used. The SCFace database contains images for 130 different subjects and frontal images of each individual were carefully annotated by the Biometric Recognition Group - ATVS at Escuela Politecnica Superior of the Universidad Autonoma de Madrid [ATVS]. The procedure to estimate P’ and Q’ from A, B, C and D is as follow: Firstly, points M and N are found as the center of segments AB and CD respectively. P’ and Q’ are expected to fall in the line that passes through M and N. Then using a normalization distance KK = |AB| + |MN| and scale constants α and β, we estimate P’Q’ = αKK and M’Q’ = βKK. From the dataset α and β were adjusted to minimize the estimation error of P' and Q'.
In a second iteration, the algorithm to detect facial landmarks has been substituted by a face alignment algorithm proposed by Vahid Kazemi 2014: One Millisecond Face Alignment with an Ensemble of Regression Trees. An implementation of the algorithm is available in dlib C++ library. This library ships with a model capable of detecting 68 landmarks, but has a compressed size of ~ 61MB. This is large enough to become a problem when packing an Android application to be deployed at Google Play store or to serve it on a web application that does client side processing (like we do with WebAssembly). Because of that, I have retrained the model to detect only 28 landmarks of the original set. This reduced the model size to 16.9MB when compressed, which can still be considered large to serve on a web application, but not prohibitive.
This application is currently provided in two forms: As a command line utility or as a Firebase web app fully running on the browser (i.e. backend only serves static files). The web app can also be built and run on Android using Ionic capacitor, but this is still not well supported. The software is fully cross-platform, written in C++ and TypeScript, and has been built on Windows 10 with Visual Studio 2015 and 2017, Linux Mint with GCC 8.x and Mac with Clang 7.0. Non C++17 compliant compilers can also be used, but boost/filesystem is needed to build and run C++ unit tests.
- CMake 3.12 or higher
- Python 2.7+ (build scripts)
- A C++17 compliant compiler: On Windows use Microsoft Visual C++ 2017. On Mac LLVM 7.x. On Linux GCC 8.x
- node.js 10.x and NPM
- CMake and ninja
- Emscripten SDK
- Clang Format
- VSCode with:
- Prettier
- Clang-Format
- C/C++ Intellisense
- Angular2-inline
- Code Spell checker
- Bracket Pair Colorizer
- OpenCV 4.0
- DLib 19.6
- Google test / Google mock (v1.8.x)
- Angular 7.x / Bootstrap 4.x / Interact.js, etc.
- Android SDK
- Android Studio
- Ionic capacitor
python build.py --test -a x64python build.py --test -a wasm webpython build.py -a wasm web androidpython build.py --gen_vs_sln -bdebugsudo add-apt-repository ppa:webupd8team/java sudo apt-get update sudo apt-get install oracle-java8-installer sudo apt install oracle-java8-set-default sudo apt-get install libpcap0.8-dev