How to train OpenCV 4.0 Support Vector Machine to recognize facial features
Jan 6, 2019
In this tutorial we learn how to train a model of support vector machine, save the trained model and test the model to check the percentage of its prediction accuracy using the latest OpenCV version 4.0.
- Knowledge of Machine Learning algorithm, SVM. (Refer links: OpenCV, Wikipedia)
- Knowledge of Feature Descriptor Histogram of Oriented Gradient (HOG) (Refer links: Wikipedia)
- Image processing algorithms such as Histogram Equalization, image scaling and image resizing.
- CMake, Make, Eclipse IDE project, Git and Linux Bash scripting.
|Languages||C++, CMake, Bash|
To install the
opencv and its dependencies refer the official documents
here. We can also follow any online tutorial we get from google
search to install the opencv from source.
pip (click on their links).
We need these tools to collect data from the web, crop faces and resize them
to smaller sizes in bulk. Follow their official documents for instructions.
When data from
imagenet is not sufficient or not available we scrape the
data from the search engine results using
Fatkun Batch Download Image
chrome extension. But this method is not encouraged to avoid copyright
and legal issues when using the images from web.
Once we collect the data we need to organize it meaningfully so we can access it programmatically and manually.
While collecting data from the web store the images in the below folder structure.
FFR_dataset/ ├── Age │ ├── adult │ ├── child │ ├── old │ └── teen ├── Emotion │ ├── anger │ ├── contempt │ ├── happy │ ├── neutral │ ├── sad │ └── surprise └── Gender ├── female └── male
We will use the same directory names in the code to access them to train, save and predict the recognition results. A minimum of 50 images in each folder is required to train the models to get good prediction results. Training more images can improve the results but not recommended as it takes a lot of time to execute that and does not give significant improvements.
Test OpenCV installation
Run a sample or a test binary to confirm the working of opencv components.
This is necessary to avoid any problems later. Some samples like the
example_cpp_videocapture_basic do not work if we don’t have the video
ffmpeg installed. Hence we run the below samples to
confirm the working of all the required modules.
- To confirm the opencv version
- To confirm the working of videoio module.
Use google test filter
videoio/Videoio_Bunny.read_position/4 # GetParam() = ("mp4", FFMPEG)to test mp4 video read using ffmpeg
- To confirm the working of videoio module. Use google test filter
- To confirm the working of ml module.
Use google test filter
ML_SVM.save_loadto test SVM model train, save and load.
- To confirm the working of ml module. Use google test filter
Its a 20 hour long process to create the code we need to train the SVM model using HOG feature descriptors. There are not enough tutorials or sample code online to train a SVM model in C++. There is just one sample provided in the official opencv repo to train the SVM with HOG, train_HOG.cpp. It uses Support Vector Regression to detect people in a video or a list of images.
We have three feature types: Age, Emotion and Gender. Four age groups, six emotions and two gender types. Hence we implement a n-class classifier to recognize each feature on a face data.
The code for the same is hosted at example_cpp_train_HOG.cpp. Strangely the prediction result is below expected values for Age and Emotion features. The prediction results for Gender detection is ok. Need to confirm with the ML experts if the calculation of HOG features is ok as I had got satisfactory results with opencv 2.4 APIs. Check the result logs at the below links.
I prefer to write python scripts to execute non-core and repetitive tasks to save time. Python scripts to do the repetitive tasks like training the SVM model with variations is created and hosted at py_train_save_svm_model.py. But I am facing a bug with OpenCV 4.0 which stops the execution of the code. Hence will update the working code soon with the result logs until then make use of the procedure described here.
- Installed OpenCV 4.0 library on linux.
- Learnt how to scrape imagenet, google image search results.
- Learnt how to crop the faces from images in batches
- Learnt to prepare the dataset to train a ML algorithm.
- Converted a real world data into a feature vector.
- Trained and tested a machine learning algorithm to classify an object.
- Created ML models ready to be used in Facial Features Recognition.
- Leave a comment to ask anything related to the blog post.
- Contact me via social media for consultation, guidance, tutorials or any suggestions.