There are a number of approaches that can be used for face recognition, where the quality of results can vary. In our research from 2016 to 2017, we compared simpler and more complex methods.
Our image database
Reliable detection and recognition of faces are one of the most important and most conducted research tasks. To validate the success of individual methods we created our own image database containing a few thousand images taken by 8 different people. To boost the quality of results we added photos of random people from open sources.
Figure 1: Image database
Detection of facial features
The first step is to detect the face itself on the image based on basic features of the face contour and shape. There are many tools today that can solve this problem. Using simpler methods it’s possible to do the extraction, for instance, using dimension reduction by analyzing major components. The extraction filters unwanted additional information from the image and highlights the most important features. In our case, we used a histogram-based gradient detector, where results are features characteristic for recognizing a face based on the number of occurrences of each gradient orientation.
Simple convolutional neural network
To recognize the face on the image based on the features collected, it’s necessary to include a classifier. For example, the historically more common used Support Vector Machine is the easiest to use, which linearly divides the symptom space into subspaces that represent individuals. In the conducted research we used a simple convolutional neural network with the VGG architecture as a classifier, containing together 16 convolutional and fully connected layers, where convolutional layers are responsible for the extraction of features from the input faces and the fully interconnected layers perform the function of a classifier at the end of the architecture.
The key step of this process was the usage of the pre-trained convolutional neural network on one of the large publicly available image databases. The images from our own created image dataset containing features characteristic for people we want to detect are used to train the last fully interconnected layers in the architecture.
The research outcome was expanded and used for creating demo tools that were presented at events. Initially, the database was expanded with 2.5 thousand images of famous celebrities. The goal was for the tested person to stand in front of a camera, detect their face in real time and uncover which celebrity from the dataset has the most similar face. Evaluating the success of this method is very subjective. However, the fact that each person's face was represented by only one image in the celebrity database played a crucial role.
Related products and services
Data integrations typically combine machine and contextual data, which is available in various structures. We will help you correctly join big dat...
Machine learning enables us to introduce higher levels of automation to many business and manufacturing processes across various industries. Artif...
Read our blog
Revealing Sensitive Documents with NER: Case Study06. 8. 2019Read more Revealing Sensitive Documents with NER: Case Study
Revealing Sensitive Documents with Entity Recognition22. 7. 2019Read more Revealing Sensitive Documents with Entity Recognition
We must know, we will know
Expert team in big data and AI
Our team has presented hundreds of insights in many possible formats. We use tools and methods developed and used by scientific teams dedicated to research.
We strongly consider the existing business environment, capabilities to execute and skill of the staff. This enables us to provide minimum risk and bring quick success to your company.
Working with the best innovators
Cloudera, Microsoft, Clever Analytics, Apache Kafka, Apache Spark, Power BI, Tableau, Jupyter Notebooks