Private Photos Detection for Modeling Preferences of Recommendation System's Users

Nowadays, the photo privacy detection is becoming an acute task due to a wide spread of mobile devices with photos published on social networks. As a photo might contain private or sensitive data, there is an urgent need to accurately determine them and impose restrictions on their processing. In this paper we focus on the task of personal data detection in a photo gallery. A novel two-stage approach is proposed. At first, text of scanned documents is processed based on an EAST text detector, and extracted text is recognized using Tesseract and neural network classifier. At the second stage, face clustering is implemented for the remaining photos to identify large groups of people (friends, relatives) whose photos also refer to personal data and must be processed directly on a mobile device. The remaining images can be sent to a remote server for processing with higher accuracy. The experimental results of text recognition and face clustering methods using various convolutional networks for facial features extraction are presented.