engleski

Multi-Class U-Net for Segmentation of Non-biometric Identifiers

Ubiquity of image and video recording devices, as well as the increasing ease of sharing multimedia contents containing people without their permission induces serious privacy risks. Despite considerable efforts in research on de- identification of such contents, potentially identity-revealing information present in soft and non-biometric identifiers is often neglected. We propose an approach for segmentation of non- biometric identifiers intended for use in a de- identification pipeline that takes into account potentially identity-revealing characteristics such as dressing style, hairstyle, personal items, etc. The proposed approach is based on an adaptation of U-Net fully convolutional deep neural network architecture.

De-identification, Semantic segmentation, Deep learning

nije evidentirano