Privacy protecting de-identification techniques in image datasets used for self-driving cars

Abstract

자율주행용 이미지 데이터 셋에서 사용되는 개인 정보 보호 기법을 개발하기 위해 차량의 번호판, 사람의 얼굴을 마스킹하여 해당 기법을 개발한다. 본 과정 중에서 민감한 개인정보로 인해 수집하기 어려운 데이터를 증강시키기 위해 객체 중심의 증강 기법을 사용한다. 해당 증강 기법은 객체 주위에 다른 객체를 overlay, underlay하여 다양한 occlusion 상황을 연출한다.|The multi-object recognition is a technology for recognizing objects by receiving RGB images as input. The image datasets used for self-driving cars include multi-objects and privacy. Privacy includes license plates and human faces, and de-identification must be carried out in accordance with the privacy protection Act. Privacy is sensitive information and it isn't easy to collect. In particular, there are no Korean license plate datasets for self-driving cars, and only the datasets provided as a cropped image exists, so it is not suitable for learning. And robustness to occlusions is an important property of real-time recognition systems. That is, a robust classifier should be able to solve the problem even if only a portion of the object of interest is visible in an image. To compensate for this, augmentation through occluded situations is essential. In this research, we proposed data augmentation suitable for image datasets of self-driving cars. This is a simple method for using occlusions effectively as data augmentation. This method is the random patched image is resized and patched, It creates object-aware random occlusion situation and augments the number of license plates. And randomly applies underlay and overlay to create various situations.