Research And Construction Of Facial Racial Attribute Dataset

In recent years,with the development of hardware equipment,all kinds of monitoring equipment in the city and life continue to emerge,massive monitoring images greatly expand the application of computer vision scenarios.In this context,it has become one of the important tasks in the security field at home and abroad to study the monitoring image understanding technology and realize the rapid response and accurate prediction of the emergency by computer.In video surveillance,pedestrians are the main targets.Pedestrian face information contains a lot of information,face attribute recognition is aimed at these information mining and has been widely concerned by researchers,and it plays an important role in face recognition and face detection and other tasks.Facial attributes commonly used in existing datasets,such as age and skin color,change with time,but as an inherent attribute of face,face race attribute does not change with time,which makes the study of face race attribute of high value.However,the current research on face race attributes has not been greatly developed,partly because there is not enough large,accurately labeled and balanced dataset of face race attributes for training of deep learning algorithms.Therefore,it is of practical significance to construct a face and race attribute dataset with sufficient data volume,accurate annotation and balanced data.It is expounded in this article to build a data quantity is big enough,with accurate data and the necessity of balanced face racial attribute dataset,face attribute recognition was introduced in detail and related knowledge,design and build a new face image dataset with racial attribute FDEA,finally in the dataset on the corresponding experiment and result analysis.The dataset construction is mainly divided into three steps.First,the face images from Celeb A dataset are manually selected and ethnically labeled.Secondly,in order to further expand the size of the dataset,a certain amount of face images were obtained from the dataset LFWA+,MORPH,UTKFace and Fair Face,and the extracted face images were carefully cleaned manually in this process.Finally,in order to ensure the balance of the dataset,a certain amount of face images are obtained from the Internet and processed.The FDEA dataset contains 157,801 individual face sample images labeled with three ethnic attribute categories,namely "Caucasian","Asian" and "African".The number of images corresponding to the three ethnic attribute categories was 54438,61522 and 41841,respectively.FDEA was also benchmarked.Eight classical deep neural networks were tested on FDEA and the results showed a baseline accuracy of 0.95,which partially validated the validity of the FDEA dataset.