Multi-task Face Analysis Algorithm Based On Deep Learning

With the successful application of deep learning in the field of computer vision,the performance of the algorithm has been greatly improved.The computer vision algorithm based on deep learning has been applied in many occasions to process pictures and videos.Face detection and analysis,as an important branch of computer vision,is also affected by deep learning,and has made great progress.However,most of the existing algorithms based on deep learning deal with the task of face detection and face correlation analysis separately,without using the correlation between the tasks to train the network cooperatively.Compared with using single model to achieve multiple face analysis tasks,separate processing needs to detect the face in the input image first,and then run a variety of required face analysis algorithms separately on the detected face region.This kind of framework makes the whole system can't be trained end-to-end,and takes up a lot of computing resources.There are a few algorithms which based on R-CNN framework can use single model to complete face detection and face analysis tasks at the same time.However,due to the lack of end-to-end training,these methods need to generate and store possible candidate areas and corresponding tags during training.This makes the training more complicate and takes up a lot of space.In order to make up for the shortcomings of existing methods,we design and implement two kinds of single model and multi task face analysis algorithms.The main contributions are as follows:(1)We propose a method which using detected key points and their embedding tags to complete face detection,facial landmark localization and gender recognition at the same time.Experiments show that our method can complete effective face detection and achieve mainstream algorithm's performance on facial landmark localization and gender recognition.(2)We propose an algorithm that can complete face detection,facial landmark localization and gender recognition simultaneously in real-time.We improve the original feature fusion network of Yolo v3.We fuse the locate feature from the lower layer of the Darknet-53 to promote the network's performance of facial landmark localization task.Experiments show that our new feature fusion network can improve the performance of facial landmark localization task.Also,compared with the method based on key points and correlation labels proposed in this paper,our method based on single-shot object detection framework has better performance on face detection and gender recognition,and can complete real-time processing.