Research On Multi-Modal And Multi-Task Face Detection Algorithms

Face detection focus on how to find and determine the position and size of faces in the image,which is widely applied in many scenarios such as security,auto driving and human-computer interaction.As the first step in the face recognition system,it has a significant impact on the subsequent tasks.In real life,the performance of the face detector is heavily affected by illumination,pose,occlusion and some other conditions.Meanwhile,increasing applications require that the algorithms work on multi-modal(Visible/Near-infrared)cameras with limited resources,and analyze the face attribute.On the other hand,the multi-modal and multi-task technology can improve the performance of the face detector.As a result,it is meaningful and challenging to do research on the multi-modal and multi-task face detection.The main work is summarized as follows:(1)A Data Fusion and Gaussian-Center based multi-modal face detection algorithm was proposed.For the reason that it is hard to collect enough multi-modal labeled data for face detection at this stage,the detectors trained on visible-modal data may degrade on near-infrared data.It is hard to make full use of the multi-modal information.To address the above issues,the modal representation processing and the face detector are decoupled.The Model Degradation Difference statistical characteristic is first proposed for multi-modal data fusion.Besides,an anchor-free detector called Gaussian-Center is designed,which utilizes the Gaussian distribution to supervise the bounding box regression task.The experimental results show that the proposed fusion algorithm can get better results than single-modal and average-weighted fusion data.The performances of the proposed algorithm are improved in both public visible-modal and private multi-modal datasets.(2)A deep learning estimated-noise based multi-task face detection algorithm was proposed.Currently,face detection and face attribute analysis are carried out separately.The progress is complex and unable to utilize multi-task technology.Meanwhile,the weights of different tasks are always determined manually,which is not able to make full use of the relationships between different tasks.To address the above issues,a two-stage framework is adopted,and a multi-task network is designed for joint learning face detection and face attribute analysis.Besides,the objective functions with estimated noise are derived in theory,and the noise is estimated by the network to balance different tasks automatically.The experimental results show that the detection performance of the multi-task algorithm is improved on the FDDB and CelebA datasets compared to the single-task method,and the attribute analysis performance is improved on the CelebA dataset compared to the average-weighted method.In future,on one hand we can further explore the method of utilizing multi-modal information.On the other hand,we can explore more effective optimization methods for multi-task learning.With the help of multi-modal and multi-task technology,we can boost up the detector performance in different ways.