Research On The Profile Face Detection In The Wild

With the rapid development of Internet and Artificial Intelligence,and as an important research direction of computer vision which closely relates to people's life,face detection has drawn wide attention in many fields.Face appearance is the most important biological characteristic of the human body,using computer to detect faces intelligently,the detection can make people's life more convenient,safe and automatical.Face detection technology has achieved a great breakthrough in recent years,products and technology on face detection have also been applied in people's daily life.For example,when face detection technology is embedded in the camera,it can achieve the function of auto-focus on the face;embedded in the beauty camera,it can more accurately locate the face area that needs to beautify,and achieve the effect of automatic beauty;embedded in the cameras that place in the city's station,customs,airports and so on,it can help detect and track criminals,thereby speeding up the public security to solve the case,and finally,ensure urban security.However,the images in the wild are usually influenced by light,occlusion,and pose,which will negative influence the performance of the face detection systems.After years of research,most of face detection algorithms can detect frontal faces accurately,yet,the detection of non-frontal faces such as profile faces,is still a challenging problem.The related problems of profile face detection in the wild will be investigated in this work,including collecting and annotating the profile face dataset,analyzing the detection effect of the current state-of-art algorithms on profile face dataset,and studying the influence on the performance of profile face detection when the face bounding box include ear or not.Finally,based on these findings,a high performance profile face detector is designed and implemented in an integrated manner.The face dataset is the most important part of face detection technology,although there are already a large number of face datasets.Yet,a benchmark face dataset for profile face detection has not been found.Thus,this paper collects a profile face dataset: celian4034,which was obtained by searching images in the wild through the Internet,filter and sort out images containing profile faces manually.Furthermore,the face datasets used in this paper also includes 360 face,celian200,LFW,and FDDB.In order to study the face detection problem more comprehensively,this paper annotates and calibrates the face region of all datasets in the way of rectangular box.When analyzing the detection effect of the current face detection algorithms on different types of face datasets,this paper uses the face detection algorithms Viola & Jones,DPM_baseline,LAEO and CNN_FD to detect faces in the face dataset LFW,FDDB,celian4034 and 360 face,respectively.LFW and FDDB represent the frontal face dataset,celian4034 represents the profile face dataset,and 360 face represents the multi-view face dataset.By comparing and analyzing the experimental results,we find that these algorithms have poor detection effect on the non-frontal face dataset,Especially,the detection rate is very low,which indicates that profile face detection in the wild is not solved completely,and remains a challenging problem.Should the face bounding box include ear information? What is the corresponding influence on the performance of profile face detection? This is an uninvestigated issue.In this paper,six different face detectors are trained through the object detection framework Viola & Jones,DPM and Fast R-CNN based on the dataset celian4034,and different face bounding box: face bounding box with ear and without ear.By comparing and analyzing the results of these profile face detectors,it can be found that the face bounding box of the training image contains the ear can improve the performance of the trained profile face detector.Moreover,DPM and Fast R-CNN framework are more suitable than Viola & Jones to train a profile face detector.Based on the above experimental results,this paper uses DPM and Fast R-CNN framework to train two different profile face detectors in the dataset celian4034 and the face bounding box with ear,then,integrate the detection results.In the process of integration,it uses a variety of integration strategies: delete small box processing,NMS processing and remove-overlap processing.We compare the integration algorithm with the pre-integration algorithms,as well as some mainstream face detection algorithms.The experimental results show that the integrated algorithm has good detection effect on both the frontal face dataset and the profile face dataset.Compared with the algorithm DPM_celian₁0 and FRCNN_celian,the integration method has improved both the accuracy and recall rate.Compared with the face detection algorithm Zhu et al.,DPM_baseline,CNN_FD and Face++,the integration algorithm achieves the highest detection rate.The datasets collected in this paper,the large-scale experimental analysis of the situation of current face detection,the research on the influence of ear on the performance of profile face detector,these research contents have certain reference value and guiding significance for the development of face detection.Finally,this paper designs and implements a high-performance profile face detector in an integrated way.The results of this study improve the performance of face detection in the wild,and have a positive effect on the research of profile face detection in the wild.