Research On Small-scale Face Detection Algorithm Based On R-FCN

With the rise of deep learning,face detection has made great progress in the field of computer vision,which is widely used in security monitoring,identity verification,human-machine interaction and other aspects.Face detection is an important basic technology for face-related applications such as face analysis,face recognition,and face reconstruction.The images in low constraint scenes may contain many small-scale faces,these small-scale faces are low resolution,the features that the detector can extract from small-scale faces are limited,which will cause missed detection and greatly reduce the accuracy of face detection.Therefore,this paper researches small-scale face detection based on R-FCN(Region-based Fully Convolutional Network).The main work is as follows:1)R-FCN is modified according to the specific attributes of human faces.Firstly,considering the occlusion between dense faces,the NMS(Non-maximum Suppression)adopted by R-FCN is changed to Soft NMS(Soft Non-maximum Suppression)to reduce the rate of missed detection;Secondly,the Anchor is reset according to the scale distribution of faces in the Wider Face dataset,so that the Anchor can cover the majority of faces and improve the detection rate.Finally,experiments verified that the improved R-FCN has better performance.2)Aiming at the problem that R-FCN is detected at the last feature map of Res Net,the extracted features lack the local information of shallow layers,which will lead to the reduction of detection rate for small-scale faces.A bottom-up feature fusion branch is built to enrich the local information of high-level features.The specific solution is: a bottom-up fusion branch is built by using corresponding element addition,and the features are normalized before fusion;the experiment is designed to generate candidate boxes and classify on different fusion layers,the experimental results show that the precision on the fusion layer after Res4 is the highest.3)R-FCN uses fixed-size convolution kernels for feature extraction,which damages the discrimination between features.In order to further improve the detection rate,a RFAB(Receptive Field Adaptation Block)is introduced after the shared feature map.This block not only considers the influence of eccentricity on receptive field,but also has the ability of adaptive receptive selection of receptive field.Compared with the basic improved R-FCN,the precision of the final method on three subsets of Wider Face is improved in varying degrees,which verifies that the proposed method is more robust to small-scale face detection.