Research On Methods Of Low-Light Face Detection Based On Feature Fusion

With the innovation of artificial intelligence technology,the development of the national economy and the needs of security and defense,face detection technology has developed rapidly.However,most face detection algorithms are limited by application scenarios,such as serious noise and low contrast of face images captured under low-light conditions,which greatly affect the accuracy of existing face detectors,especially the lack of localization ability of faces in small areas.General face detection algorithms cannot meet the needs of night monitoring and low-light face recognition.Therefore,it is imperative to study efficient low-light face detection algorithms.In view of the above problems and challenges,based on the low-light image enhancement technology,this paper deeply studies the lowlight face detection algorithm and related technologies,and proposes two efficient low-light face detection schemes from the perspective of channel feature information and feature position relationship of low-light face images.(1)Low-light face detection algorithm based on cross-fusion of high and low frequency channel features.Aiming at the problem of insufficient feature extraction by the existing low-light face detection network,this paper designs a module for separable high and low frequency channel features(HLC)based on low-light image enhancement technology and makes full use of feature channel information,and a novel high-low frequency channel feature fusion method,Cross-frequency feature pyramid network(Cross-FPN),is proposed based on this module.This method uses the HLC module to separate the high and low frequency information of different scale features,and then cross-fused separated high and low frequency features,which provides features that fuse high-frequency details and low-frequency chromatic information for the detection network,so as to improve the performance of the detection network.A large number of experiments prove the effectiveness of the algorithm,and the experimental results show that this proposed method outperforms the baseline method by 4.0%m AP.(2)Self-coordinate attention-guided low-light face detection algorithm.Aiming at the problem of inaccurate face positioning by the existing low-light face detection network,based on the feature position relationship,this paper proposes a self-coordinate attention(SCA)to guide low-light face detection algorithm.The method first encodes in the two directions of H and W of the feature space to form a position matrix,and then aggregates them into direction-aware attention respectively to enhance the sensitivity of features to positions.Considering the positional shift of features during forward propagation in the network,a corrective attention mechanism SCx A is proposed,and combined with the above SCA attention,a self-coordinate attention-guided feature fusion architecture SCA-FPN is constructed.Based on this architecture,this paper improves the low-light face detection task by 3.9%m AP compared with the benchmark method.In summary,aiming at the problems of insufficient feature fusion and inaccurate positioning during detection of existing face detectors in low-light conditions,efficient low-light face detection schemes are designed from the channel information and position relationship of features,and the effectiveness and rationality of the proposed method are verified on the DARK FACE dataset.This study provides a new perspective and new ideas for the task of low-light face detection,which has certain application value and research significance.