Research On Manifold Representation Method Of Face Image Based On Multiscale Feature Dimension

Face detection and recognition is an important research topic in the field of pattern recognition and computer vision,and it has broad application prospects in the fields of public safety,intelligent monitoring,multimedia as well as digital entertainment.In practical applications,face images are subject to a variety of factors,resulting in challenging face detection and recognition issues.Face detection is the premise of recognition,and in the case of the more complex the scene environment,the requirements for detection are more stringent.Among them,the key factor affecting the accuracy of face detection is the missed face in a complex background.In addition,information security issues are getting widely attention.So it is difficult to obtain a complete database of faces composed of multiple pictures of each person in different scenes.The small sample size problem is the main reason why many algorithms fail to achieve accurate classification.The main work in this article is summarized as follows:1.In view of the high efficiency and detection accuracy of multiscale feature fusion network,a series of optimization strategies are proposed to improve the detection accuracy of YOLOv5 for dense small faces in complex scenarios.First of all,the new activation function is applied to improve the training potential of the network,and the function is theoretically analyzed and experimentally verified.Stem module is used instead of Focus module to improve the generalization ability of the network and reduce the complexity of the calculation.Reducing the maximum pooling kernel size in SPP module to adapt to the small face target and connecting with the two-way multiscale feature fusion network;Secondly,Io U loss function is analyzed in depth,and Alpha-CIo U bounding box regression is used to improve the regression accuracy.Finally,the model training and verification experiments are carried out based on the WIDERFACE dataset,and we use ablation experiments to prove the effectiveness of the proposed strategy.The m AP index of the proposed algorithm in the hard subset reached 83.3%,and the model size is 7.075 M,which is suitable for deployment on mobile devices.2.The existing theories related to representation learning and manifold learning are hot topics in the field of pattern recognition.In this thesis,a small sample face recognition method based on multiscale manifold space is studied to solve the problem of insufficient training samples,which results in low classification accuracy.Based on the theoretical results of high-dimensional face image classification under multi-manifold hypothesis,and combined with local constraint linear coding method which has the advantages of maintaining manifold structure and reducing the complexity of coding calculations.Last but not least,a multiscale manifold space is constructed by multiscale segmentation tiles to enrich the training set and reduce the impact of tile scale variation on algorithm performance.The experiment is carried out on three datasets,and the effect of the algorithm in the case of small samples is verified by reducing the training sample size in turn,and also verified that the algorithm had better recognition accuracy in non-limiting scenarios.