Research On Dynamic Face Recognition Method Based On Convolutional Neural Network

In recent years,facial recognition technology has been widely applied in various areas of life,such as facial payment,attendance tracking,and identity verification.Currently,facial recognition primarily relies on static image-based recognition methods,requiring individuals to stand in front of a camera for face capture and subsequent matching against a database of faces.However,in real-world scenarios,there are challenges related to the dynamic nature of pedestrians and the distance between them and fixed image capture devices.These challenges include variations in facial scale,occlusions,and overlapping objects,as well as the issue of existing recognition algorithm networks being too large.Therefore,this study focuses on the research and development of dynamic facial recognition using the Rtina Face and Face Net algorithms,involving algorithm improvements,replacement of post-processing algorithms,optimization of loss functions,and the design of an overall detection and recognition system.The main research tasks are as follows:Addressing the problem of varying facial scales,occlusions,and the difficulty of balancing speed and accuracy in motion scenarios,this paper proposes an improved facial detection algorithm called D-Retina Face.To enhance the overall performance of detection while balancing speed and accuracy,a high-precision lightweight Mobile Net V3 network is employed as the backbone network for the detection module to extract basic facial features.Additionally,an ECA attention mechanism is introduced between the backbone network and the feature pyramid connection to improve the capability of extracting facial information at different scales.Furthermore,to enhance the detection ability in cases of occlusion and overlapping,an adaptive non-maximum suppression algorithm is introduced in the post-processing stage,adjusting the threshold adaptively based on the density of facial candidate boxes.The proposed algorithm is validated on the FDDB and Wider Face datasets,and experimental results demonstrate its effectiveness.Moreover,this improved algorithm also performs well in the detection of dynamic faces in videos.To address the issue of large parameter size and high computational complexity in recognition networks,which results in slow convergence during model training,this paper presents an improved facial recognition algorithm called M-Face Net.Firstly,to reduce the number of network model parameters,a lightweight Mobile Net V1 network is used as the backbone network for feature vector extraction.Secondly,an auxiliary classifier based on the cross-entropy loss function is introduced in addition to the Triplet loss function to accelerate the convergence speed during model training.Experimental results on the LFW dataset demonstrate the effectiveness of the proposed algorithm.Finally,a dynamic facial recognition model is designed by combining D-Retina Face,and its feasibility is demonstrated through experiments.