Recognition Of Multiple Faces In An Image Based On Mobile Platforms

Face recognition has a wide range of applications in the areas of access control systems,check-in systems,and self-service stations.The current applications have stricter restrictions on the scenarios used to ensure higher accuracy.Moreover,the current face recognition systems only recognize one person at a time,and do not process the face in batches,and the recognition system is inefficient.Compared with single face recognition,multiple faces recognition simultaneously detects and recognizes multiple faces,which can improve recognition efficiency.The multiple faces recognition algorithm deployed on the mobile platform can be applied to indoor and outdoor classroom attendance,customs and high-speed rail stations authentication and other scenarios.However,multiple faces recognition needs to deal with challenges such as different face poses,lighting effects,occlusion,and small face size.Due to the computing resources of mobile platforms,some high-precision face detection and recognition algorithms cannot be efficiently ported on mobile platforms.Therefore,how to implement a fast multiple faces recognition algorithm on a mobile platform with limited resources still has challenges.For mobile platforms,this paper design an efficient multiple faces detection and feature extraction algorithm which based on convolutional neural network,and at the same time optimize and simplify the network structure.Based on the lightweight network Mobile Net V2,the network structure is further compressed,making the algorithm more suitable for deployment on mobile platforms.In the face detection stage,the idea of “from coarse to fine” is adopted,and three networks are cascaded for detecting human faces to realize fast face detection.The feature extraction network maps facial features into low-dimensional vectors.The network improves the structure of the universal deep network for the characteristics of face recognition,and accelerates the network inference speed while achieving accurate results.The model of the multi-face recognition algorithm only occupies about 5 MB of storage space.In order to further accelerate the speed of network inference,CPU parallelization calculation will be utilized in the deployment to accelerate the operation of the algorithm and further improve the performance of the algorithm.The multiple faces recognition algorithm is implemented on the mobile platform,and the NEON instruction set is used for parallel optimization acceleration.The results show that multiple faces on a 16 megapixel image can be detected within 2 s,and feature extraction for each face can be completed in about 40 ms.Algorithmic accuracy is tested separately on public datasets and local datasets.The results show that the algorithm has a higher accuracy while significantly improving the running speed.The detection rates on the public dataset FDDB and the local dataset are 91.36% and 95.79%,respectively.The accuracy of small-scale face identification on public dataset Mega Face and local dataset is 95.05% and 75.34%,respectively,which has practical value.