Research On Deep Learning Face Recognition Technology Based On Embedded Platform

With the increasing degree of informationization in society,face recognition technology is more and more widely used in life.The mainstream face recognition methods are currently implemented based on deep learning,which has a high accuracy rate,but is difficult to be applied in embedded devices with limited computing resources due to its system model with high complexity and large computation.In recent years,various lightweight convolutional neural networks have been proposed,bringing new ideas to the application of deep learning face recognition systems in embedded platforms.In the above context,this paper implements a face recognition system based on deep learning and deploys it to run on a low-cost and low-power embedded platform.The main work is as follows.(1)Build the embedded development environment and study the neural network acceleration technology.We analyze in detail the relevant features of the embedded hardware platform and describe the process of building the embedded development environment,including the establishment of the cross-compilation environment,the compilation and burning of the embedded platform firmware and the transplantation of OpenCV.Regarding the neural network acceleration techniques,the lightweight network structure,network pruning and model quantization techniques are highlighted.(2)Lightweight face recognition techniques are studied.In the face detection part,the face detection network Retinaface is used,the features are extracted using the lightweight backbone network Mobile Net0.25,the multi-scale feature fusion is achieved using the feature pyramid,the face detection frame is generated by the anchor frame(Anchor)mechanism,the redundant information is filtered out using non-maximal value suppression,and the face alignment is achieved using the affine transformation of the image.In the face recognition part,MobileFaceNet,a lightweight face recognition network,is used to extract face features,and face feature matching is achieved by calculating the cosine similarity between different features.(3)Optimize the face network and quantize the model.The face detection network uses a smaller input size,and a smaller input size can effectively reduce the computation.For this purpose,the anchor frame size of the face detection network is optimized to make it perform better in small target detection.The SE channel attention mechanism is added to the face recognition network MobileFaceNet,and the number of network model parameters after adding the SE module is only23 K higher than before,and the test result on LFW dataset is 99.60%.The quantized size of the face detection model is only 0.4MB,and the quantized size of the face recognition model is only 1.4MB,while still maintaining a high accuracy rate,which is more conducive to the deployment of the embedded platform.(4)Deploy the face model and implement the face recognition system in the embedded platform.The face detection and face recognition models are deployed in the embedded platform and various modules of the face recognition system for the embedded platform are developed and implemented.The face recognition system takes an average of 55 ms from acquiring video data to outputting recognition results,and the system is functional enough to complete from face information acquisition to outputting face recognition results.In order to test the accuracy of the face recognition system,1250 image data are used for testing,and the accurate recognition rate of the system can reach 99.0% after testing,which has a high accuracy rate.Therefore,the research of deep learning face recognition technology based on embedded platform in this paper has good application value.