Research On 3D Face Modeling Optimization For Embedded Systems

3D face reconstruction is a popular research direction of computer vision.In normal circumstances,reconstruction can be realized by depth camera.However,the cost of the depth camera is high.If the monocular two-dimensional camera can perform face reconstruction without constraints,it will reduce the cost and promote the application of the face recognition technology,which has practical significance.However,monocular three-dimensional face reconstruction has following problem that three-dimensional face estimation is unstable,resulting in a significant difference in the three-dimensional model of the same individual,or too generalization.In recent years,the reconstruction algorithm based on convolutional neural network has been well studied.However,it usually consumes a large amount of computing resources,and it takes long operation time,which is difficult to be implemented on a common embedded computing platform.Considering the problems mentioned above,this paper focuses on the monocular 3D face reconstruction technology that can be applied to embedded system applications.Specifically,the main research contents of this paper are listed as follows:The neural network acceleration chip SPR2803 S used in this paper has the advantages of low power consumption and excellent performance;it can provide more than 2.8 trillion calculations per second at 0.3W power.It is expected to achieve high efficiency through certain optimization techniques on the chip platform.(1)Although the neural network acceleration chip SPR2803 S has excellent performance,the specific network structure cannot be directly operated on the platform.In this paper,the convolutional neural network model is studied,and the SPR2803 S is optimized.The customized design of the network is implemented on the platform.(2)Overcoming the insufficiency of 3D face dataset under unconstrained conditions,transforming 3D face shape and texture parameters into rich 2D image dataset,and using popular convolutional neural network for regression training,improving parameters estimation accuracy.The test results show that the 3D root mean square deviation is reduced from 1.75 to 1.57 compared with the previous one,and the face recognition rate on LFW is improved from 72.25% to 92.35%.(3)Optimize the loss function,adopt the asymmetric Euclidean loss,and eliminate the error of the average surface shape.(4)Under the premise of not affecting the operating efficiency,the convolutional neural network model is compressed to adapt the relatively small memory of the embedded system,and avoids the significant decline of the recognition rate.Test results shows although the 3D root mean square deviation of the compressed algorithm is increased 0.04,the recognition rate is decreased 1%,the model size is decreased 47%.The calculation time is reduced from 11 s to 200 ms,which effectively improves the modeling efficiency.The embedded system designed in this paper can realize 3D face modeling with low power consumption,excellent performance and high precision,leading to significant application value.