Research On Occluded Face Recognition Algorithm And Its Feasibility Of Edge Deployment

With the rapid development of science and technology,face recognition technology is becoming more and more mature,and the requirements for face recognition are also becoming higher.For example,in the case of occlusion,the traditional face recognition algorithm is difficult to achieve the desired goal,but the application scenarios of occluded face recognition are increasing,so it is more and more important to study how to improve the reliability of the face recognition algorithm in the case of occlusion.In the process of edge deployment,most platforms are still dominated by traditional algorithms,which makes it difficult for the face recognition accuracy of various common detection terminals to achieve the expected results when they encounter occlusions.Advanced occluded face recognition algorithms conflict with the issue of edge deployment,and it is imminent to improve the edge deployment feasibility of face recognition algorithms in occluded situations.First,based on domestic and foreign literature research,we design an advanced face recognition algorithm based on occlusion detection.The occlusion area location is realized by the coarse-grained occlusion block classification algorithm based on the FCN,and the encoder module fused with the coarse-grained occlusion block classification algorithm improves the backbone network of feature extraction and then realizes the feature extraction of the unoccluded area of the face.Among them,an occlusion degree quantification algorithm is proposed to solve the problem of uneven distribution of samples and accelerate the convergence.At the same time,the occlusion degree can be evaluated for the occlusion test set.In this work,two types of edge deployment solutions are developed based on the above-improved algorithm design.The OFAF acceleration framework is designed based on the FPGA platform,which improves the energy efficiency of the system and reduces the design difficulty.On the general edge GPU acceleration platform,deep fusion and inference acceleration through computational graph The layer fusion scheme achieves lightweight edge inference acceleration.Secondly,based on testing the existing face datasets,we organize and open-sources a dedicated dataset for face occlusion testing—GFRTD.GFRTD contains 27 kinds of test data:unoccluded face data,6 kinds of face image data with occlusion of a human face,and 20 kinds of randomly occluded face data.Each test data has a total of 35,806 face images of1,708 categories.a variety of occlusion test data are used to solve the problem that the current mainstream occlusion test set is incomplete and the occlusion type is single.Finally,in the test of the occluded face recognition algorithm,we verify the effectiveness of the designed algorithm for occlusion location and occluded face recognition.The performance test results of the face recognition algorithm based on occlusion detection are as follows: In the AR and LFW occlusion dataset tests,the algorithm achieves 100% and98.83% respectively,which are higher than the mainstream occluded face recognition such as PDSN and FROM.Algorithm.In the maximum occlusion block size test of the GFRTD test set,this algorithm increased the 45 pixels of the original most advanced occluded face recognition algorithm FROM to 60 pixels of this algorithm,which significantly increased the maximum identifiable occlusion block size.In the deployment test,the reasoning delay of lightweight deployment on the FPGA platform Ultra96-V2 is only 2.89 ms,and the power consumption is only 2.56 W.After quantizing the network on the edge GPU platform which is Jetson Xavier,it has reached 77% of the server GPU RTX 2080 Ti compared with before optimization,the system energy efficiency is improved by 3.2 times,achieving a balance between performance and power consumption on edge devices.