Tensor Decomposition Computing Unit Based On FPGA And Its Application In Face Recognition

In today's society,as society in the digital age,the structure,relations and relevance of of data play a greater role than the number itself.Meanwhile,with the deepening of digitalization in our life,the complexity of data's scale and correlation have also increased geometrically.In order to process these data,the tensor,which is originally used only in mechanical systems,is used to describe the association and structure of these complex data.Because the tensor can represent high-order data,it has become a research respect it various fields,and is used to describe the structure of data in data mining,image processing and etc.To study and analyze the tensor,appropriate analysis methods are also being needed.However,the frequently way to transform tensor into matrix will not only destroy the structural information of tensor,but also lead to higher computational complexity over-fitting problems and larger memory requirements.Therefore,the analysis method needs to preserves data structure and data association.Tensor decomposition is tensor analysis method of compliance with requirem.However,for tensor is a relatively new concept,not many tools supperts tensor decomposition.Meanwhile,algorithmic implementation is also limited to PC platform.As the most popular customized platform for programmable algorithms,from customized chips in mobile phones to reconfigurable computiog platforms for high-performance computers,FPGA is used as a customization module.Meanwhile,the platform of FPGA has abundant computing and logic resources,as well as strong computing power,high energy consumption ratio and reconfigurable advantages.If the tensor decomposition algorithm can be implemented on the FPGA platforms the application scope of the tensor decomposition algorithm can be greatly expanded.In this paper,the most two common algorithms for tensor decomposition,CANDECOMP/PARAFAC(abbreviated to in this article)decomposition and Tucker decomposition,will be described in detail here.Meanwhile,the algorithms are revised with the help of FPGA as the implementation platform.At the same time,face recognition is used as the application background,and the performance of CP decomposition and Tucker decomposition in application scenarios is tested.CP decomposition is used for face recognition and Tucker decomposition is used for face classification.The results show that the performance of tensor decomposition on the FPGA platform is better than that on the CPU platform,and the performance of tensor decomposition on the algorithm is also better than that of the traditional algorithm.