(1) Deep Neural Network (DNN) training and optimization algorithms for video applications (2) Optimization tool and performance simulation models in Matlab/C/CAFFE (3) Investigation report on DNN structures for video applications and hardware architectures (4) System functional specification of Intelligent Video Accelerator (IVA) FPGA board (5) Micro-architecture specifications of IVA functional modules (6) FPGA implementation and verification of IVA (7) IVA evaluation board and system for video applications (8) Performance evaluation report Customer service deliverables for customer (1): (9) Customized PCIe structured ASIC board with the implementation of intelligent video enhancement (10) Software plug-in for the customized PCIe board Customer service deliverables for customer (2): (11) Customized FPGA bit file of face recognition algorithm
Objective: Deep learning technology is one of the main pillars of artificial intelligence and is the most promising technology to enable intelligent applications, especially in video processing and computer vision areas. The major obstacle in commercialization of deep learning technology is lack of a high-efficiency hardware platform to support the huge computations required. This project aims to develop a dedicated hardware accelerator to undertake the intensive computations of deep neural networks (DNNs) and handle massive data in various video applications. Another objective is to engage target customers for mass-adoption in next step. R&D Methodology: In this project, a topology-mapping micro-architecture is proposed to obtain high performance. A re-configurable processing pipeline will be designed to support various DNN structures. In additional, the training and optimization tools will be developed to train DNN models with optimized structure for various video applications. All these technologies will be realized and verified on the FPGA platform. Impact and Benefits: The project deliverables will strongly push the commercialization of deep learning technology in video related industries. The mass-adoption of the accelerator will greatly improve the efficiency of video production workflow and enhance core competencies of relevant device makers in video production and consuming areas.